[BidClub_]
20VC · · 74 min

Inference Will Be $10TRN Market: What No One Sees | Software Investors are F***ed | Naveen Rao

Harry StebbingsDiogo Almeida

VC/PEAI & SoftwareTechnicalInvesting
YouTube ↗
TL;DR
  • Almeida argues that inconsistent judgment is the main obstacle to AI automation. Getting “Einstein maybe 95% of the time, and Mr. Bean the other 5%” leaves models unsuitable for accounting and other consequential background work. Jev’s proposition is dependable decision-making; cheaper, faster inference is secondary.
  • His hierarchy is the right task, data, compute, then algorithms. “Data matters a lot more than compute,” but useful training data often has to be created around a new objective rather than retrieved from enterprise databases. He treats GPT-4.5 as evidence against indiscriminate pre-training scale, while acknowledging that his inference about its size depends on pricing reflecting compute.
  • He expects workflow engineering to create proprietary intelligence without every company training its own model. Incumbents could benefit from a “SaaSapalooza” because they understand customer problems, although victory is not guaranteed. Harry strongly challenges his characterization of Harvey; Almeida acknowledges limited visibility and argues that copilots still fall short of autonomous execution.
  • His “mega Jevons paradox” requires successful automation before it produces higher spending. Cheap tokens initially make projects affordable; reliable results must then justify expanding budgets. He expects background decision models to dwarf text LLMs in token consumption, while conceding that Jev and competing approaches remain at “roughly 0%” on the enormous task of real automation.
  • Almeida doubts American open-source models will catch Chinese models under their respective ethical and legal constraints. “My guess is no, unless China decides to stop trying,” he says, while allowing that a capability plateau could make the gap immaterial. He would not use a Chinese model unmodified in a coding agent, but favors examining specific threats, training and safeguards.
  • He is preparing for an AI bubble burst caused by disappointed demand. FOMO and “overpromise, under-deliver” can reverse spending when customers discover capability limits. He thinks reinforcement learning can make models smarter, but says he would guess much current investment is unjustified, at least for now, and considers many new labs wasteful.
  • The financing story combines extraordinary reported adoption with Almeida’s claim that TypeSafe is profitable. Harry cites usage by about a third of the Fortune 500, Pat Grady’s claim of zero to $100 million in revenue in seven days, and a $7.5 billion valuation. Almeida says the objective was “allies to fight the largest companies in the world,” rather than maximizing money raised or valuation.
Digest · the substance, structured for research

1. Reliability blocks the jump from assistance to automation

  • Almeida distinguishes uptime from “reliability of intelligence.” Models can perform extraordinary mathematics yet remain unsuitable for accounting: “You get Einstein maybe 95% of the time, and Mr. Bean the other 5%.” The variance makes the expected outcome unacceptable for consequential automation, even when peak performance is impressive.

  • His engineering standard is effort up front followed by software that runs in the background and solves the problem indefinitely. That requires “nines of intelligent reliability.” He says companies have repeatedly failed to automate accounting because capability is not dependable enough: “It’s not a diffusion problem in the slightest.”

  • He points out that ChatGPT reliably produces recognizable conversation: it does not return random characters, and a request for a polite email will not produce slurs. In his view, the errors occur in decisions within the text. His diagnosis is objective selection: RLHF optimizes pleasing humans, while RLVR optimizes benchmark performance. Neither, in his framing, directly targets dependable background software.

2. Task design determines which data is valuable

  • Almeida’s “bitterest lesson” ranks “doing the right task, then data, then compute, then algorithms.” He says the scale-alone pre-training narrative broke at GPT-4.5. His caveat matters: assuming lab pricing roughly reflects inference compute, it remains the largest model by token cost, yet was not commercially successful. He does not claim visibility into the labs.

  • Defining a new task often means creating a dataset that does not exist. When developing RLHF, he says, human comparisons of LLM outputs were neither sitting on the internet nor inside company databases. The team had to define the instructions, recruit labelers and iterate. Enterprise data lakes may contain information without having the shape needed to improve the intended task.

  • His explanation of RLVR emphasizes hidden reasoning rather than verifiable rewards alone. People tried reinforcement learning against unit tests in 2021 without reasoning, and it kept failing. Task structure changed the result. Useful examples must also track the frontier model’s capabilities: the right difficulty depends on what will actually improve that model.

