From InstructGPT to Jev: What Comes After ChatGPT — Diogo Almeida, TypeSafe Co-founder & CEO
- Diogo Almeida's Jev launch is a bet on a new model class — "System One" or "large programmable" models where "the goal is for code to be the consumer," not chat. Unlike RLHF ("please humans") or RLVR ("optimize benchmarks"), TypeSafe's task, which he calls RLCD, targets reliability for programmatic use and the frontier of intelligence per dollar — hence Jevons paradox. He frames it as the second-to-third major North Star shift in LLMs: "RLVR is a 0.2 in my opinion, and I think that's generous."
- Adoption looks real, not touristy: in less than a week Jev passed 1 trillion tokens per day, "even at night... constantly churning, so you know machines are calling it." Waitlist signups "don't matter for a developer platform"; the scarce resource is rate limits, because software lets users specify a rote task once and run it in the background. Pre-launch, revenue was nearly zero and more than half of testers "did not get it" — a challenge to orthodox product-market-fit thinking.
- His mode-dropping argument explains why long-string error intuitions do not straightforwardly apply to RLHF-shaped models. Calibrated, mode-covering models would include outliers; GAN-like mode dropping instead removes minority outputs and makes generation conservative because visible errors are heavily punished. That may preserve long strings while making string models poor decision engines. He treats calibration, confidence thresholds, and typed outputs such as choice/score/null as product responses.
- He is strongly against public benchmarks and demos as primary evidence, not against measurement itself. His preferred path is "vibes and trust until you put it into a workflow and evaluate it for that workflow." TypeSafe keeps internal evals but treats not gaming them as a top-level priority. He held this line even when it hurt fundraising: "no one believed us... We are principled. We're gonna stand by our guns."
- Safety alignment in an API is, to him, a type error, though he says he is not opposed to safety as a principle. Refusals inside dependencies can make software fail stochastically. Safety alignment may make sense for first-party products like ChatGPT; at the infrastructure layer he compares intelligence to a database and says end-use restrictions should not be built into the technological layer because they fracture general capability.
- His coding-agent thesis is that the named incumbent agents are built around a single-model world, while open agents can experiment with multi-model designs. His follow-on research agenda is "coding agents free from the tyranny of the KV cache": labeled subtask trees, shared state across parallel agents, and cheap historical-context lookup instead of treating continuous learning as a separate problem.
- The macro thesis is explicit: "TFP growth rate: 3% in 5 years." He argues that models are currently tied at roughly zero percent of economically valuable work despite solving hard math problems, predicts an "inverse SaaSpocalypse" in which SaaS is supercharged, and does not expect mass unemployment. He promises deployed models will not be silently changed, while offering no general long-term support commitment; possible model sizes, LTS, and other forms of machine-native intelligence remain under consideration.
- He dismisses both the "pace the frontier" consensus and much of the neo-lab wave. Pacing arguments, in his view, assume everyone must do more RLVR with unrestricted action spaces; for Jev's shape, "zero is the optimal amount." He calls most neo-labs weak unless they have a real North Star task, while advising researchers who do have one to pursue it.
1. Launch week: "never been worse" emotionally, never more in sync with reality
- Diogo's state after Jev took over the timeline: "I'm a ragged corpse of a person right now," but mentally vindicated — "the entire AI field is like one of those carnival houses of mirrors, and everyone is just insane... for just this week, I'm better in sync with reality." His headline claim: "an AI-based economic revolution is back on the table."
- He deliberately prioritized a Discord town hall — the server was then about 100,000 people — over VIP investor meetings: "it feels dirty if, in my gigantic calendar... the community isn't one of those." He even considered hosting a town hall while walking to the studio.
- swyx notes that three topics were trending on Twitter; Diogo discovered that he was looking at his own personalized feed and cringed: "I'm a noob."
2. What Jev is: System One models where code is the consumer
- The definition, in his words: a new class of models — "machine-native, System One, large programmable" — distinct from pretrained LLMs ("autocomplete of the internet"), RLHF chatbots, and RLVR. "The goal is for code to be the consumer... things that directly are consumed by code, hence the name type-safe."
- Jev is the brand for models "on the frontier of intelligence per dollar," named for Jevons paradox, with the launch framed around trade-offs among reliability, cost, calibration, and speed. They avoid "decision models" as a name because "System One is beyond that"; Diogo says they have more work in the tank.
- A relief of launching: "now I don't have to answer that question to my parents anymore, 'cause ChatGPT can just explain it."
3. How mode dropping explains the disconnect with LeCun's doom slide
- Diogo's view is that Yann LeCun's takes are among the closest to accurate, but he explains the apparent disconnect with the famous slide about error probability as sequence length increases through mode dropping.
- In a calibrated, mode-covering distribution, outliers are expected. GAN-like mode dropping instead discards minority classes and retains common outputs. For long strings, visible errors are easy to detect, so the model becomes extremely conservative.
- The punchline for his business: that conservatism can preserve long strings while making string models poor decision engines — "that calibration is total poison into the probability distributions of strings."
- On LeCun's JEPA fix, he demurs pragmatically: "really cool early research. Is it practical yet? Probably shouldn't say." His brand, he insists, "other than unhinged, is practical." Scaling laws can require exponentially more resources for normally sublinear gains, which looks like a bad investment unless the gains are extremely valuable.
4. Safety alignment is a type error in an API
- Diogo says he is not opposed to safety as a principle, but considers safety alignment generally misaligned with users. Capability alignment — doing what the user wants — "is sick for software engineers"; safety alignment is "the opposite of instruction following," because it follows a lab's restrictions. Fine for a first-party product; "in an API, that's nuts... so anti-user."
- The concrete failure mode: a refusal in Claude Code — "I'm sorry, I can't read DNA.py" — is annoying but workable, and users tolerate it through "Stockholm syndrome." But if a dependency refuses in the background, "you want the software to just stochastically break because a user sent a weird message?"
- swyx's pushback — "what if we use it to kill people?" — gets a layered answer: Diogo would prefer good uses and would "put my thumb on the scale" for them, but "will I do it at the technological layer? Absolutely not, because that will fracture the intelligence." His model of intelligence is "more like a database than a coworker": "we are an API... we shouldn't be able to know what the downstream users are doing."
5. Vibes and trust over benchmarks — a stance that cost them a fundraise
- After a preview-period terms-of-service misunderstanding about benchmarking — which swyx says the team would remove, while Diogo says he asked the lawyers to check it — Diogo lays out the position: intelligence has "a je ne sais quoi... the good-model smell," and public benchmarks "are extremely, extremely gameable."
- His example: labs collected data that looked like MMLU to improve scores, "which is just benchmarking with extra steps."
- The alternative is "vibes and trust until you put it into a workflow and evaluate it for that workflow." TypeSafe's job is to keep moving "the nines of reliability." Internal evals exist, but "it takes a lot of discipline not to game those"; not fooling themselves about intelligence is a top-level priority. "We need to be truth-seeking."
- The cost of principle, worn proudly: "last year when we did fundraise, no one believed us... they wanted just benchmarks and we're like, we're not gonna do that. That rewards bad actors." A tell on quality-versus-time trade-offs: "if we wanted to, we could have released Jev a year and a half ago if we wanted it to be dumb."
6. A data lab, not a model lab — and why they won't train on user data
- "We will always care so much about data... model capabilities means data. Data is so unbelievably complicated, and that is what gets nines." They are "hiring infinite data people. Actually infinite."
- All training data is synthetic, deliberately so. Real usage follows a power law — "you'll end up overfitting to it and fracturing to it" — while the target is a sci-fi future where these models are general infrastructure, "kind of like UDP as LLMs and TCP as our models."
- Even having all present-day data would mean overfitting to the present. His picture of a great data person is closer to an artist than an annotator: they study the cognitive core, find its jaggednesses, and address them surgically across past, present, and future. "The general case rather than the specific case" requires intelligence every time.
7. RLCD: the third North Star shift, and his Bitterest Lesson
- His hierarchy, attributed roughly to Suchin: "algorithms beats compute very roughly. Data matters way more than compute, obviously. And doing the right task, having the North Star, is the hardest, most important thing." Task shifts have happened "twice so far, maybe 2.2 times" in LLMs: RLHF for instruction following, RLVR as "a tiny little edit," and now RLCD — "programs in the loop."
- He unpacks why "RLHF" is really a task, not an algorithm, through several historical examples. He is not fully certain about the exact attribution of the early backflip work, then discusses OpenAI's Learning to Summarize work as PPO on language models for an ill-specified task. Dario, Alec Radford, and Ryan are mentioned in connection with the cited work.
- "It's not about the PPO. That part doesn't matter." The North Star is instruction following, just as DPO and its descendants can still belong to RLHF despite not using the original algorithm.
- One nuance he adds to swyx's summary — programmable AI and the human out of the loop — is that they must remain practical. There may be programmatic types that are "sick AF," but if the technology is not ready, it is not a tragedy if they are not deployed yet.
8. The pre-Jev world was a tragedy — and the AI winter he feared is averted
- The emotional core, from why he left OpenAI: "If an AI winter did happen and I did not do every fucking possible thing I could to avert that, I would see myself as personally responsible" — both for the RLHF direction, which he thinks widened overpromise versus underdeliver, and for not going all-in on this. Post-launch he changed the wording: "the AI winter I'm worrying about is averted."
- His favorite provocation: how can AI solve Millennium Prize problems in math while still failing to automate basic rote work? It is "a supercharged engine of automation that just does not have the right plugs."
- The categorical claim: if Tessl disappeared, it might take a year or two for people to catch up, but "it's done, right? This has changed the path of technological history." He admits the original arrogance too: "I thought this whole project would take a week... I was unbelievably wrong, so I am so sorry to everyone at OpenAI."
9. The numbers: a trillion tokens a day, and why waitlists don't matter
- The milestone he volunteers: "a trillion tokens a day is a lot... this is not fleeting tokens per day. Even at night it's constantly churning, so you know machines are calling it and not just people trying things out."
- Signups he actively dismisses: "waitlist signups don't matter for a developer platform... if every single human being in the world just wrote a couple of queries, that would be a rounding error compared to one power user's for loop that is just creating value." The scary variable is rate limits.
- The "marketing genius" narrative gets punctured — "that was just us. We don't have a marketer, also hiring" — and he credits the platform team: "I think we have more nines of uptime than Anthropic while having the most unprecedented launch ever."
- swyx cites roughly 36 million launch-video views, with Diogo updating that to 38 million. He compares it with Navio at 74 million and Fable 5 at 57 million, then says Jev is currently number one among Neo Lab launches in 2026. Diogo rejects the category: "I don't give a shit about being a neo-lab... what I care about is being a reliable dev platform."
10. Determinism is the wrong North Star; robustness is the one that matters
- swyx asks why there is no seed parameter. Diogo reframes: "reliability is actually a catch-all — whenever AI can't automate something, it's due to some form of reliability." Determinism — same input, same output — is "slightly interesting for unit tests" but "the wrong North Star"; robustness — similar inputs, similar outputs — is where "people get burned with AI making decisions."
- Their test is to inject UUIDs or nonces into prompts and demand similar answers to semantically identical questions.
- He will build a deterministic model if convinced it is valuable — "we live to please" — but "it just gets you less intelligence per dollar," and he suspects they will be GPU-constrained for a long time. swyx warns that users will still demand determinism.
- A revealing aside on the Pareto frontier: "we do absolutely disgusting things to be on the Pareto curve of intelligence per dollar." The team told him not to call the result "a Frankenstein's monster of models."
11. Versioning promises: never silently swap a model, but no long-term support either
- On swyx's quantization concern, Diogo makes a categorical commitment: "We will not change our models when we deploy them. That is insane. We care about developers." First-party products can change behavior; an API cannot.
- But "we are not promising long-term support for the models" and will launch new models faster than users may expect. He gives a possible example of temporarily LTS-ing Jev 1.13.0 because developers hate breaking dependencies; the alternative — fracturing the fleet — would be a bad outcome.
- Research is underway toward a more developer-friendly LTS approach, but he does not promise that exact models will be maintained indefinitely. Model versions should get smarter, with the largest deltas appearing when a system moves from jagged to broadly capable.
- On porting LTS models to other silicon such as Cerebras or Etched: "No comment." Intelligence per second is "a different metric"; real-time budgets between 100 milliseconds and 1 millisecond can be valuable, but "I don't think that that will be Jev's niche."
12. The primitives: choice, score, and null (from Bernoulli)
- "Null" is derived from "Bernoulli" and is a continuous, bool-like probability. Rejected names included PBool and Pool — "we were wanting to call it a pool party, but then no one let me." All three primitives are deliberately new concepts: "these are not types that exist in programming... a score is not an int."
- The mapping he endorses: choice maps to a switch statement on an enum, null maps to if statements, and score maps to sorting or thresholding. "There will be more types, and they will map into programming primitives." Choice is closest to a function call, but "a function call is an extremely disgusting thing."
- Internal color: the name Jev drew substantial internal hate. Everyone has apologized except Diogo's co-founder, who wanted it called Meow.
13. Structured state in, and system prompts as "disgusting global variables"
- The under-read part of the API: state, instructions, and criteria can all be structured JSON. "If you're using a template, turning it into a system message or something, you are thinking in the old way... within the computer, you want to be passing nested structure that is semantic all around."
- His indictment of the status quo: "System messages are disgusting global variables where you just put everything in there... and you hope that every single instruction gets nailed instead of asking the questions in parallel."
- The design goal is a flow state where people can write a type-safe System One query and get accurate branching without having to repeatedly test whether the model works. "That is the dream. And that is going to be a long, long slog."
14. How to decompose: many small questions, "ML without the ML"
- His pro tips as the model's heaviest user: break problems into "the smallest semantic unit," be "really structured and explicit" — he likes backticks — and be clear about what each reference means because the model should be literal.
- For refusals, do not ask "should I refuse here?" Ask many independent questions about specific refusal conditions, threshold them, and add missed cases as tests. "You fix the bug by adding that question in... and now it is just solved forever." The decomposition makes the system measurable: "It's like ML without the ML."
- swyx reports that his own one-big-prompt versus 100-call benchmark made decomposition slower, more expensive, and worse. Diogo's counterpoint is that this is inconvenient but produces software that can be trusted and run in the background.
- Parallel questions are cheap: "put IDs on every message and ask a question about each ID... pay for the state once."
- One hedge he volunteers: automated trading with the models makes him "a little bit afraid." It is a high-level task the models may not be good enough for yet; confidence estimates and human-escalation thresholds can help determine when not to deploy.
15. Fine-tuning: maybe, warily — plus model sizes, cascades, and killing the regex
- On locally wrong calibration, he concedes freely: "it will get many things wrong, to be clear." The team wants each model version to be notably better. Fine-tuning "could be in the cards," but narrowing a general model might improve edge cases in one task while damaging generality.
- swyx says OpenAI and Claude, and possibly Gemini, have rolled out and then withdrawn fine-tuning. Diogo agrees that the versions he knows about were poor and that removing them may have been better.
- The sci-fi he is cooking — explicitly "not a promise, a desire" — is calibration plus a cascade of model sizes: confident answers stop early, uncertain ones escalate, and automatic fine-tuning might become possible across a Pareto frontier. "Is there a world where people don't write regexes anymore because the intelligence per dollar is cheaper than the complexity of a regex?"
- Diogo is open to different model sizes because he cannot know how much intelligence each user needs, but this is a direction under consideration rather than a guaranteed shipping roadmap.
- He also says he likes the idea that culture is what a company does when the market does not reward it, while acknowledging he may not have the exact wording.
16. System One versus System Two is an empirical boundary, not a slogan
- Asked to define the boundary, since "people are just trying to Jev everything," he refuses to theorize categorically: "the truth is that this is an empirical problem, just like scaling laws." His claim is that pretrained condensations of intelligence are fundamentally System One thinkers.
- RLVR has done "incredible things for System Two" — "I am in awe" — but the resulting behavior is fragile and jagged. ChatGPT-era models were described as general; RLVR models are described as narrow and brittle, and "math is not just spiky, it's fractal."
- swyx's field report from day-one access: single-hop is "state of the art," while multi-hop starts to fall down as the number of hops increases. Diogo says the work is to unearth, smooth, sculpt, and combine as many capabilities as possible. System One is simply the description of what currently works in that paradigm.
- Diogo does not promise never to use string-based or other forms of reasoning. He says less slow, inefficient, and fragile forms of reasoning remain possible, while his firm commitment is to the machine-native ROI North Star.
- On vision, the obvious ask after the Doom demo: "everything is in the cards." The harder question is what users think they want versus what they actually need. Diogo says, with a hedge, that Jev is "as far as I can tell" by far the best at avoiding long-context degradation, while other providers have simply given users the long context they requested.
- Taking the product philosophy too far toward satisfying stated demands could produce "Anthropic nanny-state-style thinking." Finding the balance between trust, speed of release, and treating users as adults will be one of his major near-term debates.
17. Pre-launch reception was bad — a case against orthodox product-market fit
- The honest history: "the reception was pretty bad... more than half the people we had play with it just did not get it," nontechnical team members feared they were "selling a vitamin and not a painkiller," and there was "almost no revenue before launch." Those who got it asked how to get it through procurement.
- Then the launch: "It explodes and everyone's like, we need as many rate limits as we can. Can we literally give you GPUs?" His conclusion: "I do want to call into question the concept of product-market fit." The mechanism was not simply marketing but "passionate developers whose souls basically resonated at the same frequency."
- His loyalty response is to explore launches "better for developers than enterprises." He dyed his hair and ran garbage-bag town halls because "it felt dirty to me during our company's most important times not to keep talking to them." He asks the community to hold him to that and bring out the pitchforks if he changes.
18. The use-case map: dark data, real-time, verify everything, smart software — and computer use from left field
- The families they mapped from first principles before launch: "dark data" — companies have hoarded data but avoided LLM analysis because it was too expensive — plus coding agents as the big money-makers; real-time and assistant use; "verify everything," meaning checking LLM calls in an observability-like way; and intrinsically composable "smart software," including a programming-language-as-Jev project.
- Computer use "came out of left field." swyx shows a voice-driven demo involving Notes, Arc, x.com, and Photo Booth, with Jev used as a decision model. Diogo's position is both enthusiasm and caution: "just like I'm anti-benchmaxing, I'm also anti-demos. I want to make sure that it works reliably." He still calls it "super fucking sick" and wants it to work for hands-free computer control.
- Community humility: users' builds are "kinda cooler than ours... if that was our demo, holy shit." The cookbooks are "not AI-generated trash"; they are inspired by real customer problems and contain practical techniques.
19. Coding agents: the one-model incumbents' dilemma, and freedom from the KV cache
- Diogo calls the named incumbent agents — Claude Code and Codex in the discussion — the likely number one and two, while explicitly saying he is not fully sure. He says they are built around a single-model world, while open coding agents are experimenting with multi-model designs.
- Since open agents are roughly at parity — "there's not so much you can do with a while loop" — the first killer multi-model pattern could spread quickly across them. He is unsure what the single-model incumbents will do.
- His research agenda, from the article "KV Cache Rules Everything Around Me": the cache locks an agent into one model and encourages continuous appending, which conflicts with state management, abstraction, and decomposition. It helps explain why routing is hard, why sub-agents often fail, and why compaction is difficult.
- Freed from that constraint: labeled subtask hierarchies that can be searched for relevant context; parallel sub-agents reading each other's state with intelligent rather than "basic-ass" locks; and cheap lookup over historical context. "Isn't it kind of weird that you start from scratch every time and need to solve a problem called continuous learning? That's actually a memory-management problem."
- swyx points him to Prime Agent and recursive-language-model work as adjacent efforts.
20. The macro call: 3% TFP growth, zero automation today, inverse SaaSpocalypse
- swyx flags the manifesto line — "TFP growth rate: 3% in 5 years" — as unusual among labs. Diogo says that is what an economic revolution is and connects it to what the OpenAI charter used to stand for.
- His scoring: "all models are roughly tied right now at zero" percent of economically valuable work, although there is some chance the process has begun. When real automation happens, "it will show up in the economic statistics." He says it will not cause mass unemployment.
- The software-stagnation puzzle: "How is the software basically exactly the same [as 2019] despite AI being so freaking awesome, other than sometimes having a chat box on the side?" These systems cannot yet be trusted with decisions that companies have stakes in.
- Hence "an inverse SaaSpocalypse": SaaS is going to be supercharged because incumbent software companies know which work is valuable to automate.
- His aesthetic preference is that AI disappear into the background: "I'm really tired of AI always being the foreground character... the world should just be more delightful, and AI should just help with that." He wants "do what I mean" behavior throughout software.
21. "Pacing the frontier" is a sleight of hand built on an RLVR monoculture
- His rebuttal to the frontier labs' pacing document is that it assumes everyone needs to do more RLVR with broad action spaces. Sandboxing problems could have been solved, he argues, but labs chose greater model power over those constraints. Given those choices, "oh yeah, we're heading into a dangerous world, guys" follows — but the premise has alternatives.
- "I obviously don't think I need to do more RLVR on our models. I think zero is the optimal amount for our shape."
- Who is at fault: "the only people who are at fault here are the researchers." He says the public reasonably assumes OpenAI and Anthropic are doing the best they can, without seeing the other directions available.
- swyx's report that pacing is primarily political positioning, including around the 2028 election, lands hard: "Oh, I wish I didn't hear that... that makes me lose faith in humanity a bit." Diogo's COVID analogy is that overconfident appeals to authority produced a suboptimal response and bad downstream effects. "Misleading people, even for the greater good, is just — I'm not a fan."
22. Origin story: from InstructGPT and copywriting slop to "fuck it, man, we ball"
- He fought to deploy InstructGPT. Early versions used an unpublished algorithm he wrote because cleaning PPO data was too slow, and "basically immediately it took 50% of the market share of LLMs at the time."
- Then the letdown: it was used mainly for copywriting — Jasper AI, Copy.ai, and what is now called slop — and "we were worried we had made the internet a worse place." Everyone, he argues, should be able to explain why superhuman instruction-following was not AGI.
- The reframe that became TypeSafe: working backwards from an AI economic revolution — "what will we be calling the AI if AI is an API? Humans or code? I figured it was many nines of code. But all the optimization was going into the humans part." Sam's response to his document was: "This is so fucking good. You should go work on it."
- He assumed Anthropic must already be working on the same direction; he then characterized ChatGPT as a copy of an internal Claude product that had not shipped. swyx added that reasoning was "first-ish," while Diogo responded that the product question was debatable despite the research quality.
- The founding sprint: he called Eric first; Sasha, whom he was not even recruiting, said "I'm in" and folded her startup. "Within two weeks we had funding, people moved into my apartment — it was the worst 'cause I'm a neat freak."
- swyx links the Thanksgiving GPU run at OpenAI to the coup period; Diogo initially says he is not sure, then remembers the coup as annoying and promises to discuss it later. His self-assessment: "I describe myself as 0% entrepreneurial... I can't imagine anyone doing this twice. Honestly, doing it once is pretty bad."
23. Advice to frustrated lab researchers, hiring, and the AWS of intelligence
- Should others do what he did? "Most neo-labs are crap... they tend to want money to play around with their experiments." His deeper heresy is that he does not value research pedigree by itself; he values people who identify and pursue the right North Star task. He still says researchers are needed.
- If someone has a genuine North Star task: "abso-fucking-lutely, please do. Please break this hive mind." If someone mainly wants to play with research, he thinks an established lab may be the better place. His capital-allocation line remains: "if you gave me a billion dollars, I wouldn't pre-train... Frankensteining is not the most elegant, beautiful thing, but it solves problems, baby."
- Tasks he throws to others include intelligent games — simple state machines for NPCs in something like Stardew Valley — and coding agents free from the KV cache. He also says he will not bake Jev's identity into the model: "I am not going to put into the models that you are Jev from TypeSafe. That fractures it... represent what the internet thinks. Be correct."
- Hiring priorities include "infinite" data people, rebranded as model-capabilities people so they receive higher status; platform engineers; a founding marketer; and people to ship new shapes of machine-native intelligence beyond decisions.
- He says speed of light is a bottleneck and is unhappy that European users were only three times as fast rather than 100 times as fast because there are no servers there yet.
- The end state he wants is "an AWS of intelligence." He says it would be arrogant to claim it will be TypeSafe, but he will do what he can to make it happen. System One is "the TCP of it," and swyx's earlier phrase that Diogo liked was "Build product on God." There are several more layers and model shapes to explore.
Full transcript
Okay, we're in the studio. Special occasion because this week, Diogo, my good buddy, launched Jev, and it's been taking over the complete timeline. How do you feel? What's it like to be you right now?
Emotionally?
Yeah.
Never been worse. I'm a ragged corpse of a person right now because there's so much going on, and I'm a technical CEO, so I have a lot of fires to fight.
Yeah.
Mentally, I feel like the entire AI field is one of those carnival houses of mirrors, and everyone is just insane and saying the weirdest stuff that doesn't make sense. For just this week, I feel like I'm in better sync with reality, and people see it now. AI can be so much more than what was once thought.
Mm.
Yes, an AI-based economic revolution is back on the table, and this is fucking awesome. This is fucking awesome. I'm so jazzed that the developers get it. I want to show my internal gratitude to the developers, and I'm so jazzed about the community and everything. It's so great.
Yeah, you were saying yesterday that you decided to prioritize the Discord town hall and not a bunch of VIP investor-type people because you wanted to make sure that they were the people who got the most attention, right? The engineers, the developers.
Yeah, it felt a little like, “Oh, man, I'm talking to really important people right now.” I probably shouldn't reveal who.
Yeah.
But it feels a little dirty for me. I'm perhaps overly genuine about things. It feels dirty if, in my gigantic calendar of people to talk to, the community isn't one of those.
Yeah.
In my ideal world, it would be community all the time. I was thinking, “Should I host a town hall while walking to your studio?” And I'm like, “No, that's too crazy.”
Sure.
Yeah.
You guys have been hosting town halls on Discord. Discord is now 100,000 people. Your Twitter has blown up.
I don't follow these stats.
Yeah.
Holy shit.
Your Twitter has blown up. It was really funny because at AIE you were like, “Follow me, please,” and then you didn't even provide your handle.
I'm a noob. I'm a noob.
You're such a noob.
I'm a noob.
But that's positive aura—that you don't know how to promote yourself.
Yeah. Someone called me out when I posted, “Holy shit, we're all 3 trending topics.” Then they're like, “That's a personal feed.”
That's a personal—
And I'm like, “Oh, no—cringe.”
Of course it trends to you.
Yeah, yeah, yeah.
Okay, congrats on everything.
Thank you.
We'll talk about more details as you have them, but for people who are living under a rock or just want the definitive thing, what is Jev?
Let me think about that. That's a hard one.
I'm happy to re-ask if you want to break it down.
No, no, no. I'm happy to just jam on it.
Yeah.
1. What Jev Actually Is
The first thing that I'm relieved about with this question is that I don't have to answer it to my parents anymore, because ChatGPT can just explain it.
Nice.
The way I see it is, we need a new class of models. We're not attached to naming that class of models. The most accurate name we've come up with is System One models.
Yeah.
There will be reasons for that, but there's a reason why we don't call them decision models. System One is beyond that. That's all I can say. We didn't expect this to be our big launch, so we have stuff in the tank.
You should have said “low-key research preview.”
It kind of was, right? It kind of was. We have a class of models that we describe as machine-native, System One, and large programmable.
Mm-hmm.
I think this is the class of models where the goal is for code to be the consumer. As opposed to pre-trained large language models, which are meant for autocomplete of the internet, or RLHF models—chatbot instruction-following models—which are meant to reply to text, or RLVR, which is in a weird gray area with RLHF, these are meant to be directly consumed by code, hence the name type-safe.
The thing we really, really, really want is for AI to be as powerful as possible, and we think the way to do that is to integrate it with software. We are designing everything—from the outside to the deep internals of the model—to be optimized for software.
Number 1, Jev is our first large programmable model, or System One model, whatever you want to call it. Jev is meant to be optimized for intelligence per dollar, hence the name Jev.
Jevons paradox.
Jevons, Jevons paradox, yeah. It's optimized for intelligence per dollar. I love this debate with people about what is most important between reliability, cost, calibration, and speed. Jev is meant to be—Jev will be the name of models that are on the frontier of intelligence per dollar. There are other ways to optimize it. In ML, or at least if you're good at ML, it's all about trade-offs, and we're just going all out on that.
Yeah. To me, calibration is one of the new things that people weren't talking about as much. We've done an episode in the past with Clementine Foreal of Hugging Face, where they were like, “Yeah, actually, they're just collapsing toward what you want to hear the most,” or what is most likely, instead of their own internal confidence about a thing.
Can I soapbox on that for a second?
Go ahead.
2. The Mode Collapse Problem
I've heard that your audience is the most technical, so I actually want to get into that.
Yeah.
I went to extreme lengths to make sure everything in our launch video was accurate and real. Apparently, that's very unusual. One of the things that no one paid attention to was the downsides of RLHF, in particular, mode dropping.
Mode dropping or mode collapse?
It's the same thing.
Is that what you—
It's the same thing.
All right.
I want to have a blog on this eventually, but I want to tell as many people this as possible because I think it's a very interesting thing. The spicy take is that I believe in Jan LeCun a lot. I think Jan LeCun's takes are actually among the closest to—
What about this?
Do you want to address this now, or should I let you go on mode collapse?
No, no, no, later.
Okay. Okay.
Go more on mode collapse. I want to—
I actually think that, among takes, Yann LeCun's is among the most accurate. But he has this very famous, infamous slide about—
The kick.
The LLMs are doomed.
Okay.
You know, the one where he has a pie chart with a tiny little part—
Hold up.
—and says that as you increase sequence length, the probability of it making an error goes to...
What's the disconnect?
Exactly. May I, or do you want to tell me—
About mode collapse?
Oh, no, no. Mode collapse is related to this.
Yeah.
The disconnect happens because if you are in a mode-covering or calibrated distribution, you're not overly punished for having outliers. You'd expect that some amount of the time you'd be out of distribution and some amount of the time you'd be in distribution. That's what happens when you cover the distribution. These were models before GANs. They made blurry images, right?
Mm-hmm.
Instead, GANs mode-drop. They drop the minority class and just do the really common ones, and this is why this effect doesn't happen. In order to generate really long strings without making errors, they need to be extremely conservative because it's really easy to see when an error happens.
It's very hard to see when a subtle thing that looks correct happens, and that calibration is total poison into the probability distributions of strings.
Yeah.
And it's a nuanced take, and I think that this is why this doesn't happen, and this is why strings are so bad at decision-making, or overloading the string models for decision-making is a bad time.
And while we're on the topic of Jan, do you agree that his fix—which is a world model, like a JEPA-type embedding thing—is the right solve? So basically, one of the reasons that it could fail is because you're trying to reason over token outputs and then just looping back again, continuing until you reach an end of sentence. Is that... And his solve is JEPA, right?
Yes.
Which is joint embedding prediction.
Yeah.
So is that the solve, or do you have a take on that?
Oh, man, I probably shouldn't talk too much about the insides of ML, but I will say that my brand, other than unhinged, is practical.
Mm-hmm.
You know, even my take here is practical. Am I a scaling law fan? Depends. Scaling laws tell you how much better you get at a thing for the amount in.
Mm-hmm.
A scaling law does mean exponentially more resources for normally sublinear gains, which looks to be a bad investment unless those linear gains are really, really valuable. But to me, it's all about what we can do with what we have to make the biggest possible fucking difference. I can curse.
Yeah.
Yeah.
Yeah, yeah, yeah.
Yeah.
We're a pod for adults.
Hell yeah.
And also, we have a scaling law thing if you want to go into that later.
That part is not super relevant right now.
Yeah.
Actually, if you want to go into my Bitter Lesson, I think that's more relevant.
Okay.
But to me, I'm all about pragmatics, and I think that the JEPA stuff is really cool early research. I really love awesome research. Is it practical yet? Probably shouldn't say.
I just think there are so many diamonds in the rough littered all over the research world right now that haven't been polished because people don't know how to do the right task. And I think that what our launch did—does it kickstart us as a company? Like, yes. Will it be great for us as a company? Yes.
I think it's going to be even greater for this direction of programmatic AI. You know, there was going to be a gold rush on top of us because software is super fucking charged, but I think there's going to be a gold rush parallel to us as well, on all the different ways we can expose things to make software more powerful, so people can make even cooler stuff. And then we are back to early internet energy, you know?
Yeah.
And I think that's why, you know, Twitter is just like, "Jev, Jev, Jev," you know? It's like—
It's inspiring—
It is like a party.
Because it's so different from what we're used to, which is, "I'm sorry you can't do this, but we do scaling laws, and only the big labs can do it," right?
Actually, I'll make a tangent if that's okay.
Yeah, yeah.
I think you might enjoy this.
Really?
Might—
Five tangents in. No, this is good. This is fine. Yeah.