  • Synthetic data is “more of a tool than an idol to pray to.” Jev uses a lot, but Almeida says the important distinction is how it serves reliable intelligence. Harry sees a vast market for data suppliers; Almeida agrees data remains necessary but cannot estimate sustainable spending once customers must operate as real businesses.

3. Jev’s defense is attention to the failures benchmarks conceal

  • Almeida says people misunderstand Jev by emphasizing price and speed. The central question is: “Why is it 97% accurate? What was the missing piece?” Sometimes the answer is a smarter model; often it is prompting, language ambiguity or developer education. Winning a benchmark does not eliminate the obligation to investigate failures.

  • Harry’s pushback is direct: OpenAI and other providers want enterprise revenue, so how could they neglect reliability? Almeida replies, “You have no fucking clue how little they care.” His argument concerns incentives: monthly active users, and now analogous coding-agent metrics, reward different behavior from a dedicated decision model optimized for dependable judgment.

  • He currently feels there is “no real competition,” while expecting that eventually to change and hoping serious rivals emerge. His own change of mind concerns hype: launch excitement initially seemed crucial, but after other products launched, people Almeida spoke with said they had tried them and found them lacking. That made “underpromise and overdeliver” look more viable than he had expected.

4. Reliable AI could change software rather than decorate it

  • Almeida’s frustration is that writing software has changed dramatically while software itself largely has not. His analogy is the blue LED: adding it to red and green did not simply enlarge the LED market by 50%; it enabled screens and efficient white lighting. The missing component expanded what could be built.

  • He hopes AI becomes a “fourth fundamental logic gate.” Adding a composable primitive alongside the three fundamental logic gates software is built on should expand possibilities by much more than 33%—potentially “hundreds of thousands of times.” His ambition is to turn previously impossible automation into an engineering-effort question.

  • Today’s enterprise AI interfaces are logical responses to unreliable models, he argues. Companies avoid putting them inside consequential business logic, instead attaching assistants whose outputs users must check. “All of the burden of checking” lands on the user. Enterprises have incentives to adopt AI; aggressive adopters have also been badly burned.

  • Almeida rejects a “SaaSpocalypse,” preferring Martin Casado’s “SaaSapalooza.” Incumbents have workflow knowledge and resources for substantial engineering, although they will not necessarily win. His example is TurboTax: understanding exactly where an interface should change is valuable expertise that frontier labs lack. The advantage comes from proximity to problems, rather than distribution alone.

5. Proprietary intelligence can reside in application code

  • Almeida calls AI coworkers the “horseless carriages” of this era. Automation should exploit its own strengths rather than reproduce a human role. His industrial example is twisting a bottle cap onto a bottle: automate the narrow operation and solve it permanently, without waiting for a complete humanoid coworker.

  • Harry cites companies building proprietary models through Fireworks. Almeida allows that narrow models could become cheaper and more specialized, especially when capabilities plateau. But he warns that internal teams face incentives to demonstrate superiority and defend their strategic relevance. He does not know how good those particular models are.

  • His preferred route to company-specific intelligence is software surrounding models. Engineers dislike updating dependencies; once automation reliably works, they should want to leave it alone until an upgrade offers a substantial benefit. The database analogy is deliberate: the important intellectual work lies in how the database is used, rather than simply which database a company owns.

  • Harry strongly disputes treating Harvey as a thin wrapper, citing legal specialization and deep product functionality. Almeida acknowledges he lacks a license and current visibility. His understanding is that larger autonomous-project ambitions became a lawyer copilot. Domain engineering is valuable, but he wants to move beyond double-digit efficiency gains toward background work taking weeks—potentially organized very differently from human work.

6. Coding progress does not establish broad gains in intelligence per dollar

  • Almeida rejects “dollar per benchmark” as a meaningful measure. General pre-trained intelligence per dollar of compute, to his understanding, is “barely moving right now.” Coding performance can improve substantially because coding is increasingly represented in training—not because the underlying model has become uniformly more intelligent across every activity.

  • He places the largest coding-agent jump around November last year, without remembering which Opus release. Subsequent discussion of Fable 5.2 or Opus 5.5 looks like diminishing returns to him. Writing, by contrast, does not seem better over the preceding nine months, though he explicitly allows: “Maybe I’m writing with them wrong.”