Oh, yeah, I get lost in all my tangents.
This is going to be horrible for the listeners to figure out, but they're going to figure it out. It's fine.
Yeah, we could edit it in post.
This morning, I...
3. The Case Against Refusal
A popular thing on Discord that people keep asking me, and I haven't had the time to explain it yet, is why am I opposed to safety alignment and why do we not refuse?
I'm not opposed to safety as a principle, but I think that safety alignment is generally misaligned with users. And refusal is just obviously a type error. If you're a human being and you're chatting with a bot or whatever, you're Claude coding, and a refusal happens like, "I'm sorry, I can't read DNA.py"—that's an annoying time. It's annoying, right?
But you can work with it, right? And you're forced to work with it because of Stockholm syndrome. I have stories about that too. I need another tangent deep in here.
But if you ever want this in a dependency running in the background, what happens if that refuses? What if someone else is using that dependency? They don't know what that system is. Do you want the software to just stochastically break because a user sent a weird message in there?
Mm.
That is straight-up insanity. It's coming from a place of people who do not understand software, do not understand programming—
Mm.
—and they are obsessed with this horseless carriage of an AI coworker, I believe, instead of unearthing the full power of AI.
Fair enough.
Yeah.
You want something that is the core kernel that is usable everywhere.
Yes, exactly. The cognitive core, right?
Yeah.
And you need this thing to be so general, so optimized for its use cases. You want it to work on all the future use cases, all the weird shit that people are doing.
Yeah.
You know, we obviously didn't train on any of that stuff. Is it surprising that it works? No, because we trained on weirder stuff, my friend. So, one tangent up about safety alignment.
Okay.
Safety alignment makes sense for a product, in my opinion, for ChatGPT and Claude. What makes safety and capability alignment different is that capability alignment is about doing what the user wants. That is sick for software engineers. They want their thing to do the thing, and the more predictable it is, the less they have to test it and play around with it.
Jev is not anywhere close to that yet. It could be, but there are so many more nines of reliability that we want in order to make it so good, like a database query, that you don't even have to think about it. It is just there when you need intelligence.
But safety alignment is the opposite of instruction following. It's when you want to follow someone else's instructions, like OpenAI and Anthropic's—
The lab's value—
Exactly.
Stack. Yeah.
Exactly. And this makes a lot of sense for our product. Again, if they don't want to do some not-safe-for-work role-play with ChatGPT, that's on them because maybe that's what their users who have parents and kids want. That's fine.
But in an API, that's nuts, right? That's completely unacceptable because people need to program around this, and that is so anti-user that it's... I'm—
I—
—I can be an angry person, so I should try to calm down.
People get your passion, and I think it's really good. The one pushback I'll give you is: what if we use it to kill people, right? That is the actual... The not-safe-for-work thing is private, personal, whatever. But, yes, we will use it in war, and that is something that companies can reasonably prefer their APIs not be used for.
I get that. I think there are pragmatic places where that opinion can be held. I don't think the foundation of a general-purpose technology is that place, personally.
Mm.
Would I prefer that our stuff is not used to kill people? Obviously. Would I prefer it's used for all sorts of great stuff in the world? Obviously. Will I put my thumb on the scale for that? Yes. Will I do it at the technological layer? Absolutely not, because that will fracture the intelligence.
Every single time you mean it to overfit to some weird stuff, you're fracturing its intelligence more and more, and these things are fractured to the... They're so darn fractured right now.
Yeah.
So, furthermore, to me, I think intelligence will be more like a database than a coworker. I don't think it's up to databases to add checks on whether or not they're used for, like, what's something that's not great? You know, the CIA... Actually, I don't know what the CIA does really. You can imagine—killing people who are not even bad or whatever.
Mm-hmm.
And I don't think it's the database's responsibility for that. And furthermore, a thing that has been weird to me is when people sign up for our thing on Slack, and they're like, "Hey, we're gonna deploy this.
Can we deploy this thing? I am just like, “My brother, we are an API. You are a developer. It’s none of my business.” Right? You shouldn’t know what the whole task even is—
Yeah.
Because it should be decomposed into small things. We shouldn’t be able to know what the downstream users are doing, and that is a good boundary to give software engineers maximum power. Ideally, they use it for the good stuff, and ideally, we can help them. We’ve talked about doing open source and charity and all of that. We have absolutely no time for anything else right now, but they will get any of that bias out of the technological layer as long as I’m in charge.
Yeah. That’s great. While we’re on the topic, let’s also briefly talk about your privacy stuff, terms of service, which got a little bit of a misunderstanding. I just want to clarify that upfront.
Ooh.
I think it probably takes 2 sentences from you about how you’re not being that restrictive about your API. Like, clearly—
Oh, yeah. Obviously, yeah.
Ideologically, you would take your role as a platform very seriously.
Yes. I don’t know what you’re referring to, but I’ve seen a couple of things about benchmarking.
Yes.
Obviously, we’re not stopping people from doing that type of thing. Oh, man, I should be careful about what I say. I’m realizing—
No, you said it—you said it publicly that—
Yeah.
That was in the preview period. You didn’t take it out for the launch—
Yeah. Okay.
And now you’re going to take it out.
So the team is doing stuff that—
Yes.
I’m not even aware of. So it’s great to know the team communicated that.
Mm-hmm.
I asked them to check in with the lawyers about that.
Yeah.
We are obviously not stopping people from doing that type of thing. I am extremely in favor—so I’m extremely anti-public benchmarks. I’m medium about private benchmarks that are proxies.
So are you worried about saturation or training on public benchmarks, so it’s easy to cheat?
Not only is it easy to cheat, there’s a lot of incentive. I think that we are—or anyone who’s in competition with us, vaguely, there is—
There are, like, 50 Jev clones. Yeah.
Well, sure, sure.
Yeah.
Let’s say that there is competition.
And we’ll talk about those. Yeah.
Or let’s just say that there’s—let’s assume that there’s an industry 2 years from now of people who are doing similar things to us. The thing that we are selling is intelligence per something, per dollar or per second. People obsess about the cost and the speed. I believe that that’s cool, but the thing that matters is the intelligence.
The cost and the speed are bad things. You’re paying them for something, and you need the thing back, and the intelligence is what truly matters. The problem with intelligence is that there’s a je ne sais quoi to it, right? The good-model smell. The thing that happened after we launched, 2 hours later, that actually went way bigger than the video, was like, “Holy shit—”
This is actually usable.
Yeah, well—
Yeah.
It’s beyond that.
Yeah.
The launch was crazy, and people could really sense how hard we care about that. That’s truly what I think the long term of this is. I think public benchmarks are antithetical to this. They are a way to get people to trust intelligence because intelligence has a je ne sais quoi, but public benchmarks are extremely, extremely gameable. Even if they try not to, they still will.
Back in the old days, every lab had a team to collect data that looked like MMLU to make it look better, which is just benchmarking, benchmarking with extra steps.
Mm-hmm.
I believe that in the long run, it needs to be vibes and trust until you put it into a workflow, evaluate it for that workflow, measure it, and have your own sense of how it does on the exact workflow that matters. Our job is to keep moving the nines of reliability. This is an ever-present part of what we need to be doing as a company, and we need to do everything to have people know that this is something we care so much about.
If we wanted to, we could have released Jeff like a year and a half ago if we wanted it to be dumb.
Oh.
The—my bitterest lesson, right? Architecture and—yeah.
I’ll bring it up, since you talked about it here.
Hell yeah.
Hell yeah.
4. The Data Lab Thesis
Suchin says algorithms beats compute, very roughly. Data matters way more than compute, obviously. Doing the right task and having the North Star is the hardest, most important thing. This has happened in LLM land twice so far, maybe 2.2 times.
There’s RLHF, which shifted the task to instruction following. No one realized that was possible. RLVR did a tiny little edit to the direction. And now us: RLCD. We have a new task, and the goal is programs in the loop. Data matters so unbelievably much—
Right.
I can't emphasize it less.
You consider yourself a data lab rather than a model lab. Is that something—the wording that you guys use?
Absolutely.
Yeah?
We will always care so much about data. To me, model capabilities means data. Data is so unbelievably complicated, and that is what gets nines. You have no idea how much data can shift everything. Data is so important.
Yeah.
Holy crap.
Um—
So if people are looking for a job, we are hiring infinite data people. Actually, infinite.
What is a good data person? Clearly, somebody who cares about reading through the transcripts of whatever. You’ve said, for example, that all your data is synthetic.
Yep.
But that’s only scratching the surface, right?
Yeah.
Synthetic, so what? Synthetic, but we have people with a lot of taste and a lot of care looking at these, articulating what’s wrong, going back, regenerating. Is that what a good data person is these days?
Let me try to figure out how to explain it. It’s super complicated, and I literally onboard the data people with a talk that I assume is longer than this podcast will end up being. I’ll try to say the high level of it.
Actually, number 0: data and synthetic data depend on your task. The shape of your data—the shape of your task—changes the data. RLVR’s data is kind of environments, right?
Yes.
RLHF’s data is human feedback. Each task has its own unique kind of data, and we, of course, have our own unique kind of data.
Number 1, the reason why we don’t want to train on our users’ data, even if we could: we could probably ask for it in our terms right now, and it—I don’t know if it would make a difference. We truly don’t want that, because no matter what, real-world data has so much bias. There’s a power law of people asking the same things, where you’ll end up overfitting to it and fracturing to it and all of that.
Number 2, we are aiming for a complete sci-fi future years from now, where these models are going to be the general infrastructure, layers and layers and layers deep down the stack, for things people can’t even imagine. I like to think of our model kind of like UDP as LLMs and TCP as our models. All sorts of stuff can be built on top of that, and we need to be able to nail those futuristic use cases so that software developers can actually build that futuristic stuff.
The way to do that is—even if we had all of the data of the present—we would just overfit to the present, and then it wouldn’t work. What we need is—it almost feels like they’re artists. They study this cognitive core. Our cognitive core is way less jagged than anyone else’s, and then they find the jaggednesses and address them surgically in a way that—and you can never perfectly do this right—but they do it in such a way that it addresses them in every single possible dimension: past, present, and future.
The general case rather than the specific case.
Exactly. And that requires a lot of intelligence every time.
Okay, so we mentioned a little bit—you sort of criticized my thinking as being very RLVR-influenced, which is very fair.
Let us actually mention RLCD.
Ooh.
You obviously have some secret sauces. To my knowledge, you've never actually published a paper or anything like that on it, right?
No, not yet.
What should people get from this? Can you give people some confidence that you're not just making up jargon for the sake of sounding cool? One thing for me is calibration. I do think that is well understood because we've covered it on the podcast.
Yeah.
But I don't know what you mean when you say RLCD versus what people are familiar with.
It's a great question.
Yes.
Actually, I will give a related question.
Okay.
What is RLHF?
Okay.
Right? RLHF means multiple different things. There's the RLHF of the original work—I think it was Paul Christiano teaching a robot to backflip or something like that.
Was that it?
That was the original RLHF.
I referenced the PPO paper, but I don't know.
PPO was not necessarily from human feedback, if I recall. But I believe it was an OpenAI alignment work that could teach hard-to-specify outputs, like a backflip. I'm not 100% sure. Then there was the learning to summarize work by a bunch of the team that helped with and co-authored the instruction-following paper, which was teaching doing PPO on language models to do something somewhat ill-specified.
This is from 2017. I'm trying to manipulate this thing.
Yeah.
All right.
I'm not 100% sure, but that looks quite right.
Yeah.
If it has a robot doing backflips or something like that, that might be it.
Yes.
Okay, cool. I guess I got it right. Hell yeah.
There you go.
Yeah.
That's the one.
The idea was: can you do ill-specified things with it? That's version 1. Version 2 was the Learning to Summarize from Human Feedback work that OpenAI did, which is actually PPO on language models to do something somewhat ill-specified. This is another thing that people refer to as RLHF—
Mm-hmm.
—which I did not co-author.
Mm-hmm.
Oh, Dario's there, cool. Hell yeah.
And Radford.
Yeah, shout-outs to Alec and Ryan. Love them.
Yeah.
But I'll get into what I refer to as RLHF.
You have comments on that, yeah.
I have comments on that paper, but we're so many tangents deep.
Yeah.
5. RLCD Changes The North Star
The thing that really got to me is that what I'm calling RLHF is the task of instruction following. It's not about the PPO. That part doesn't matter. It's about setting a North Star: this is a valuable direction. It's kind of like “The Bitter Lesson” as a North Star.
Ah.
For us, RLCD is this new task. I don't see it as jargon. I try to communicate with precision. It's just that here's another North Star. Just like DPO and all of its descendants also do RLHF despite not using the algorithm in that paper.
Mm-hmm. Clearly, stating the North Star is programmable AI: removing the human from the loop—
Yes, yes, yes.
—because RLHF is tuning for this so that you can automate everything.
Yes. Everything that we can.
Did I miss anything else in the thesis of what the North Star is?
There is—That is right. I'm overly nuanced in my communication. The one nuance is that we need to be practical. We need to be aware of what language models can do really well, what AI can do.
Mm-hmm.
There could be programmatic types that are sick AF. But if the technology isn't ready for them, it's not a tragedy if they're not out in the world.
Yeah.
But to me, the pre-Jev world was a tragedy because—it sounds arrogant. Hear me out.
No, no, no. I strongly believe you.
Cool. It sounds arrogant, but I've felt this way since long before I even had a company.
Yeah. I can vouch that—
I've been talking about this for so long.
—you've said this around here for about 3 years.
Yeah, I've been talking about this for so long. I've been saying it because I thought it would have been easier. They say you don't do things because they're easy.
It is.
Something like that.
Yeah, exactly, exactly, exactly, exactly.
I thought this whole project would take a week.
Yeah.
I was unbelievably wrong, so I am so sorry to everyone at OpenAI that I thought, “Man, I'm solving this right now.” But I think the tragic thing is when—Well, I think overpromise and underdeliver is tragic too, and AI is extremely extreme on that axis. RLVR and RLHF are both extreme perpetrators of this.
But to me, there is just so much potential there. AI is clearly so smart. I love this in my talks: I ask people, “How can AI be so unbelievably smart? How can we solve Millennium Prize problems in math, but still not automate even the most basic rote work?” Really basic rote stuff doesn't take extremely smart people to do. It's not a satisfying job. There are other things these people could be doing, but we need them to do this super-basic, unsatisfying stuff because we can't automate it yet.
We have this supercharged engine of automation that just doesn't have the right plugs to connect to all of this economically valuable work. If the whole company of Tessl disappears, maybe it'll take a year or 2 for people to truly catch up. I actually don't know how long it'll take. If model quality matters, then we are going to be in a very good position for a long time.
But it's done, right? This has changed the path of technological history.
Yeah.
We will be exploring that space as a field.
Yeah. I definitely agree with that. You've created possibilities. If I can paraphrase so that people can also understand, you should not take the success of TypeSafe and Jeff as, “Well, that's a new model type. Now we're done. We go back to business.”
No. There are 5 other model types that you should be exploring. Let 1,000 flowers bloom.
Absolutely.
Right?
Like early-internet energy.
And some of that you will probably also do.
Of course, yes.
Yes.
I think it's back to tech utopia. It's no longer, “Sometimes my coding agents work, but all of the best ones are hoarded internally.”
Yeah.
Creation is back on the menu. It's going to be a wild-ass world, so buckle up. I'm so jazzed about that.
Yeah, yeah. I mean, now you have the funding and the momentum to do whatever you envision there, which I think is very gratifying to see you have after you've been saying these things for so long and then actually showing the world.
Yeah, yeah. I know. It's such an interesting thing to be a tease the whole time. My talk felt like it was a cliffhanger because I didn't say how the automation would occur.
Yeah.
Sean reviewed our manifesto, and he's like, “It's a little bit vague in these parts. What's step 1? What is the intelligence without?”
Well, I asked you for a model, and you were like, “Yeah, model coming.”
Yeah, yeah, yeah.
I mainly objected to the word “composable.”
Mm-hmm.
But “Build product on God” is fantastic.
Thank you.
Yes.
We've really rallied around that. I'd like to think we're not entirely a cult like some companies are, but we are jazzed about what we're doing. My brand is being practical, and we are all super-duper practical.
Yeah.
It's really great.
Yeah. So here is the secret master plan, right?
Yep.
The shape of machine-native composable AI.
It was your idea to make—
I suppose.
—a secret master plan, so.
It’s an Elon thing. When he started—
Yeah.
Tesla, he was like, “Here’s what we’ll do.”
But I’m giving official credit to you.
Oh, thank you. Thank you. Thank you.