7. Benchmark parity leaves open-source competition unresolved

  • Almeida considers Anthropic models the best despite Chinese models matching benchmarks and Anthropic not occupying the frontier of many published benchmark comparisons. Public intelligence measurement is poor, he argues, and people without time to try models are especially vulnerable. Labs have internal proxy measures but little incentive to publish their best measures.

  • His distinction is about incentives rather than a duty of honesty. OpenAI and Anthropic are “mostly truthful,” he believes, because they can genuinely compete at the top. Second-tier labs face pressure to deceive themselves; he cites Llama 4, Google and Chinese labs. Pursuing the “American open-source leader” consolation prize can encourage particularly extreme benchmark optimization.

  • Harry raises Chinese-model backdoors and prompt injections. Almeida says, “I would not use a Chinese model in my coding agent vanilla.” But that is one threat model: retraining, measuring bias, guardrails and separate decision models can change the risk. He cites a use case for Jev that the transcript marks ambiguously as “Jev [?]”: classifying whether tool calls are safe, including whether they might phone home to China.

  • Asked whether American open-source models will catch up, his answer is explicitly a guess: “No, unless China decides to stop trying.” He attributes the difficulty to his understanding that American models are held to a higher ethical and legal baseline, rather than to a lack of capability. A hard enough plateau might nevertheless make the remaining difference irrelevant. Treating one unmodified coding-agent model as the only possible architecture imposes unnecessary “horse blinders.”

8. An AI bubble can burst through weaker demand

  • Almeida dislikes the new-lab playbook of hiring former frontier researchers, buying GPUs, experimenting and eventually getting acquired: “Throwing money into the furnace.” He sees value in reinforcement learning because models are getting smarter, despite diminishing returns. He would not make those investments himself; perhaps a 1% moonshot probability works for venture mathematics, but his expected-value judgment is that current spending is not justified, at least right now.

  • He has believed AI is a bubble “for a while” and plans for a burst. Harry sees no demand shock; Almeida argues FOMO itself generates demand. Customers who overestimate capabilities can reverse course when results disappoint. He points to businesses that cut workforces on AI expectations and subsequently had to backtrack.

  • Harry’s counterargument is that even frozen model capabilities leave enormous value unexploited. Almeida agrees with the spirit—“the models are so fucking smart”—but disputes whether companies can reliably extract that value through current interfaces. The experience of aggressive adopters getting burned complicates the assumption that remaining headroom automatically becomes durable demand.

  • Harry cites Cognition as a fourth-place coding-agent business with $1.6 billion in revenue. Almeida declines to forecast the market’s eventual size: successive AI booms have repeatedly risen and fallen, and this cycle need not rise indefinitely. For TypeSafe, efficiency also offers a way to stay on the sidelines of neocloud and GPU-market turbulence.

9. Jevons requires reliable projects before expanding budgets

  • Harry asks how extremely cheap usage can support a $10 billion revenue business. Almeida’s answer is a causal chain, not automatic elasticity: lower costs make a project affordable; reliability makes it succeed; demonstrated value justifies more projects and eventually a larger budget. Simply reallocating an unchanged AI budget across cheaper calls is insufficient to establish Jevons paradox.

  • His deliberately humbling metric is real automation: “Jev, Fable, and Astra are all tied at roughly 0%.” There are isolated successes, but the remaining opportunity is enormous. He sees background AI as an industry still being created, potentially much larger than coding agents making software perhaps 20% more efficient to write.

  • Almeida predicts “mega Jevons paradox” and decision-model token consumption dwarfing text LLMs. Asked whether energy and infrastructure can support it, he answers yes: Jev serves efficiently rather than following a subsidized playbook. Further efficiency research remains necessary, alongside developers discovering useful applications. He identifies those technological and application boundaries as the central bottlenecks.

10. The financing bought allies rather than a spending mandate

  • Harry’s introduction supplies the headline figures: Jev is used by about a third of the Fortune 500; Sequoia’s Pat Grady says it went from zero to $100 million in revenue in seven days; the valuation is $7.5 billion. Harry identifies Andreessen’s Martin Casado as leading the round, with Sequoia also investing.

  • Almeida describes an inbound process without a deck. Instead, investors reported how portfolio companies were using Jev, giving TypeSafe a valuable view of adoption it could not otherwise see because it does not inspect customer data. Changes in investor excitement over the first days and weeks became an additional market signal.

  • He says TypeSafe is profitable and cash in the bank is increasing. Investors expressed interest even at three times the valuation, but the company did not need to maximize either money or price. “What we need is allies to fight the largest companies in the world.” Raising capital does not imply a plan to spend it immediately.