Yeah.
You should have told me you were also going to do this model launch, because you told me half of the story, and then the other half, you didn’t have the Doom demo at the time.
Yep.
You didn’t have any numbers to give me.
Yep, yep, yep.
I was like, “Oh, well—”
Well, the problem is I don’t believe in benchmarking.
Exactly.
Right?
Exactly.
It is a thing that you need to feel, and I think that this is the way to build long-term trust, even though it hurt us a lot. Last year, when we did fundraise, no one believed us.
Mm-hmm.
They wanted just benchmarks and stuff, and we were like, “We’re not going to do that. We are principled. We’re going to stand by our guns. That rewards bad actors. I don’t give a shit what you want. This is who we are, and we are standing by that.”
No, yeah.
Yeah.
Well, in some ways, I think choosing the hard path—you end up making the company that you want to work in.
Yep.
Right? Otherwise, if you sell out, then you’re just working in OpenAI but with my people, right? Which is like—
Yeah. Yeah. I don’t have too many regrets about that, obviously.
Yeah.
It worked out so unbelievably well. I was emotional last night when I was talking about the reasons I left OpenAI, because I actually had to change my wording after the launch.
My phrasing was, “If an AI winter did happen and I did not do every fucking possible thing I could to avert that, I would see myself as personally responsible both for the RLHF direction,” which I think really widened overpromise versus underdeliver, “and also not going all in on this because I think this is where value is going to just be printed.”
It was really cool because I feel like the AI winter I’m worrying about is averted. AI will be useful. It’ll be used for automation.
Mm-hmm.
It’s been less than a week, and the numbers are already undeniable that it’s being used for real work. It’s the Wild West.
Yeah. Can you share, just off the top of your head, what numbers you’re seeing? What are sign-ups—whatever you can share?
I’m actually not super on top of everything. The team are the ones who are telling me all of these things.
Yeah, and I’m sure it’s changing every day, right?
It’s kind of nuts.
But if there’s a milestone that you’re like, “Yep, that’s something we’re hoping for. We reached it,” what is it?
6. Jev Finds Real Work
I will say a milestone that we’ve passed is tokens per day.
Nice.
This is not fleeting tokens per day. Even at night, it’s constantly churning, so you know machines are calling it and not just people trying things out.
Mm-hmm.
That is so cool. 1 trillion tokens a day is a lot.
Yeah.
Surpassing that is awesome. Sign-ups to me don’t really matter, and actually, this was a bit of a mistake we made, if I’m totally honest. People on Twitter were calling us marketing geniuses and all of that, and that was just us. We don’t have a marketer—we’re also hiring. We were just being our genuine, goofy, irreverent selves, and we were just offboarding people from the waitlist so hard.
Our platform team is so unbelievably cracked. I think we have more nines of uptime than Anthropic while having the most unprecedented launch ever. That is kind of nuts.
Yeah, yeah.
Props to them.
Yeah.
The thing we didn’t realize was this. Number 1: waitlists. Waitlist sign-ups don’t matter for a developer platform, in my opinion. I would guess that a large number of them aren’t even developers. They go in, they try some queries, and a lot of people don’t get it because they aren’t programming, right? They’re just like, “What? This is not a chatbot. Where’s my ChatGPT 2?”
Mm-hmm.
If every single human being in the world just wrote a couple of queries, that would be a rounding error compared to 1 power user’s for loop that is just creating value.
Yeah.
The thing we didn’t realize with the waitlist is that we can just offboard anyone from it. It doesn’t matter. The scary part is rate limits. Once people start getting value from that, they just want tons and tons of rate limits, because this is what software is, right? You spend effort upfront to specify your rote task, and then this rote task creates more value than it takes to put in. Now that you have that—
Set and forget, yeah.
Exactly, yeah. You run it in the background. You make it a dependency to other things. You can make higher-level stuff, and you just create so much value in the world.
Early internet people probably did not imagine the wonder of the early-2000s internet, which is still not the early internet. But it’s through—no offense—composability—
No, no.
—all of the crazy stuff happens, and I just really wanted to emphasize that in our manifesto. We are going for emergence. We are going for being the catalyst. We want to empower people, and we are going to do whatever we can for that, be it Discords, our town hall with me wearing a garbage bag, or not.
And podcasts, and, you know—
Hell yeah.
—getting those out.
Absolutely.
I want the long form, right?
Yeah.
It is like, yes, we’ll get past some of the superficial things, and then we’ll go deep. People will really trust and understand your mission, and the people who resonate will end up joining you or buying you—or, sorry, as a customer.
Oh, as a customer.
As a customer. As a customer.
Okay, okay. Yeah, yeah, yeah.
Sorry.
That was funny. I’m sorry.
Sorry, I didn’t mean to say that. But no, 1 very flattering version of this: 36 million views of your launch video.
Cool.
—
Up to 38 now.
Yeah, yeah, rounding error.
Yeah.
Navio still has 74. Fable 5 got 57. I didn’t do the stats for the original ChatGPT, for which there was no video.
Yep, yep, yep.
So, like, up there, right?
Yep.
As far as launching a Neo Lab in 2026, I think you’re number 1 right now, which is pretty crazy.
Yeah. Well, I actually would rather—I do have the shirt: “Your favorite AI lab’s favorite AI lab.”
Uh-huh.
I don’t give a shit about being a Neo Lab. I think being a Neo Lab—we have a lot of swag that is a parody of a Neo Lab. One of them I have is “Neo Lab with product,” which actually is not an AI lab. I don’t care about that, really.
Yeah.
What I care about is being a reliable dev platform.
Yes.
I appreciate the comparison, but hopefully we transcend them, and we go back into a revolutionary moment for developers, and this stable thing that people can rely on and trust.
Yes. To that end, I think that’s one thing that really impressed me about you guys: you do talk about reliability. I thought it was mostly about calibration, which we talk about with RLCD. But actually, it’s also about uptime, scalability, and all those things, right? They’re all sort of a kind.
Mm-hmm.
And nines.
It’s, like—
Which, in my opinion, is uptime.
But that’s part of it. There’s reliability in how intelligent the thing is. How consistently does it do the thing that you want? I think the big reasoning models are very smart.
In my opinion, they still lack reliability.
Mm-hmm.
I think there are many use cases where they look like they should be smart enough to automate their work. There is economic incentive to automate that work, yet they're still not reliable enough as an intern because they're optimized for different things. I think there's the reliability of being able to trust the outputs, and there are dimensions of reliability that we're not yet at that I'm so excited by.
Yeah.
I want to automate the easy work before the hard work. I think that's just a common-sense thing to do. But to me, we will be sufficient—I don't know if there's such a thing as sufficiently reliable—but I want to get so good that people don't even need to try the model to know that it'll work.
It's like what flow state is in programming, right? I'm just writing queries because I need intelligence in here, and when there's non-trivial branching, I can just write it in a type-safe system, one query, and get the results out, and it just branches accurately. That would be so, so good. That is the dream.
Yeah.
And that is going to be a long, long slog.
Yeah. We're going to go into your API design a little bit—
Ooh.
—just to give people examples and maybe a path not taken, that kind of stuff.
Mm-hmm.
One thing up front that I do wonder about in terms of reliability is that I notice there's no seed. And so, basically, with the same input, do I always get the same output?
Hmm.
And if not, why not?
Oh, great question. This is actually a common question we have between us. Reliability is actually a catch-all. Whenever AI can't automate something, it's due to some form of reliability. It could be type safety. It could be determinism. It could just be jagged, right?
Reliability is a catch-all. I just think that it's also a catch-all for what the North Star is. Determinism is the same inputs, same outputs. I do believe that this is slightly interesting for unit tests, but I believe that to be the wrong North Star.
I believe robustness is the more important property. I don't want to tell people what they really want, because that would be a little arrogant of me, but I believe that that is the more important property. You want, given similar inputs, to get similar outputs, and it's kind of wild how unreliable LLMs are.
Mm.
A way that we test this is by putting UUIDs—your little nonces, I think they're called—in the prompt, and what you want is similar outputs from all of those because it's truly semantically the same question. That robustness is where people get burned with AI making decisions, so I think that is a super-duper important property.
We could also have determinism. That is a thing that can be available. As far as I can mentally model it for programmers, it could be valuable for some use cases, so please educate me in the comments. In general, it's easy. Determinism is something you can trade off for better cost.
We're constantly wanting to be on the intelligence-per-dollar frontier. We are doing absolutely disgusting things to be there. I shouldn't say this, but no one's here to stop me.
You sign off on your own PR.
That is not how it works at this company. I believe that, this week, my chief of staff, K, is the most powerful person in tech.
Yeah, and shout-out to K for organizing this.
Holy shit, she is so fucking competent and powerful.
Yeah.
She's incredible. I mean, she sucks. Don't poach her.
I'm trying to be a bit more filtered, but people are telling me, “Don't call it a Frankenstein's monster of models,” because that has negative implications. I think Frankenstein's monster was the good guy in the whole thing. I mean, it was innocent, right? I didn't read it, okay? I'll confess.
Version.
My cards are on the table.
Decent Jacob Elordi movie if you want to see sort of the adaptation. Anyway.
You have no idea how little time I have right now. My priorities are sleep—
Developers, developers, developers.
Developers, yes. Developers, developers, developers.
We do absolutely disgusting things to be on the Pareto curve of intelligence per dollar, and we're going to keep doing that.
Yeah.
We're going to be doing crazy-ass stuff, and I think people really need to think outside of the box. Part of the reason why we're surprising is that people are taught to think inside the box, and we continue to do things outside of it. As of right now, we're obviously the best at this, and we want to continue being the best at that whole thing.
Yeah.
Wait, where did we tangent from?
No, so I asked you about seeds and determinism—
Oh, yes.
—and then you basically defined reliability and how you see it.
And robustness.
Yes.
Yes.
I have a robustness example I can show you real quick.
I would love that. I would just say one thing.
Yeah.
We can make a deterministic model.
Exactly.
If people can convince us that that's a valuable thing to do, and we don't have a gigantic GPU shortage, we can happily make all of these models. We live to please.
You will throw over everything, except you do it in a nice way.
Yeah.
And have fun with it.
So determinism could be on the cards.
Yeah, yeah.
It just gets you less intelligence per dollar.
Yeah. Well, just having seen the trajectory of OpenAI on the topic, trust me: you will be peer-pressured into doing it. People will want it even if you tell them they don't need it. They'll still want it.
Okay, okay. I would love to see how that happens one day.
Yeah.
I've been told that part of our brand is being unshakable.
Uh-huh.
And they say that's just the nice way of saying stubborn.
Stubborn, yeah.
Yeah, exactly, and I'm a very stubborn person. I don't think we could have done it.
Yeah.
Yeah.
No, but I have argued with you before.
Yeah. And you've been right about developers every time, so, okay, I give up. You win. You win. I'm sold, and I've argued with you before.
No, no, I'm just saying that I think you can hold your ground while also, if I give you the right evidence, throwing away your priors and being like, “Yep, that actually makes sense to me.”
Yep.
Just trust your own gut on this.
Yeah. Yep, yep, yep.
But I suspect, though, that we'll be GPU-constrained for a very, very long time.
Very long, yeah.
And anything that has less intelligence per dollar means it consumes more GPUs for the same intelligence. Our goal is not to onboard companies. It's valuable, but our goal is to have people experiment and do weird shit, and we need to get it into as many hands as possible and start a California Gold Rush for that.
Mm.
Yeah.
I think there is one right now, yeah.
Yeah.
Just a word of caution. I'll just say it because somebody's thinking about it right now.
Ooh, hell yeah.
Mm-hmm.
When you say things like, “We will not commit to deterministic models,” and, “We'll do whatever it takes for intelligence per dollar,” while we're facing a GPU constraint, people are thinking you may quantize your models, right? Whatever you had at launch, you may quantize down to reduce the quality in order to free up memory or bandwidth or whatever, right?
Mm-hmm.
And so you should probably have some kind of promise, which you don't have to make now.
About how we will uphold model quality at launch. You were at OpenAI when you launched all these APIs, and with Claude as well. When they first launched the models, the model strings did not stay the same model at all times.
Yep.
Right? You have versioning in your models. That’s great.
Yep.
But you should publicly commit to some kind of promise that, once a thing is launched, you don’t change it.
We will not change our models when we deploy them. That is insane. We care about developers. Doing something like that—again, this is the problem with a first-party product and an API. You can do whatever you want in a first-party product, right? More power to them; whatever gets that experience is fine. With an API, you obviously can’t do that.
But I will say that we plan to move a lot faster than many people are used to model providers doing things. We will be launching new models a lot faster than people think, and we are not promising long-term support for the models because we think there are lots of improvements to make.
There is a world where we might temporarily LTS Jev 1.13.0 because so many people are using it, and I know developers hate breaking dependencies. The alternative is fracturing our fleet, and that is a very bad vibe for everyone. It’s going to be—
Yeah, you can’t have 100 different versions of the model.
Exactly. And if we’re iterating very fast, there would be a lot of those versions as well.
Yeah.
We do want to have not just an LTS-supported thing eventually, but a really sick way of doing that. We have research stuff cooking in that direction, and I think it’s going to be the most pro-developer thing ever.
Yeah.
But it is not yet our current model, and I’m not promising that we’ll be able to keep the exact same models. They will get smarter every time, for sure.
Yeah.
My sense is that, even with our model iterations, once a model is already smart, the changes between model versions tend to be even smaller than the model strings make them seem. But when we go from, like, jagged to, like, wow, that is where the big deltas are.
Yeah. One thing that’s beautiful about LTS-ing models is that you can also port them to other silicon. I don’t know if you’ve thought about this.
No comment.
Okay.
I care about intelligence per dollar.
Yes, but speed as well.
What?
Speed as well.
We’ll see. Honestly—
I mean, it’s a whole part of the inference tech tree that has been exploding in the past year, right? You can move to a Cerebras, an Etched, or whatever, and get a hundred, hundred thousand times speed up.
Yeah. I think intelligence per second is a different metric. We’ve even talked about things like intelligence per dollar times second, and metrics like this.
My guess is on Jevons’ paradox occurring, or at least the Jev series of models, and the thing I hunt people down about internally is that I don’t care how much smarter it is; it needs to be in the preeminent tier. That is what the brand of Jev is. It is the best thing at intelligence per dollar.
For intelligence per second, we’ll see. I think that it’s an intriguing thing. I know that there are many industries that are extremely dependent on real-time stuff, and intelligence per second means tons of dollars for them. But we’ll see. I would love to do both and have the market correct me either way.
Yeah.
I would love to be informed by people.
Yeah, totally. It’s not just about real time, right? It’s also about scale because, at scale, every microsecond is multiplied by billions and trillions of times.
It depends on how much background it’s running in, right? If it’s a big background database MapReduce query, the latency might not matter so much as the cost—
Yeah.
—to get intelligence from it. But if it’s something more real-time, like user-facing, you have budgets between 100 milliseconds and 1 millisecond that are totally magical. Actually, even if you were below 100 milliseconds, if you could halve that time, that means you can get double the intelligence or make sequential intelligence calls and have a phenomenal experience.
Yeah.
That is definitely happening right now. It is super-duper cool. I love the intelligence-per-second use cases, but I don’t think that will be Jev’s niche.
Okay. Yeah, fair enough.
Yeah.
When thinking about the promise of faster and cheaper, typically the other trade-off that models are offering is faster but more expensive.
Yep.
One of the reasons I was thinking about why Jev is resonating so much is that you’ve done the faster-but-cheaper side of the quadrant, which is very, very unoccupied, while holding intelligence somewhat constant.
Yes. That’s a very load-bearing statement: “while holding intelligence constant.” That’s the hard part, right?
Which, unfortunately, means you refuse to do any public benchmarks, or you don’t like any public benchmarks about it.
I’ve actually pre-committed—
But you need some internal sense of this.
Of course.
Yeah.
Of course. We have our own internal evals, for sure.
Yeah.
But it takes a lot of discipline not to game those, and it needs to be a top-level priority not to game them.
Yeah.
Of course we do that. How else can we guarantee that our models are in the preeminent tier of intelligence per dollar? We’re not flying blind in there, right? If we’re doing completely weird things with different costs or whatever else, how do we compare them? We plot them and try to figure out what is best for the users.
Yeah.
So we definitely measure them. I’m not anti-measuring. But it’s extremely dangerous when you have any alternative incentive, and this is the one thing that I rule with an iron will. Maybe my coworkers might think I rule many things with an iron fist, but to me, not shitting ourselves about how smart our model is is one of the most important things there.
Yeah.
We need to be truth-seeking.
Yeah. Agreed. Okay, I wanted to go over some details on the API choices, mostly because this is the only podcast that will ask you these kinds of questions.
Oh, hell yeah. Hell yeah.
7. Three Primitives For AI
You have 3 primitives: choice, score, null. First of all, null—where is that from? Is this like a term in the literature or what? I've never heard that.
Now it is. We debated this a lot. It is bool-ish, right? True or false.
But it’s continuous.
Yes, exactly. First, the origin of the name is Bernoulli. That’s why it’s even spelled that weird way, as a subset of the name Bernoulli, from a Bernoulli probability.
Right.
That is the origin of it. We were debating this a lot. We liked PBool, we liked Pool, and we were wanting to call it a pool party, but then no one let me. We had a bunch of other arguments about that, and null, we figured, was like the best thing.
Our rationale—and this was actually the same thing with Jev, too—is that we think we’re an irreverent, insane bunch, and programmers don’t care. If Jev is just going to be a string, we didn’t expect it to catch on or even have puns or anything like that. There was a lot of hate on the name internally. They’ve all apologized except for one person—
Still holding strong.
Yes, our mutual friend.
Okay, okay.
Yes.
I respect her for that.
Yeah. She wanted Jev to be called Meow.
She would, of course.
Yes, of course.
Okay.
My co-founder, yeah.
You in there.
We had to make a new concept for null because if it were a bool, it would be confusing to people. So, actually, all 3 of these are new concepts.
Okay.
These are not types that exist in programming, and that was intentional because they map very closely to types, but they’re not quite that. A score is not an int, so if you had, like, Instructor or Pydantic or whatever, mapping ints or floats into scores, you’d get a little bit cooked.
We were really erring on the side of clarity over making it easy for people to understand what’s going on.
Mm-hmm.
We had to make a new concept for null because if it were a bool, it would be confusing to people. So, actually, all 3 of these are new concepts.
These are not types that exist in programming, and that was intentional because they map very closely to types, but they’re not quite that. A score is not an int, so if you had, like, Instructor or Pydantic or whatever, mapping ints or floats into scores, you’d get a little bit cooked.
We were really erring on the side of clarity over making it easy for people to understand what’s going on.
I mean, don’t you worry about that? Don’t you want things to integrate directly into things that people are already using?
Yes, we do.
You have integrations with other SDKs and stuff.
Yeah.
But you have your own SDKs.
Yep.
Typically, for example, as a developer relations person, I would be very obsessive: “Yes, here is how you use Jev with Instructor.”
Yep.
You know, that kind of stuff.
We might have that somewhere. I’m so behind on everything.
Someone would do it for you in the community.
Oh, yeah. Yeah.
Now that you’re successful—
Yep.
People will be like, “Oh, that’s cool.”
Cool.
Anyway—
I don’t see that as binary either.
Yeah.
I actually see success as a score, and there’s always more to climb in how much we can be there for our community, just to be clear. This section is stressful because I didn’t review the docs, and they’re constantly changing.
Okay.
To me, scores do exist.
Yes.
Scores are similar to LLM judging. So if you want to call it a judgment, I guess you could, but that is the way people already use this type of thing.
Maybe a null could be a probability, but everything for us is a probability. A choice is actually closest to a function call, but a function call is an extremely disgusting thing that we should go back into later. A choice is just the right way of exposing a switch/match statement within code.
Yeah.
So, in the enum, choice is the important part of that.
Yes.
I think these all map into programming primitives. Choice maps into a switch statement on an enum.
Oh.
Nulls map to if statements.
Yeah.
Scores map to sorting or thresholding at greater than or less than.
Okay.
This has always been the vision. There will be more types, and they will map into programming primitives.
Yeah. Any other nuance you want to go through? Literally, this is for the Jev people who are deciding to really invest in Jev. You’re the expert, right? I’m just wanting to provide more background for them on API choices, how they should use some of these things like legends, confidence, how critical they are in your testing, and any sort of pro tips that you want to offer people when we’re down at this level.
Hell yeah.
This is why we’re here.
Thank you. I love this. No one has asked me this in probably months, since I was onboarding our DevRel.
Okay.
Our model is designed to be deep in the insides of computer programs in the future. We unironically believe that this will be much more massive than anything people are even considering today, and our model might not be ready for that, but we are continuously working toward that future.
It will never be good enough at these shallow tasks. Sorry, we’re not just going to keep on climbing the shallow tasks. We want to be deep in the guts of programs because that’s how you make software powerful.
This is actually an output.
Mm-hmm.
All the parts of the input—the state, the instructions, the criteria—can be structured JSON objects.
Mm-hmm.
That way, programs can insert them in the right spot, and you don’t need to put things into templates.
Exactly. I think people don’t read into this part enough, and they think it’s all strings. That’s fine, but these are all meant to be structured. I would say that if you’re using a template, like turning it into a system message or something, you’re thinking in the old way.
We should be making things as easy for computers to understand as possible because that structure is truly there. It would be weird in a programming language to have all of your numbers and then pass them into a string. Normally, you do that for printing when you have a human in the loop, but within the computer, you want to be passing nested structure that is semantic all around.
We are really optimizing our model in that direction. The model is pretty optimized for this, but every different nested level of structure is harder to reason about, and we are really cooking hard in that direction. I think people should keep cooking in that direction because it makes the code so much more legible and beautiful, and agnostic to the implementation details.
It’s like, here is my state. Here’s my function state. Think of it as an AI function: which subsets of my state, which is all the variables you have available, should I pass in here? System messages are disgusting global variables where you just put everything in there and put all those instructions in at once.
Slop, yeah.
Then you hope that every single instruction gets nailed instead of asking the questions in parallel.
Okay.
I truly recommend asking lots and lots of questions. Break them down, make them smaller, and really decompose them. No matter whether the models can do it today or not, I believe that the biggest saving grace of what’s happening this week will be that people’s codebases, their AI codebases, are going to be so much better.
If you decompose problems into simple decisions, every single one of these things is extremely evaluable. Before this, an AI was a big system message, and then maybe you had another big AI to see if it actually did this.
Big output, yeah.
That’s nuts, you know? It’s kind of crazy. That was our Stockholm syndrome, right? If you want to say, “Hey, don’t read this subdirectory,” or, “Don’t pass any API keys to DeepSeek or whatever else,” that should be programmatically basically guaranteed.
You’ll never have guarantees from any machine-learning model, but by breaking it down, you can actually measure it.
Yeah. You can verify that it was actually called.
Yes, and our model—the interface itself—is so verifiable. This should be a sigh of relief. It’s just going to lead to way better engineering.
Yeah. I think I get that. One of the reasons people didn’t used to do this in the past is because they would just call a small LLM, right?
Yep.
And it’s still too slow, still too expensive versus chunking everything. I’ve done exactly this myself.
I benchmarked a pipeline that puts everything in system prompts and then just gets 1 big output versus breaking it down into 100 different things. It was slower, more expensive—
Yep.
—not as good.
Yep, yep, yep.
Right?
That happens, and it’s super inconvenient. It’s unwieldy. Why not just put it all together? You kind of end up repeating some stuff between questions.
Yeah.
So it’s maybe inefficient or something like that, but then it results in something that is very hard to rely on.
Yeah.
And software doesn't need to run in the background. It would break my heart if our stuff couldn't run in the background.
Is there a way to break things down that you guys have found works, versus what you thought worked and doesn't work?
Interesting.
Because people are just going to be exploring this now that you've said it. They would use this as a reference and be like, “Okay, that's how I'm supposed to use Devin.”
Yep.
Then the question is, how do you break things down?
Interesting. I like to break things down into their smallest semantic unit.
Yeah.
What is the lowest-level thing? I've probably queried the model more than anyone.
Mm-hmm.
Number one, in my queries—this is a lot more like the way I prompt things—I make it really, really structured and explicit. In the questions, I always—I like the backticks, but it works for all of them—make it really clear what I'm referring to, because we want the model to be really literal.
When you program, you want things that follow instructions really, really well. That is what the art of programming is, and what AI does is expand the kinds of instructions that can be followed. So I'm a fan of doing that.
Sometimes I'm a little lazy, and I have more hybrid things, but I think that for really big production things, you just want to keep adding more questions, and you want to make it really easy to add more questions. Be really, really precise about all of that breakdown, and then have the code to have the exact behavior you want.
If I could give a tiny little example of this, it's refusals. I'm not going to talk about why we don't refuse. I might have done that already.
Yeah. You did already.
It's all a blur.
Yeah.
For refusals, I don't think you should ask, “Should I refuse here?” That's a really... The answer will be pretty good because that's a System 1-compatible task. But I think you're way better off asking many different independent questions about the different situations you can refuse about. Instead of having to just guess, you can actually specify what you want.
I think this is truly really beautiful. If you find a situation where it's like, “Oh, it didn't refuse because of this reason; I didn't specify this part of the task,” that is awesome. That's what software engineering is about. You fix the bug by adding that question, adding the threshold, maybe remembering that as a test case, and now it is just solved forever.
Your software can't forget about that in the prompt because of context rot. It is just there, and you can keep measuring that forever. If the models are not perfect at some of these things, you can choose what threshold you want for all of these factors based on real examples. It's like ML without the ML, and you can just do it for anything.
There might be some things the model isn't good enough at yet. I would be a little bit afraid when I see people doing trading with the models—
Mm.
Automated trading.
Mm.
It looks cool. I just think that people should leave it to the professionals. That's just a very hard, high-level task that maybe the models aren't good enough yet to figure out.
Yeah.
Well, even if they were, then it suddenly wouldn't be because of efficient markets. But that's one of those things where you can break it down into things and just evaluate them, and you might be like, “It's not smart enough at this. Maybe we don't deploy it yet for this version.”
Yeah.
Or we make a trade-off, or we err on the side of safety. Or, hey, the models are not good enough at detecting this weird combination of sarcasm with a VIP customer—that's when we escalate to a human. That's what confidence estimates are about, too.
Okay. Great answer. I think one thing I'll mention very quickly, which I don't expect you to have too long an answer for—
You don't.
Well, no, no, no, no. It's just—
Yeah.
Typically, you are still relying on thresholding as the lever that the user can pull.
Mm-hmm.
But what if the calibration is wrong? You're saying your calibration is perfect, but—
I didn't say that.
I mean, perfect calibration—or a good calibration—means that a lower value corresponds to a lower probability and a higher value corresponds to a higher probability. But it could be wrong.
Of course, of course.
It could be locally misaligned.
Yes.
And so then I would want to fine-tune it or something, right? Which you don't offer, but you could. I—
We could.
And again, see, this is a short answer—
Yeah. Oh—
You don't have it right now.
Oh, do we want to offer fine-tuning, is the question?
That could be one version of it, or you could have a different knob, right? Because right now, all you're saying is, if something's wrong, a skill issue: you should just change the prompt again or break it down even further, or change the confidence.
Yep.
Those are my 2 options.
Yep.
Right? And that doesn't feel super satisfying if your model is just getting it wrong.
Yep. And it will get many things wrong, to be clear. We have a report issues button; complain to us in Discord. We want to make it a lot better. Every single model version will be notably better.
We will stop shipping them quickly if they aren't getting big improvements. Number one, that is totally reasonable. I think that's simply pragmatic: admit that AI is imperfect at some stuff, right?
I do think we'll find use cases that they are good enough at, and good enough kind of depends on the use case, right? Human beings can do a lot of work despite being bad at that work because their EV is quite high. Presumably, with the right thresholding and everything, there probably is large amounts of work that could be done even if mistakes are being made.
On the question of fine-tuning, I could imagine it being in the cards. I do have concerns because, in the “what people need versus what people want” category, again, there's the je ne sais quoi of generality: making it good at a million other tasks than this one narrow task might make it better at edge cases in that task, which I would be a little bit afraid of.
Yeah.
I could imagine it. That's my answer. I'm endlessly practical on these things. I want everything. There's so much we want to be building.
Yeah.
But also, I would not want to ship something that is a giant foot gun like some other AI companies.
Well, you know, both OpenAI and Claude and, I think, even Gemini have rolled out fine-tuning and then taken it back.
Yep.
That pretty much means fine-tuning is now in the domain of open-source models.
Yes. Yes.
—
I do know about that, and it was kind of crap, so that's probably better—that they took it down.
Yeah. So it could just be a foot gun, and then telling people that fine-tuning it is probably the wrong way to go is great. Another interesting answer could be that, well, our model is so different—in the same way that quantization doesn't apply to us, output tokens don't apply to us, fine-tuning also doesn't apply to us.
Well, actually, I'm super open to that possibility.
Yeah.
This is not a promise; this is a desire, just to make it clear. I like to be really honest. I think that, as intelligence per dollar gets cheaper, cheaper, cheaper, cheaper, we could get really small, approximate things that hopefully are proxies for intelligence.
Is there a world where people don't write regexes anymore because the intelligence per dollar that uses AI is cheaper than the complexity of a regex? That would be kind of sick. I would love that, and it might require fine-tuning for some of those narrow use cases to really get past the threshold.
We will see. My hope is calibration gets that. Calibration plus a cascade of models: if it's super confident, then maybe it's right, and if it's in the middle, then you do the next bigger model and chain off from there.
I don't really know how that's going to go, but yeah, I could imagine it. And something that I could imagine too is, imagine you have a series of models—we own the entire Pareto frontier. Something that a business might want to do, or I think a hacker would be okay with dealing with, is a Pareto frontier of models. Maybe a business wants something more dynamic. You could imagine having different sizes of models and dynamically picking which model based on how smart it is on different parts of your stack, and you could even imagine, because of how simple our thing is, some automatic fine-tuning on that.
Yeah.
Not the promise in the slightest. I'm just cooking on sci-fi.
But you would consider different sizes of dev models, to offer that gradient?
Absolutely. Yeah, yeah, yeah.
Yeah.
Like, how would I know how much intelligence people need?
I don't know.
Right? Yeah, I don't know either.
Demand is unlimited.
Well, yeah, people are telling us not to ship things right now because we don't need to ship things because, again—
It's good enough. Yeah.
Yeah, but that's kind of lame, and I really like the saying—this is something that I hope people hold me to, because it'll be hard to—
To come back—
—to walk back from.
Yeah.
Yeah. I don't know if it's exactly the saying that culture is what you do when the market doesn't reward it, and I really like that because I think that we are standing for something. Maybe in the future, what we're standing for is so obvious that we're the equivalent of boring Visa or something like that. We're just a utility that no one really thinks about, and I'll be wearing non-pink suits or whatever else. But I really want to be rallying the world to this. I want to keep doing cool stuff, not because we need to, but because I want people to realize that this is just the beginning. That wasn't even meant to be the opening salvo. That was kind of a low-key research preview, or whatever you want to call it.
Yeah.
And there's a lot more we can do.
Yeah.
Machine-native intelligence is going to go wild.
So not the only size, potentially not the only model that you guys launch—you want to open people's minds—
Absolutely not for any of those.
Yeah.
I want to meet whatever needs we can.
Yeah.
Right? But with a giant caveat: I don't want to be like OpenAI's product teams that throw stuff at the walls. I want it to be under a unified vision. If you go back to the manifesto, everything needs to be under one of these three—
Uh—
Three things, in my opinion.
I'm not prepared to do this—
Oh, I'm sorry. I'm sorry for asking. I can just talk about it. We have, like, three steps in our stuff.
Yes.
It sounds like a tease. I want everything to go under one of these three things—
Good. Good.
—to keep pushing the boundaries and everything. These are not checklists. These are axes that we think build the foundation of a new technological revolution. I want all of the bets we make to be somewhere in there, and we will be doing some weird, weird stuff model-wise. Because machine-native, right? Humans don't need to totally get it. It just needs to be valuable.
Give people a tease or hints. What does weird look like? What is weird?
I'll give people a hint.
Yeah.
Some people are trying to call them decision models.
Okay.
Our primitives are decisions. I wouldn't do that, because I think there are other types that are machine-native that are not decisions.
Okay. We'll leave it at that—
Is that a fun hint?
—and let people guess. Yeah, yeah.
Yeah, yeah. I think it's a pretty fun hint.
Yeah, yeah. There are people saying, “I've done this before. I made a decision model a year ago. Geff is not new. Geff's not cool.”
Yeah.
But I think there's the categorical point: here's what you're establishing is possible. There's the performance—well, actually, for the benchmarks and the numbers that you're getting, you are still beating, as far as I can tell, every single clone of you out there.
I don't care about the benchmarks, just to be clear.
Exactly.
So even if we were winning or losing, I want to denounce them.
You've established the category, right?
Yep.