  • Andreessen appealed through philosophical alignment and enthusiasm for technology. At Sequoia, Almeida emphasizes Sonya Huang’s distinctive strategic perspective more than the institution itself. That fits his unconventional company value: “Make great decisions slowly.” He wants the reasoning behind consequential choices, rather than business activity performed on autopilot.

11. A different research culture could unlock different software

  • Almeida’s hardest hiring problem is the data domain team. A new task requires a new shape of data, creating something like an internal research subfield. He wants brilliant people who may not identify as ML researchers to realize they can learn from its experts and directly make the model smarter.

  • TypeSafe deliberately pays well without matching the largest lab packages: “We don’t want mercenaries.” His concern is that extreme compensation attracts hired guns and encourages candidates to perform cultural alignment. The trait he wants across roles is “technophilia”—a shared love of technology that sustains the ambition to deliver “technological shock and awe.”

  • His proposed “software researcher” role would challenge the status ladder placing ML researchers above ML engineers and software engineers. Exploring AI-powered software and new interfaces is his most underhyped direction. Robotics is his instinctive choice for most overhyped, because reliability and other prerequisites remain far away.

  • His five-year speculation extends to principal-agent problems, including electorates and representatives or shareholders and companies. AI might help express and enforce intent more directly. He argues that people may agree on underlying aims while disagreeing over the detailed rules needed to implement them.

Full transcript
Diogo Almeida

Almost everyone who’s paying attention to AI knows reliability is the absolute largest bottleneck. You get Einstein maybe 95% of the time, and Mr. Bean the other 5%. Data matters a lot more than compute. I don’t want to be called a neo-lab because, from my perspective, most neo-labs are incredibly wasteful.

Anthropic models are just the best, even if Chinese models match them on benchmarks. It feels like there’s no real competition right now. I truly believe we’re going to see a mega Jevons paradox. AI is so clearly awesome, yet software itself hasn’t changed. Everyone has been wearing mono-model horse blinders for so long.

I’ve believed AI has been in a bubble for a while, so I’d like to keep planning for the possibility of a bubble burst. What we need are allies to fight the largest companies in the world. We don’t need the money or the valuation.

Harry introduced Diogo Almeida, founder of TypeSafe, maker of Jev. He said Jev was used by about a third of the Fortune 500, and cited Sequoia’s Pat Grady saying it went from zero to $100 million in revenue in seven days. The company had been valued at $7.5 billion in a round led by Martin Casado at Andreessen, with Sequoia also investing.

Diogo Almeida

I’m so looking forward to this. We had a brilliant exchange following Eric from Modal’s show. I actually spoke to him last night about you, so I’m very excited for this. Welcome to the show.

Harry Stebbings

Thank you. It was a very small exchange, but I’m glad you liked it. It was really fun, and I didn’t realize it would become this.

Diogo Almeida

We said it would be a spicy discussion. Before we get into the spice itself, I want to lay some foundations. In 60 seconds, for people who don’t know what TypeSafe and Jev are, can you give us the 60-second intro?

Harry Stebbings

I’d argue that Jev isn’t a specialized model. My pitch for people who aren’t purely software developers is more about the state of AI. How can AI be so smart at so many things, like solving Millennium Prize Problems in math, yet not transform software at all? All the SaaS we use is basically the same as it was a decade ago, and we’re not automating almost any entire job.

We keep failing at things like drive-throughs. There’s AI customer service, but it can’t be trusted with the big decisions. How can that be the case? It seems so good at some things and so bad at others. What we’re doing with Jev is trying to bridge that gap and make AI too good to be true for automation, instead of just human-in-the-loop assistance.

Diogo Almeida

I spoke to Sonali at Sequoia, and, speaking of full end-to-end automation, she asked, “Why is reliability the biggest obstacle to automation?”

1. The Reliability Bottleneck

Harry Stebbings

I’ll give a caveat about reliability first: I don’t mean uptime, availability, or throughput. I mean reliability of intelligence, which I’m still trying to find a good word for. It’s this non-jaggedness of intelligence.

Almost everyone who’s paying attention to AI knows reliability is the absolute largest bottleneck. Here’s an example: How good are you at math? How good am I at math? There’s no chance we could even help with a Millennium Prize Problem solution, I believe. I haven’t read one, but I’d wager that’s true. And yet these same models can’t even be used for accounting, which is much simpler stuff. The question is, why?