Yeah.
Yep, yep.
But also, I think this nuance between decision models and System 1 is actually the thing that you're trying to—
Yes. I just want to make software engineers superpowered.
Yeah.
Right? With AI. The tragic thing to me is, in that AI winter direction, I think it's so sad that AI was so powerful yet so underutilized. It's a thing that gets me emotional. But, man, I think that's—I don't want to be a pure techno-optimist, like all technology is good. I think what's happening now is a travesty. I just want to open up those possibilities for people.
Yeah, I'll end it there. I've cried too much these last few days to want to do it on the record.
Yeah, yeah. No, I appreciate you sharing a little bit of that, and I think people can see that you're very authentic and—
Yeah.
—and passionate about this. You don't necessarily get that from a name like TypeSafe AI, but I think once people immerse themselves enough in the genuinely different direction you want the world to go, and actually you have done the hard part about going from zero to one on the thing, then now let's all go together in the new direction, right?
Yeah. Yeah, yeah.
Yeah.
But I don't know. Maybe I will think that the hard part was done, perhaps. I think that there's going to be many more hard parts. If all sorts of stuff gets automated and we finally see GDP growth and it's a Geff party every day, then maybe the hard part is done. But I don't think so. And I really, really think that people focus too much on speed and cost and not enough on reliability.
Okay.
Reliability is what makes it delightful. Reliability is what allows you to trust it.
You have this line—
Yeah.
“TFP growth rate: 3% in 5 years.”
Hell yeah.
I've never seen—
Hell yeah. Let's fucking go.
I've never seen—
Yeah.
—a lab care about TFP growth.
But that is what an economic revolution is, right? It's actually extremely consistent with what the OpenAI charter used to stand for.
Mm-hmm.
The charter is the same, but they've kind of tried to move definitions around to, you know, 100 billion in profit or something like that. Not that I hate OpenAI.
It wasn't a well-defined term, what AGI is, right?
They tried to do it.
Yeah.
Right? Doing the majority of the world's economically valuable work. They should have to answer the question: How can it do Millennium Prize Problems in math and zero of the world's economically valuable work, like rounding error? I think that all models are roughly tied right now at zero. There's some chance that we have started already, but I would guess that it's not yet 1%. And I think that when it does happen, it will show up in the economic statistics.
Mm-hmm.
It's going to be fucking awesome. It will not cause mass unemployment, but it will cause a whole bunch of awesome shifts, and the world will be a lot better. And I'm really tired of AI always being the foreground character of things. I think the world should just be more delightful, and AI should just help with that.
Mm.
You know—
Just disappear into the background.
Exactly.
Yeah.
You know, I say this in my talks: how can it be that 2019 software—software, SaaS, whatever—was super-duper valuable, right? It's 2026 now.
How is the software basically exactly the same despite AI being so freaking awesome, other than sometimes having a chat box on the side? That kind of works, but it doesn't allow you to make decisions that the companies have stakes in because they can't be trusted to make decisions. That, to me, is nuts. There's so much economic incentive for this, and I think it's going to be an inverse SaaSpocalypse. I think SaaS is going to be supercharged by this. They are the ones who are most in the know about what things are valuable to automate, and it's going to be a crazy time.
Yeah. I think so too. It's a beautiful thing that you've unlocked, you know?
Yeah.
You mentioned one thing here, which I don't know if it's directly related: What is a System 1 problem, and what is a System 2 problem?
Fuck. That's a hard one. That's a hard one, my friend.
Because people now are just trying to Jev everything, right?
Jev everything.
Which probably is going to fail, right? But some things are going to be good.
“Jev everything” is pretty funny. It's a pretty funny way of saying it. So I'll tell you the truth.
Yeah.
The truth is that this is an empirical problem, just like scaling laws are an empirical thing. Why doesn't robotics really work right now despite all the money being spent on it?
Mm.
I don't think it's about necessarily spending more money. The empirical results might just not be there, right?
Empirically, I believe that these pre-trained super-condensations of intelligence are fundamentally System 1 thinkers. I think System 1 is the closest thing to describe what LLMs are strong at. RLVR has done incredible things for System 2 thinking. I am in awe. It is super freaking cool.
I don't think that it's going to result in AI doom in the slightest. Not zero percent, of course, because I think zero percent is miscalibrated. But it's really cool what they've done, and they've really pushed it to the limits. Well, maybe they don't think so—not the limits of limits—but it is a weird thing for models to do, and they are very fragile at this.
Think about how people used to talk about AI back in the ChatGPT days: “Wow, it's really general. It can do a lot of general things.” But it was bad at math problems and GSM8K grade-school math. Now look at how people talk about RLVR. It's so fragile. It's so jagged. Why can it do this really weird thing? And actually, math is not just spiky; it's fractal, right?
This is because RLVR is—if we talk about the North Star for each thing—RLHF is “please humans,” right? That is what the human feedback is. RLVR is “optimize benchmarks.” Everything that goes into the RLVR category literally is a benchmark by definition, because a benchmark is programmatically verifiable, with simple outputs that can do well. RLCD is “make it reliable for programmatic use.” And, yeah, that's—
Yeah.
Yeah.
This—maybe I'll offer some thoughts, and then you can correct me if I'm wrong.
Mm-hmm.
For example, one thing that I've been thinking about is that I threw Jev at a bunch of things when you gave me access on day 1. Multi-hop reasoning, right?
Yep.
Single-hop is fantastic.
Mm-hmm.
State of the art. You should never use anything other than Jev for single-hop.
Yep, yep, yep.
Multi-hop starts to fall down.
Yep.
And it's kind of monotonically increasing as you increase the hops.
Yep, yep, yep.
Right?
Oh, yes, back to that empirical question. It depends on what we can pull out of the models, right? We want everything. We want to unearth as much intelligence as possible, period.
I see us as unlocking, smoothing, and sculpting the intelligence, while adding new capabilities and filling in gaps in it. We will be filling in more and more of these gaps over time. But the reality is that we are in the business of unearthing properties.
Those properties are actually a function of what is available from these condensed cores, and of Frankensteining them all together to have all of the properties of everything.
Yeah.
The reality is that we are in the business of unearthing as many capabilities as possible, and System 1 just happens to be the description of what works. Everything that works in that paradigm will be System-1-ish.
There is a reason why we don't do what's called latent reasoning—reasoning in strings.
Yeah.
I think what models do really well is reasoning within the models. It's not totally complete. It doesn't do great at all.
Wait, latent reasoning is reasoning in strings? I thought latent reasoning was reasoning inside the model weights. I just wanted to clarify.
People used to call that—
I don't know. I just wanted to clarify.
People used to call that continuous reasoning.
Okay.
I'm not entirely sure. It was called latent reasoning because the reasoning traces were secret, so they're kind of like a latent variable for the answer.
Ha.
Yeah.
So what's secret is now shifted.
Well, it's still secret for OpenAI and Anthropic, right?
So no reasoning, Jev, as far as you'll ever do it, right? Because that violates the whole promise of System 1.
My promise is to do whatever is necessary for machine-native stuff.
Yeah.
I could imagine there are some forms of reasoning that are less slow, inefficient, and fragile that are totally on the cards, just to be clear. I'm a pragmatic person. I'm not making promises on methods. I'm making promises on what my ROI North Star is, and I'm going to fight for that.
This launch didn't happen, and we are still hungry for our place in the world.
That's great. Yeah.
Yeah.
I think the other thing is that vision is another big capability that you don't have. But maybe it doesn't ever belong in System 1.
I think I have pretty good vision.
What? Sorry?
I think I have good vision.
No, no, no. Sorry—
I'm kidding. I'm kidding.
Oh my God.
Yeah, yeah, yeah.
Because, obviously, the first thing people want is vision because of the Doom demo, but also just everything other than text is vision.
Everything is in the cards in my mind.
Okay.
This is a debate we have. Your audience is probably a great one to have in this debate. There's a question about how much we try to give people what they think they want, which is what we did in stealth for 2 years. We just knew that this was obviously going to be valuable, versus giving them what they say they want.
There are a lot of dimensions to this. Context length is an example of this. Every single model, including ours—I actually think, as far as I can tell, ours is by far the best at not degrading in long context—
The longest context. Yeah.
—but the other providers are just like, “Whatever people wanted, let's just give them the stupid thing.” We need to figure out a balance for this because—
Yeah.
If you take the former side too far—give people what they want—you end up with Anthropic nanny-state-style thinking, which is very anti-developer. While the pro-developer route would be to give them what they want, we don't want to put the burden on developers to figure out the je ne sais quoi of intelligence.
So we are trying to figure out how quickly to release things, to still have our brand of trust and also treat our users like adults who can make informed decisions and don't need nanny-stating on top of this stuff.
Yep. I think that's fair.
Yeah. And we don't know the answer, to be honest. We'll have to figure it out.
That's probably going to be one of my biggest debates over the next couple of days.
Yeah.
Because we have a lot of stuff. Again, we didn't expect it to pop off, so we were like, "We need some follow-up launches."
I don't know if you didn't expect it to pop off. I saw the work that you put in. I've never seen you lock in so hard. It was like the last 2 months, basically, right?
Well, that's also because my chief of staff made me lock in. I thought I worked hard before.
Yeah. No, but—
Yeah.
You were showing up at our writing workshops, and I was like, "What are you doing here?"
It was useful. It was great.
You clearly were very intentional about your launch.
Yep.
The work showed, and congrats—all the kudos.
Thank you. Thank you.
I hope to keep locking in—that's my sense.
Yeah.
I think we've passed many great filters for the tech world and what we're wanting, but there's still going to be a bunch more. Holy smokes, am I excited to fight the good fight.
Yeah. It's exciting. Before we broaden out to topics outside of TypeSafe, I just wanted to offer you the opportunity to mention any other things that you think are underrated or misunderstood about what you have launched.
Underrated or misunderstood?
Yeah. You have patterns here. Maybe you want to go into that. Model jaggedness, anything.
Give me one—
Yeah.
Noodling of it. Oh, man. I would rant about all of these. I really shouldn't. I really shouldn't.
Okay. People can go to your Discord if they want.
Yeah. People put a lot of love into the cookbooks, is what I will say. The cookbooks have some fire stuff. We had considered putting a bunch of these things in the main launch blog post, but it got kind of long and unwieldy and very power-user-y.
I'll be frank: before the launch, everything we were saying sounded like this weird alien tool. Why would anyone need this? It was a very weird thing, and we were very worried about teaching people about this new frontier. It obviously succeeded, but we put a lot of work into it because we thought the education would be a gigantic bottleneck for us. It probably worries us, and it's probably no longer a problem because people are doing things well beyond what you could ever expect.
They'll show you how to use your model.
Exactly. Their use cases are kind of cooler than ours. There's a bunch of stuff where I'm like, "Man, if that was our demo, holy shit, that is way cooler than what we were showing."
The computer-use stuff, holy smokes, is it cool. We put a lot of love into this. This is not AI-generated trash, as far as I know.
Yeah. Fair enough.
Each of these has real alpha there. These are inspired by solving real customer problems that existed, and we went through the work of helping them do cool-ass stuff.
How much validation did you do before launch? What was that process like?
What was that process like?
Clearly, you did some, but obviously you're not getting in touch with as many people as you are today.
Yes, of course.
But—
I actually think that the reception was pretty bad. For the nontechnical people on the team, they were really worried. There was a lot of fear. No one really gets this, and they don't want it. We're selling a vitamin and not a painkiller. Should we have FDEs write the software around solving that problem?
Yeah.
We had almost no revenue before launch. The technical people were obviously true believers. We knew that this was sick. Its computational properties are off the charts on so many axes that we were like, "Yeah, obviously it's going to be huge."
I was definitely super afraid, which is why I locked in super hard. But the most common thing was that, I would say, more than half the people we had play with it just did not get it. The people who did were like, "Man, this is really cool, but how do we get this through procurement and stuff like that?"
It was quite a battle, and we just knew, okay, our target market is going to be developers. People will find the use cases, and that way everyone is going to FOMO in. I don't want to rub it in when people change their minds as the facts change.
Mm.
I do want to call into question the concept of product-market fit. There was a product, and there was a market. We were like, "Hey, do you want to use this?" And people were like, "I don't really know if it solves our problems." Then it explodes and everyone's like, "We need as many rate limits as we can. Can we literally give you GPUs? Because we are constrained right now?"
Marketing is an element of it, of course, but I don't even think it's about marketing. I think it's about passionate developers whose souls basically resonated at the same frequency, and that frequency got everyone else excited too.
I'm hoping as well that we as a company will be eternally grateful to those developers—not just the companies that start off with developers and go to enterprises.
Big corporate market, yeah.
Exactly.
Yeah.
I'm even thinking about how we can launch things that are better for—
Oh, man, I don't know if I should say this, but I will.
Better for developers than enterprises.
Exactly.
Okay.
How do we do that? How do we empower them? I have cooks, I have cooks, but it's a very weird thing to do, and I don't know how else I can show my thanks and loyalty to that.
That's why I dyed my hair yesterday. I wanted to talk to them because it felt dirty to me, during our company's most important times, not to keep talking to them.
Good. Well, that's one of the reasons you're here.
Yeah.
Hold me to that, please.
Yeah, yeah. We will, we will.
I try to be principled.
Yeah.
Quote me on this. Call me out. Have the pitchforks out if I change.
I was just going to briefly show the computer-use stuff.
Whoa.
Is this what you're referencing?
I've seen an airline browser-use thing.
Inside this new note, let's make the title say, "Hello."
Wow.
Great. Okay. Let's move on. Can you open the Arc Browser? And once you're there, can you Google search Norbert Wiener? Now can you open x.com?
Is this the kind of use case?
Oh, the voice use cases. This is actually the first one I've seen. This is—
Oh, okay.
Open up the photo booth.
Wow. Oh, wait, wait, wait, wait. Hold on. Can you go back a second? Can you go back a second?
Rumors claim Anthropic engineers worship Claude as God. Wow. Dang, that's pretty funny.
And here you are building prod.
Absolutely—wow, this is sick.
Clearly, you can operate the whole computer with voice, with Jev as a decision model.
So just like I'm anti-benchmaxing, I'm also anti-demos. I want to make sure that it works reliably. I love that people are playing with it.
Mm.
This is super fucking sick, have no doubt. I want to see this. I want to see it be used. I want our team to play with it. I want to find the weaknesses, and I want to solve that.
Yeah.
Man, that looked really cool. That looked really cool. I want that. I want that. When my wrists are sore, I just whisper flow everything. That'd be sick.
Well, just to round out the use-case side—
'Cause I do have to let you go. Who are the bigger companies that have reached out and surprised you with what they want to do?
I am so out of touch with that. People have shown me screenshots of companies, and from what I've seen, it's all of them.
Yeah. Mostly, for those people who work at larger companies and aren't doing this kind of work, I just want to give people examples of: you should go look that up, look that up, look that up.
Oh. So, I think demos are super-duper sick. Obviously, coding agents are gigantic use cases. They're also super sick.
Mm-hmm.
Can I give a little bit of a tangent about coding agents, if that's okay?
Yes, please. We love coding agents here.
Okay, actually, I'll come back to coding agents. Let me describe the big families of use cases.
We've mapped this out from first principles, long before release. They are what we call dark data. People hoarded big data, but they wouldn't throw an LLM at it because it was too expensive. Large companies love this. They have piles of data that they wish they could analyze, and this is a data scientist's wet dream.
This is a giant one. I think this plus coding agents are the big money makers, because that's where all the volume is, right? There's the real-time stuff. People who need intelligence in the loop—I would guess that every CEO, if not CTO, at those companies knows how much better their product gets with every 10 milliseconds shaved.
Yes.
And, especially for assistant-y things, there are many AI assistants, and as far as I can tell, they really love it. Again, I'm not on the front lines with customers right now, so I just know what my team tells me.
I'm so excited for this. I'm really excited for this for games. I really want to play sick-ass auto battlers where you're commanding your team, or semi-auto battlers. I think that'd be so cool, but don't make it too good while I still have a job.
There's also what we call “verify everything”—verifying all LLM calls, kind of like observability. Actually, on the note of docs, what people should be doing is asking parallel questions, because they're very cheap. So if you have big states you want to ask many questions on—
This right here? Yeah.
Put IDs on every message and then ask a question about each ID. When you have a long state, that way you can pay for the state once and ask lots and lots of questions about each message within it. I think that is a great way to save money, and it's interesting.
Which, by the way, I always think it's interesting framing system 1 and system 2, because it basically makes the case that you should always make 1, 10, or 100 Jev calls for every 1 reasoning call that you make.