These models are extremely high variance. People love this saying I’ve coined: You get Einstein maybe 95% of the time, and Mr. Bean the other 5%. The problem is that the expected value doesn’t work out. Companies quite wisely learn that you can’t actually automate things. I’ve seen so many companies try to automate accounting over the multiple AI hype cycles of the last four years, and they keep failing because the math doesn’t work out.

It’s not a diffusion problem in the slightest. Reliability is what gets you there. Maybe you can trade off some level of Einstein for getting it more consistently, with the same functional intelligence. That, to me, is Jev’s superpower.

Harry Stebbings

How do we get there, then? What’s the solution to removing Mr. Bean, as much as I love Mr. Bean?

Diogo Almeida

I love Mr. Bean, too.

Diogo Almeida

I love when agents are insane. “Bad Vibes Forever.” It’s a very fun story, and that’s not what I want in my software that runs in the background.

As an engineer, I want to write a thing. I want to spend a lot of effort up front. Maybe I don’t realize it’ll be a lot of effort, but it always is. Then I want it to run in the background and solve the problem forever.

You just need nines of intelligent reliability to do that. I’d argue that the way to get there… I wish I could bring up my slides here, but I’ll try to use my hands. I call it AI’s bitterest lesson.

Sutton coined the bitter lesson, which is roughly that compute matters more than algorithms.

Diogo Almeida

This is the scaling era we had from 2019 to 2022, to 2023: just scale, and the models get better.

2. Data Beats Compute

Diogo Almeida

This kind of plateaued at GPT-4.5, and then we kind of stopped scaling as a field because it stopped paying off. I’d argue there’s a bitter lesson: data matters a lot more than compute. This is obvious to anyone who’s done applied ML, but people in the peanut gallery don’t seem to realize it. Data is obviously more valuable than compute.

The most important thing by far is your North Star: What do you define as the right task? Before Jev, the two right tasks that existed in LLM land were RLHF—

Harry Stebbings

Mm-hmm.

Diogo Almeida

—which is reinforcement learning from human feedback, meaning “please humans,” and RLVR, which means—

Harry Stebbings

Mm.

Diogo Almeida

—score high on benchmarks. Neither of those actually translates to software running in the background.

Mm.

And I would argue that this is the root of all the problems. It’s not that the AI isn’t capable, because it’s clearly super capable. The problem is that—and you’re experienced in the real world. Everyone who’s experienced the real world knows this: you get what you optimize for, right? We just happen to optimize for the human in the loop.

Harry Stebbings

Oh, thank you so much for bluntly giving me so many follow-ups in one shot.

Diogo Almeida

Oh, can I say one thing? Can I say one thing about this, too? On the note of reliability, ChatGPT never fails at being ChatGPT. When’s the last time you got random characters, right? It’s kind of odd that it… Or even if you ask for a polite email, it’s not going to start saying—I probably shouldn’t curse on a podcast—but it’s not going to start saying slurs, right? Humans might make that error sometimes.

The problem isn’t that it’s not impolite or that it’s sycophantic. The problem is that it makes wrong decisions within the text. It’s interesting that it can be reliable at some stuff, but not others. If you optimize for the important stuff to be reliable on, which is decision-making and judgment, that’s how you get real reliability.

Harry Stebbings

You said there about the plateauing of compute in terms of the benefits that accrue. First, have we got to that stage? Are you serious? I thought the whole point was that—

Diogo Almeida

On pre-training?

Harry Stebbings

Purely on pre-training, sure. But more and more compute still leads to better and better output, no?

Diogo Almeida

Yes, but you need a lot more data than people once thought. Maybe people have warped this to seem like they were right the whole time, but the narrative back in the day was just, “Scale to the moon and you will get AGI.” GPT-4.5 is what broke that narrative.

I obviously don’t know what’s going on behind the labs, but if we believe the labs were pricing roughly based on inference compute, then GPT-4.5 is still the largest model ever by token cost, and it was not commercially successful. I think this is obviously the case because you can’t just scale compute without really thinking about it. People are throwing a lot of compute at training, but you need a lot of data with it.