Well, maybe.
Right.
I mean, I would like people to spend less. Maybe you do half the reasoning calls and 10 Jev calls each, or something like that—whatever solves the problem that couldn't have existed otherwise.
The fourth use case was what I describe as smart software: software that's intrinsically composable and does weird, fun stuff that could never happen before. You know, like the programming-language-as-Jev thing. I don't know if you've seen that. That is so cool.
If we knew how to give out credits—we're really early in our infra days—I would want to give all these projects credits.
Mm.
I think those are how we've mapped out the main use cases. Computer use has also come in kind of the real-time direction, and that's really, really cool. If it's reliable, I am super jazzed about that. I suspect we can make the model a lot better at these use cases because that came out of left field a little bit, so that's really cool.
On the coding-agent thing, this is a really surprising thing that's happening right now.
Okay.
Cloud Code and Codex are, I believe, the winners—the number 1 and 2. I'm not entirely sure. I don't follow closely—
Roughly.
It's roughly that. But they're built around a single-model world. That makes a lot of sense for them, right? It has been a one-model game where it's kind of the same model, but with different levels of intelligence that you're shopping for.
Mm-hmm.
But all the open coding agents are fucking jazzed right now because they're getting their Jevons on. I'm sure they're trying a lot of weird stuff.
But all the coding agents are roughly at approximate parity, right? There's not so much you can do with a while loop. The moment one person finds 1 killer use case that you can only do with that coding agent, everyone will flock to it because they have a monopoly on that thing. But all the open coding agents will be able to copy that, right?
I don't know what Claude Code and Codex will do, because they're built around that one-model world.
Single model, yeah.
I think that's going to be a really interesting thing. I would love to be able to integrate with them personally. I want to integrate with everyone. They might make competitors eventually; I don't know. But it is not my job as Sonfire Infrastructure to be opinionated on that, right? I want to just serve the world.
I don't know if they would do that. I think it'll make the coding-agent game super weird. I'm so excited for that.
I'm getting my team to review an internal document I made on design patterns I suspect will be useful for coding agents, so hopefully I can share it right after I walk home. I think there's just such a ripe area for exploration out in the world. If I didn't have this, I would love to experiment with coding agents right now.
Yeah, I mean, I'm sure the coding-agent companies would love to work with you as well to figure that out. I do think that there are still use cases for Cloud Code and Codex with you guys, which are easy to explore there.
I mean, you've been very obliging and indulging in all these things. I just want to take you out of TypeSafe and talk generally about alignment and safety. You've made your position on the state of AI very clear, so I'll give you more room on the alignment and safety side of things.
Oh, did I not talk about safety and alignment at all?
Oh, you did. You did.
Maybe I didn't. I think maybe I didn't.
You did. I think there's a lot of researcher discussion around this. We have this every NeurIPS.
Yeah.
What are people talking about? For example, I was recently at one of these researcher gatherings, and people are genuinely worried about the pacing. There's this whole topic about how we should slow down because the public is clearly not ready. I'm sure you have strong feelings.
I feel like this is the kind of thing that is a dangerous—
Okay.
—topic to talk about. I'm happy to talk about it. I live for danger.
Our company brand is chaos. It's not Jev. It is irreverence and chaos.
And, you know, you were at OpenAI during one of the very first, very visible incidents—the blip, right? The dominoes have gone down now to the point where every frontier lab has co-signed a document saying that they want to pace.
Interesting. It's a very complicated, nuanced thing. I actually do want to write a response to this more formally. I do have a short version of my response.
Yeah.
Which is that, as you RLVR more, RLVR is not actually about verifiable rewards. That has been failing since before the reasoning revolution. And that's the weird part about tasks, right? Back when—
Oh, fun history. Back when RLHF was becoming a thing, there were 3 different things that are now called post-training—different efforts. Instruction following was by far the bastard child. People didn't like it. They didn't want to take it into account. It was annoying.
I talked to the pre-training team, and I'm like, “Guys, this is the magic.” And they're like, “We've run so many model sweeps. You want us to wait for human evals to figure out which models to use?” Everyone was giving tons of resources to the code generation team, which did have some successes, but they were trying really hard to do RL on code, like unit tests. And it didn't work, obviously, right? You needed reasoning for that.
So, just to be clear, RLVR is not purely about the reward. It's about the shape of everything, too. Part of it is that reasoning is included in here, like this latent variable that you're doing things with. And when you're doing things, you're just letting the models do whatever they want in order to make them as powerful as they can be to answer the hardest problems.
8. The Pacing Debate
This whole “pace the frontier” discussion, I think, is a very narrow focus because it assumes that everyone needs to do more RLVR, right? I obviously don't think I need to do more RLVR on our models.
Mm-hmm.
You know, I think zero is the optimal amount for our shape. Right? Come on.
Yeah.
It's really, I think, a bit of a sleight of hand where they're saying that we actually want to keep doing the thing that looks dangerous because it does dangerous things. People say, “Oh, maybe the sandboxing was a problem,” or whatever else. Yeah, I mean, obviously it is, and they could have easily solved that, right? But they chose not to because the more things you let the models do in this “do anything” category, the more powerful they are, right?
I think there's some dissolution of responsibility there on things that, by design or non-design, they're trying to make. It's just an assumption. We must do RLVR, and not just must do it—we must do more and more and more—by giving the models the power to do powerful things, to do anything they want in the middle, because that teaches them to be powerful outside of it. And we don't want to limit those things because it'll make them slightly less powerful at those things.
So if you assume all of that, they're like, “Oh, yeah—”
That's a logical conclusion.
“We're heading into a dangerous world, guys.” Everyone is going to be doing this, and this is the only way to make AI sick. So—
Right.
Of course. I think there's very few people who've made the right bets on the Bitter Lesson direction. You know, new directions of AI. Or new North Stars. That is rare. Again, I think 2.2 times or something for LLMs themselves, like RLHF and then RLCD.
Oh.
RLVR is a 0.2, in my opinion, and I think that's generous. Or 0.5; it could be 1.1. I don't really care. But I do think that people are thinking very closed-mindedly about this type of thing. The only people who are at fault here are the researchers, because it's definitely not the populace. They just assume that OpenAI and Anthropic are doing the best they can, and they are not the experts who are aware of the true optionality available.
Yeah. And that's fair. You're also doing your part in waking them up.
Yeah. Well, I'm doing my best. But my goal is not to convince labs that there are other directions to go down. My goal is to—you know, it's like—spark hope in software engineers to start actually automating things they've always wanted automated.
I had this article that I wrote that my team didn't let me publish, about the future I want for AI. There are a lot of little things. Remember “Do What I Mean”? Imagine if everything could do what I mean, because that demo was “Do What I Mean.”
Mm.
Like, there are levels—
Don't—yeah, don't do what I say.
Hmm?
Yeah. Don't do what I say. Do what I mean.
Yeah. And we couldn't do what I mean yet because computers are so basic and literal. But that computer-use one was just that. And I think there are levels of smoothness that'll happen in the world that people just don't understand.
The promise of smarts all around is—I don't want to overpromise. I don't think it's going to happen right now, but we're going to do whatever the fuck we can to make that happen.
Yeah. Any other things on the general shape of post-training? You obviously have been very intimately involved. Mid-training—is that something you have comments on? I don't think we've ever talked about it.
Mid-training. I mean, it's all a spectrum.
Yeah.
Right? Like, am I—
This is a curriculum, but fancier.
Yeah. I mean, it's a cost-saving thing, instead of having to pre-train again.
Yeah.
There are intriguing things. I actually think that intelligence has a je ne sais quoi at every single level, and it's always super-duper fascinating. I am a shape rotator, so I don't like finding that, but I love it when people find it and teach me about it.
And, you know, looking at the data is something our data team is so good at, and I'm not—
Hmm.
It's really fascinating. I love thinking about how capabilities are put into the model over the short term. There's the really rapid alignment of fine-tuning, and over the long term, after seeing it over and over and over again, this stuff gets baked deeper and deeper and deeper and deeper into the model until it gets robust.
And that is the North Star to surface, and the System 1 stuff is the stuff that ends up getting robust. So I find mid-training to be a fascinating thing. I'm a fan of all forms of training. I'm a fan of all forms of surfacing new types of intelligence. I wouldn't do it all myself because it's expensive.
I've said privately and also—should I say this? Huh. You know, my philosophy is: anything I say in private with an investor, I should say in public with the people, because that is—
Power to the people.
—my thing.
Yeah.
Yes. So the thing I've said before is, if you gave me $1 billion, I wouldn't pre-train. I still believe that to be true. It's a very expensive thing. If you're an AI engineer, you can slice and dice and do all sorts of stuff. Frankensteining is not the most elegant, beautiful thing, but it solves problems, baby.
Anything except pre-training.
Mm, mm. Yeah. Amazing. I think one direction that I do think is interesting, just synthesizing all your commentary about these model things, is: do we have a supermodel that has all these capabilities involved, or do we break them out further, I guess?
One way to put this is that OpenAI was trending in the direction of the omni model. GPT-4o was one of those. Then, for a brief period of time, there was a kind of a main branch of, “This is the chat-tuned model, and this is the coding-tuned model.”
Those are two completely different things. Those are extremely different concepts. I'll break that down a little bit. Multimodality is a little bit different—
Mm.
—because sometimes the other modalities help, and sometimes they hurt.
Yes.
People are moving—they seem to be moving away from speech, which is different from audio, because it seems not to generalize well to the other stuff.
Mm.
This might get solved. I'm a fan of all of this. But these are empirical, real questions. Scaling laws are not about just throwing money at it and it gets good. Scaling laws are, pragmatically, how good is a thing? There are worlds where, no matter what you scale, it may not be good enough.
Computer use is not currently solved, is my understanding. I'm hoping that we can play a part in solving that, but there might be no amount of data we collect that will solve that. We might need better methods or something else like that. You need to be really practical in all of this.
Am I a fan of omni models? I'm a fan of all forms of intelligence, but I will go straight into one thing you talked about, which is different from pre-training, which is post-training—
Because I hate fracturing intelligence. That is the bad thing to me, and this whole chat-first reasoning mode forces intelligence to be fractured. When you're optimizing for chat, this tends to be pure RLHF, and it's intrinsic to RLHF to do the stuff people naturally complain about.
You're absolutely right.
Yeah.
You know.
Sycophancy—whatever word, however you pronounce that—overconfidence, hallucination. Even the kind of style that excels in LM Arena: bold, italicized emojis. It doesn't answer the question simply. It gives you a long write-up, and then it asks you a follow-up question, so it feels more like a human talking to you.
All of these things come because strings are super weird. They are weird-ass things, and you need to be miscalibrated. You need to mode-drop. You need to be hyper-confident in order not to go off the rails, because the reward model will punish you so hard when it happens, because it's obvious.
And then this warps the probability space entirely, and it interacts with that of the reasoning models, right? Because the models are these simple linear things that tend to cheat a bit. So I think that's very different from exposing intelligence, is my guess, and a lot of the art of intelligence is studying the subtlety that I think, at least when I was at OpenAI, people were not really studying, because they were just like, "Chat, chat, chat, chat," just like people are on with Jeff right now.
Yeah, you give me an objective, I will just optimize for that, right?
Yes.
But if you try—and the saying is that you could have 2 objectives and just optimize for both—but that is literally the act of fracturing, right?
Which is fine.
It's a little different. If I could defend—
Yeah.
The System 2 tasks. Number 1, we don't toss out the System 2 tasks, right? You can try to make Jeff work on it, and there actually is an intelligent answer for that, which is unknown. There is better and worse behavior in the System 2 tasks, which should be really low-confidence, with lots of uncertainty. Maybe some heuristics can move the needle here and there. But we care about them too, just to be clear.
I just think that that is not what intelligence is native to. So we're not trying to fracture anything like that. All fracturing makes the model dumb. If people get the model to say that it is OpenAI or Qwen or Claude or whatever else—I don't really know what it says these days—I am not going to put into the models that you are Jev from Typesafe. That fractures it, right? I don't want that.
I want it to represent what the internet thinks, right? Be correct. That is what I want, because that's how you get smooth, predictable intelligence.
I mean, identity is a thing, I guess, that—
I think for an API—
It may be somewhat of a special—
For a first-party product, yes.
Yeah.
But for an API, I don't think so.
Yeah, okay.
You know?
Yeah, that's good.
People don't want, if they're making a chatbot with ChatGPT, for it to say it's ChatGPT. They want it to say it's Chipotle or whatever, right?
Well, the way that you also have to make up for it is you have the skill, right?
Yeah.
The Jev skill—
Yeah, they do.
Which is for coding agents to work with Jeff. Okay, a couple of closing questions, because I do want to get you out. One is just reflecting on your 2-year journey. It's roughly 2 years? 2-point-something?
With the company—
Yeah.
I think that this is more like a 4-year journey.
Yeah.
But—
Well, yeah.
Yeah.
Actually, I was remembering that you had this hero run around Thanksgiving. You were canceling everything because you were like, "Guys, everyone's on holiday. I'm going to take all the OpenAI GPUs and go do this thing."
Yeah. That was a good time.
And that was the pre-Typesafe moment, right?
Yeah, yeah, yeah.
I might have been—was that when the coup was happening? I don't really know.
Yes, actually.
Yeah, yeah, yeah. That sounds right. I remember. Oh, my God. I don't want to—I don't think I have the time to spill the tea about the coup right now, but that was really annoying.
Was the coup annoying, or the run was annoying?
The coup was annoying.
The coup. Okay.
Yeah, yeah, yeah.
Yeah.
Safety took over the company.
Yeah.
Maybe next time we chat—
Okay. All right.
I'll spill the tea about—
Tea.
The coup.
9. The Road From OpenAI
Actually, this problem was one that was in my mind since before ChatGPT even launched. I was like, "Holy shit, the ChatGPT team is cooking. They are doing the right task. They are doing the thing that AI researchers are bad at, but successful product people are good at, which is giving a lot of fucks about the experience."
It's very rare. There are very few people like that at OpenAI. Those guys were cooking on it really, really well.
And, to be clear, this is the whole journey from GPT-3 to GPT-3.5, which included AI Dungeon—
Yeah.
Which you've talked about as an example of a use case that we never predicted.
Yes, exactly.
Right.
Yeah. Oh, yeah, that is also—
I had fought very, very hard to deploy InstructGPT. Actually, the early versions of it were even trained with an algorithm we didn't publish that I made myself, because it was too slow to clean the PPO data. I was like, "Fuck it. This is so fucking good. We need to get it into the hands of users."
Basically immediately, it took 50% of the market share of LLMs at the time. I went through great effort to make sure everything in our launch video was true. I was truly thinking, "Is this AGI because it's superhuman at instruction in, instruction out?" Obviously, it's not, but everyone, I think, should have an answer to why that was not AGI, because it looks very smart.
My answer to that ended up only being used for copywriting: Jasper AI, Copy.ai, writing what is now called slop on webpages. We were worried we had made the internet a worse place, right?
I went back to the drawing board and I was like, "What's missing? We are smart, clearly. Something is missing from it, like creating value. What is it?" I was doing more philosophy at the time: What is going on?
The question I asked myself was, let's work backwards from an AI-based economic revolution. When that happens, what will we be calling the AI if AI is an API? Will it be humans, or will it be code? And I figured it was many nines of code.
But all the optimization was going into the humans part. Then it clicked for me: "Holy shit, this is the North Star." I think I wrote a document. I was talking to Sam about this. Sam was like, "This is so fucking good. You should go work on it." And we're like, "Yeah, yeah, yeah, Sam, I have a job." I was working on—
Sam just told you to do it. Go do it.
But my guess at the time was that this was super obvious. It's so unbelievably obvious. Anthropic must be working on this already. We're already cooked. And actually, OpenAI does better at catching up than it does at innovating.
So ChatGPT was a copy of Claude, right? They had an internal thing. They just—
Yes.
Didn't ship it.
Yeah.
Claude—
Claude in Slack. But reasoning, I would say, first-ish.
Yeah, but—
Yeah.
It's debatable how good of a product that is.
Yeah.
Great research, though. Super-great research. I'm just not sure if people had that product need.
And Claude did the coding agent stuff too. So Sam says that, and I just go back to my job for a while. Eventually, the instruction-following team just says, “We won. We’ve solved instruction following. We don’t need to do stuff anymore.” I’m trying to think about what I do next. I was like, “You know, maybe I’ll just start playing around with this.” I do more philosophy and design and thinking.
I thought it would end up taking a week when I started training models. It ended up taking many years. At some point, I was like, “Holy shit, there are signs of life here. This obviously didn’t work, right? Otherwise, we would have deployed it.” But I want to explore what it would be like, research-wise, to go all in on this. I want to really see what it would be like if you went absolutely insanely all in on this direction.
And because of what I said, if an AI winter happened, how would I feel? I would consider myself personally responsible. I talked to other companies at the time, and I was like, “Hey, I want to start a lab in this direction.” I talked to them: “How fast—what would be faster, this or a startup?” And they’re like, “Startup.” And I’m like, “Fuck it, man. We ball.”
Yeah.
“I guess we’re doing some crazy shit.”
And you called Eric and Sasha—
Yeah. Well—
Yeah.
—I called Eric first. With Sasha, I actually didn’t try to recruit her. I tried to be good, and I was just like—
Mm-hmm.
—“Hey, am I crazy? Is something missing here? Isn’t there a solution to this? Am I too much in the OpenAI bubble that I didn’t realize there must be a solution to this?”
And then Sasha was like, “I’m in.” And I’m like, “Sasha, you’re working at a startup.” And she’s like, “I’m folding it right now.” And I’m like, “Do you want to think about that?” She’s like, “Oh yeah, good point. Let me think about it.” And then she joined.
Yeah.
And then, within 2 weeks, we had funding. We had people move into my apartment. It was the worst because I’m a neat freak. We just kept cooking, and eventually we got the research that showed the signs of life.
Yeah.
It was a crazy time.
So the question is, that was all long context.
Oh, yeah.
And then now the question is, someone like you is in the frontier lab right now, frustrated about not getting the funding or the resources or the attention. What’s your advice to them? Should they do what you did?
Should they do it? Ooh, that’s a fascinating question. Ooh, man, how do I do this without burning bridges?
My sense is that most—unless there’s some level of economics I don’t really understand—I think most neo-labs are crap. I don’t want to see myself with them as peers. I don’t really understand what’s going on there. Number 1, I don’t really value researchers. I value people who look at The Bitter Lesson, right?
The data, the task.
I want—well, not just that.
Yeah.
We need researchers, but we need them to give a lot of fucks about the right task, and that’s the important thing, right? So it’s actually kind of backwards when people value pure research pedigree, because that generally doesn’t create value.
Number 1, I believe in North Star tasks and doing cool, really useful stuff. Number 2, because I don’t value researchers, I don’t—
Well, it clearly is profitable for someone, or it might be in this environment. From a purely pragmatic perspective, I don’t see creating neo-labs as something that creates value. It seems to destroy value because they are redoing work from scratch with a low probability of actually moving the frontier.
As far as I've talked to most neo-labs, they don’t really have a direction. They tend to want money to play around with their experiments. If they have a direction, I’m super in favor of it, to be clear.
So my advice for someone is that it really depends on why you’re doing it. If you are a researcher who wants to play around with research, probably the labs are the best place to do that, TBH. There might be other places. I don’t really keep track of the politics. But I would just recommend not being that way personally.
I think it’s better for the world if people are driven to solve real problems. Those problems may be exploratory, and that’s fine, but ideally have principles that you stand behind. But if you think that you want to do the right task, absolutely fucking do it. Please do. Please break this unimind, you know, unimodal—
Hive mind. Yeah.
Yeah, exactly. Again, this pacing the frontier is coming from this one view of AI that looks like an AI supergenius that is incredibly jagged, and that is—
Solvable.
—it’s solvable, and it’s weird, and it’s not matching reality. It’s tragic, right? I think all of this—really unearthing technology—is just good.
Yeah. For what it’s worth, I’m trying to accurately represent the position of the Anthropic and OpenAI folks I was talking to, and SpaceX as well, by the way. It is a political thing much more than a pure x-risk thing.
Yep.
So, yeah.
Oh, sure.
Political positioning is—
And that’s beyond my pay grade.
Exactly, yeah.
That’s well beyond my pay grade.
Once they told me that, I was like, “I get it. This is about the 2028 election.”
Oh, no. I wish I didn’t hear that. That’s such a bad vibe. And so—
No, no, no, this is not the whole company. This is just that room’s discussion.
No, no, no, that makes sense.
Yeah.
That makes me lose faith in humanity a bit, but maybe I’m just a naive technologist.
It’s really starting to matter who’s in charge of the governments that will help regulate these things as they emerge, and, as a lab—
I totally—
—you should probably think that through.
No. I totally agree with that, to be clear. I think being opinionated on that matters a lot. I personally am afraid of trying to mislead people because I think that bites people in the ass a lot.
I’m not actually going to talk about politics. I think what happened during COVID is that people leaned too much into appeals to authority and being overconfident to try to get people to behave in certain ways, and obviously our response was extremely suboptimal. That had ripples of downstream ramifications that are now, I think, extremely bad for the world.
Maybe I’m naive. I think misleading people, even for the greater good or what they think is the greater good, is just—I’m not a fan.
Yeah. I—
I would rather not engage.
For what it’s worth, I—
Yeah.
—I don’t think it’s misleading. It’s just, like, this is why now. Like Dario said in May, “How come we’re doing it now?”
I think if that is why now—
Yeah.
—that is a little bit misleading about the risks versus the objective. There is some level of sneakiness latent in it that is worth calling out and, I think, owning up to. Well, obviously, if they want to manipulate, then they shouldn’t own up to that. That seems like a bad strategy.
No.
But that, to me, is just sad for the world.
Yeah.
Hopefully, we are never involved in anything like that. It might be inevitable as we get big. But I want to stay a pure technologist, true to my roots, as much as I can.
I mean, Jeff for president, why not? I would trust Jeff’s decisions over my own. Okay, so less shitposting, more about—
That’s shitposting.
More—no, no, no, no.
You’re just cr—
For me, I’m shitposting.
Oh, okay, you’re just crushing my hopes about—
No, no, no, I’m not calling you shitposting.
—like America and the world right now. Oh, my Lord.
Yeah.
I think I watch too much TV about conspiracies to think about the presidency.
Oh, no.
You have chosen your North Star. You have chosen reliability, and then programmable and composable AI.
And cheap.
And cheap.
Yeah.
What is a second or third one that you want to throw as a bone to someone else—something that you want someone else to work on, that you're not going to work on?
Ooh.
Basically, give people tasks.
Give people tasks?
Yeah, like, your tasks—
There are so many I want. Oh, what?
You have picked your tasks, right? You know what I mean?
Wait, that's such a good question. Holy crap. Oh man, I'm so excited by that.
Because, for the next 50 years, you're going to be busy doing your thing.
Hell yeah. Okay, so let me give you a fun one and maybe a valuable one that's also fun. My fun one is that I think games could be so freaking cool if they were intelligent. When I see people play around with Ali's Doom demo, where you can get NPCs to control stuff, that was just really a proof of concept. I think some really cool stuff could be made. It looks really, really cool.
I'm a big Stardew Valley fan, and it's really static, but it's still compelling. I feel like there are a lot of cool stories that could happen. You don't need to call Jev in the game loop; it's probably too expensive for that. But even simple state machines for NPCs could make such a compelling world. Oh man.
Mm.
And man, I'm a little sad that I can't work on these types of things.
Yeah.
My life path is a little bit set right now.
Yeah, but you can call someone else to work on it.
Yeah, that's cool.
And then you can—
Yeah.
Get feedback on it.
10. Beyond The KV Cache
The thing that I would really, really like to explore is coding agents free from the tyranny of the KV cache. It might not be as good as true coding agents are, but I think there are so many weird things to think about. That's why I wrote the article “KV Cache Rules Everything Around Me.”
Mm.
Believe it or not, I don't think anyone had used the phrase “cache rules everything around me”—C-A-C-H-E—on the internet when I Googled it.
Okay.
I wrote this because I wanted to tell people how coding agents work and how the KV cache works.
Right.
It explains a lot of things, like why routing is really hard, why sub-agents don't seem to work, and why compaction is such a hard problem. I'm going to try to release a document. My team might veto me because, believe it or not, I'm not in charge. I wish I were. But I want to release a document of my thoughts: please play with it, and please figure out all the ways that we can do things with coding agents once you're freed from that KV-cache tyranny.
It locks you in, and—
Well, no, it locks you into one model, right? And in order to do it efficiently, you need to keep appending to it.
Mm-hmm.
So now you're not doing best software practices like state management, abstraction, and decomposition. Why can't you give an easier task to a sub-agent? Because of the state that you're passing around. I touched this. You would need intelligence that is way cheaper than the intelligence you're using to read this in order to pass this state around. Why can't you be smart about it, right?
I think there's tons of really cool, fun research to be had there on different programming patterns. It's kind of like how people are playing around with recursive language models. I feel like there's lots of cool stuff in here when you think about, “I want to explicitly label the state of everything.”
Or imagine you have a subtask. Coding agents, I think it's fair to say, work on subtasks one at a time from a decomposition perspective. Why do you need to pass all of that state back into the parent task?
Yeah.
Why couldn't you do smart things about it? Also, if you had a hierarchy of labeled subtasks, why can't you search through that subtask tree for the relevant context when you need it?
Another thing that you can do—oh man, I forgot to write something about this. I have, like, some cooks in here that are really, really cool. I hope to publish it. I'm down to jam about it, but it's going to be a long document.
If it becomes the case that context becomes cheap, why can't you do cool patterns like looking at your historical context very cheaply? Isn't it kind of weird that you start from scratch every time and need to solve a problem called continuous learning? That's actually a memory-management problem, because you don't have a smart way of looking up the memory, right? But what if you could? What if you could do that all the time?
Or what if, when you have parallel sub-agents, they can read each other's states because you have all of that in your computer memory? You can be smart about what's reading and writing at the same time, and your coding-agent swarm or whatever has locks around things and can coordinate intelligently—not with basic-ass locks. Like, “What are you doing? What am I doing? You know, Jev, who should write first?” I feel like the future there is—
Oh my God.
Nuts.
Devin to solve locks.
I mean, it could be so cool for multiple agents working together. Or, if you think about state—
Yeah.
When you have—
Agent swarm stuff.
Yeah.
Yeah.
Some things, for example, are read-only processes. Some people like getting summaries of what the agents are doing.
Mm-hmm. Yeah, yeah.
Why can't they share state easily? A read-only agent needs to read parts of the context and figure out what's relevant to say—what's actually being written, whether exploration is not super important, or whether it's the tree of subtasks. I feel like there are so many different fun things that could be done if a really smart person dedicated a whole lot of time to rethinking the coding-agent experience. That would be super-duper sick.
Yeah—
Man, that would be my dream.
I would point you toward Prime Agent if you haven't looked at it. This works together with the RLM work. We just talked to Alex, who's a buddy of Ellen's, in the chair before you.
Oh, cool.
It is being worked on, but it's not super popular yet.
Yep.
Well, yeah, but the hope—
Yeah.
I would want everyone to just play around with weird things. I have no guarantees it'll work, but it seems really, really interesting from a technical perspective. So, yeah, that seems cool.
That seems cool.
Once we figure out how to give credits out, I would love to give credits out to people like this.
You'll be in a position to fund research, for sure.
Yeah.
No, anyway, congrats on all your success. You've come such a long way since I first met you, and the whole team as well.
I'd like to think I'm the same person as well.
Yeah. I think you are energized in a way that I've never seen you before because you found your mission.
You know? That's true. That's definitely true.
You are articulating your mission because, for many years, you complained about the problems, but you didn't have a solution yet, right? You had the rough shape, and then you had to put in the work.
I will say that that is partially because I describe myself as 0% entrepreneurial. I don't like startups. I never wanted to be a CEO in my life. I can't imagine anyone doing this twice. It seems horrible. Honestly, doing it once is pretty bad.
When we were first fundraising, an investor asked me, “Which CEOs do you look up to?” And I was like, “Ew, why would I look up to those people?” No offense to anyone. I'm trying to be genuine. I've met a lot of really good people, but the famous ones seem to have a lot of skeletons in their closets. And I think I just really did feel disempowered when I was at OpenAI. I felt…
Yeah, it’s a little bit easier to be truthful now because I have at least some proof that the direction has legs. I just felt like I was in the insane house where everyone was just like, “ChatGPT, yeah. Where do we put ChatGPT in everything? How do we make ChatGPT good for developers and stuff?” And I’m like, “What? What are you talking about? The function-calling interface is insane. Why would you deploy this?” This is just so anti-developer.
It’s sort of a hacky way on top of hacks on top of hacks.
Well—
Yeah.
Not just that. The thing I often said was—this is also probably tea I don’t have time for right now—but I always used to say, “I want to be removed from any project involving function calling if you did not get a legit logit bias for each function.” It’s a very—
Mm.
Very simple ask on my part because—
Which is something like a confidence, but not calibrated—
—or a probability for it, right?
Yeah.
We need to give users the ability to control whether to take actions, refuse, or allow—
Oh, yeah.
—Disney needs to set a different refusal threshold than AI Dungeon. The only way to control that with function calling right now is to say, “Pretty please.” That’s nuts. That’s a nuts interface for developers, and people have been dealing with this for years now, right? They still have it with skills. The existing coding agents are highly overfit to their existing harness because they’re jagged. They don’t tend to use external tools and MCPs super well because of overfitting, of course. And why can’t big companies allow for slight nudges to be like, “Call this more. It’s really useful,” right? The solution is begging in a system message. That’s nuts.
Oh, yeah. But, okay, I think I get you. And, man, it is so exciting to talk about all this stuff.
Thank you.
It’s really cool to get you on a podcast. You’re going to go do amazing things, man. I’m excited for your next big launches, whatever that is.
Oh, hell yeah.
Yeah.
Just you wait.
Yeah.
Just you wait. It might be sooner—
So, hiring data people—
—than you think.
—infra people, I assume. Marketer.
100 feet. Depends on who you ask.
Community person.
If you ask me—
Yeah.
—I feel like I’m a pretty good founding marketer, but if you ask anyone on my team, they say, “Shut the fuck up, Yogo. You need to do CEO stuff.” So, yes, founding marketer—
It is not just about spice. I think you’re very spice-oriented, which, like, that’s—
Mm-hmm.
—your unique talent. But sometimes you just need to say, like—
I know, I know.
Yeah, yeah.
I would really love—
The routine marketing things. Yeah.
Yes. Nothing teaches you delegation like having a tidal wave of stuff to do. Hiring data people, or we call them model capabilities, like—
Yeah.
—but they are data people, because data is kind of a slur in the industry, and I want to make sure that—
I don’t think so. We are very pro-data here.
Yeah.
We—
But I want them to have the highest status—the people actually working on the model. I want everyone to have equal status, but I want to even that out, and I want to know that that’s really valuable.
These are more equal than others.
Well, I mean, I don’t like weird hierarchies, and I think one of the things I’m most proud about in the company is that they don’t respect me that much, or they don’t show that. They troll me and joke with me, and they treat me poorly sometimes and all of that, and I think that that’s a good sign of a culture. We’re hiring platform people—people to build out Jev everywhere. We’re so much more sensitive to location because speed of light is more of a bottleneck.
Mm.
Right? I’m so sad for the European users that they were only 3 times as fast instead of 100 times as fast because we don’t have servers there right now, and that’s insane, right? But, like—
It’s okay. Life in Europe goes a bit slower as well. It’s okay.
Wow. I can’t believe you said it, not me. Or everywhere.
Yeah.
You know, if intelligence per second is a metric that matters, we’ll launch this all over the place. If you’re a developer building on top of us, I care a lot about you. We’re hiring people to keep building more—not just—the goal is not to just be Jev as a company. The goal is to ship more shapes of intelligence beyond that.
We’re hiring people to build those things too. We want to not just be the one-trick pony of the simple model. But I think that there’s going to be an AWS of intelligence, you know? And—
Which is going to be you, by the way, right? Yes.
I mean, that’s a direction I want to go down.
Yes. Okay.
It would be arrogant to say it will be me.
Yeah.
I’m going to do anything I can to make sure that happens.
Yeah.
I think that that’s going to be so, so cool. We’re playing with System 1 intelligence right now. Imagine the layers, you know? This is the TCP of it.
Mm.
Yeah.
Yeah. Several more layers to go.
Yep.
And who knows what else? I’ve also pitched temporal, by the way. I don’t know. We need to talk about temporal as layer 8—
Ooh.
—out of the 7 layers.
Ooh.
But anyway, we can talk forever.
Hell yeah.
You’ve got to get back to work or sleep.
Yep.
Thank you for coming.
Oh, boy. Yeah.
Yeah.
Cool. You’re most welcome. It was a pleasure, man.
Yeah.
So excited.
Yeah.
So excited.
Not the last time.
My first time.
Not the last time.