This is why you see so many “get rich quick by selling data, RL data to labs” type of things. You can’t really do very much without that data. If you saw—oh, man, I believe it might have been Dwarkesh—he had a recent Twitter post that showed what matters more: data or compute or algorithms. Data was many, many times more effective. It was something this week. This is something everyone has known since the beginning of machine learning. It’s not even specific to AI. It’s just that the compute-scaling era was a weird outlier in this.

Harry Stebbings

When we say the value of data, or how it’s so much more primary than compute, is that synthetic data? Is it proprietary enterprise data that everyone has in their companies? Is it AI systems generating their own experiences? Is it certain types of data, or is it just broad data? Can you help me understand that?

Diogo Almeida

I wish I had my visuals here. Let me work backwards again from my bitterest lesson, which is, in order of importance: doing the right task, then data, then compute, then algorithms. This is what we extremely, extremely learned when we made RLHF. The quality of your underlying algorithm matters very little as long as you do the right task. A GPT-2-sized model could beat something like GPT-9, which doesn’t even exist.

In order to have data, you actually need to know the right task. The data for the task—or normally, when you create a new task, the data doesn’t even exist. When we made RLHF, human preference data for comparing LLM outputs didn’t exist anywhere on the internet. It wasn’t in a company database. We had to define it, understand the instructions, get human labelers, iterate, and figure out what that data should look like.

Harry Stebbings

Mm.

Diogo Almeida

Doing the right task is what unlocks the data in the first place. The problem is that bootstrapping a new task is hard because, by construction, the data doesn’t exist.

Harry Stebbings

Mm.

Diogo Almeida

It’s a similar thing for us, and it’s also similar for RLVR. Before RLVR existed—and I actually want to point out a tricky thing about what the right task of RLVR is—RLVR means reinforcement learning from verified rewards. Sorry if I’m going a little bit too advanced here.

Harry Stebbings

No, no, no. Go ahead.

Diogo Almeida

RLVR is not at all about the verifiable rewards. The big part of RLVR is this latent chain of thought, or hidden reasoning, so the models can think in private and then answer. Even in 2021, there were verifiable rewards. People tried to do RL, like PPO without reasoning on unit tests, and it kept failing. So it wasn’t at all about the verifiable rewards. It’s about the structure of the task, and it just turns out that once you’re doing RLVR, the models are a lot better at this.

What good data means is now a function of how good the models are. This is why the people who make data need access to cutting-edge models, because what actually makes the data good is whether or not it improves those models. They need to be at exactly the right level of difficulty to push the models forward. There’s a lot of nuance and on-policyness of data.

But going back to the original question, is the right data stored in corporate data lakes? I’d argue no, because it’s probably not shaped the right way. People are definitely trying to do this, so a lot of data vendors are trying to yoink this data and sell it. Maybe they can massage it to make it work. But the most important thing is focusing on how we use this thing to push our task forward. If you forget that, you are Goodharting a local objective.

3. Data Businesses and Synthetic Data

Harry Stebbings

How do we use this thing to push our task forward? Are the data-providing companies, Surge and Macaw, good businesses? Are they services businesses now, which have to go end-to-end, where they really get into implementation and the effectiveness of the model, not just data serving like they did in the early days? Are they good businesses?

Diogo Almeida

That’s more of a you question than a me question. Do you think they’re good businesses?

Harry Stebbings

I do, personally, because I think the market is just so large for the amount of money that these LLMs will spend on data. And I think if you believe we’re going to move into a world where companies own their own intelligence, that data will always need supplementary data. So then the TAM becomes moving from LLMs to everyone.

Diogo Almeida

Interesting. That might be a fun debate: whether or not companies will own their own intelligence. I’ll actually argue that we’re still in hype-wave territory. There’s always a trough of disillusionment, and then the reality. I don’t think we’ve hit the reality part of it yet.

In the future, I believe both RLHF and RLVR are immensely valuable. I don’t know if people will invest as much as they are right now, but I think each of them results in its own… Even if RLVR never existed, ChatGPT would be an incredibly valuable consumer brand. You’ll need data for that. Data will always be necessary, but how much that company will be willing to spend on data when it actually needs to be a real, sustainable business—I don’t know what that number is. It’s very hard to ballpark.

Harry Stebbings

What value does synthetic data hold in model improvement?

Diogo Almeida

I don’t think there’s one form of synthetic data. I think synthetic data is more of a tool than an idol to pray to.

It depends on the hands of the team. If you read the research, you’d think synthetic data is mostly overhyped because it’s only used to make small models catch up to large models, but it can’t be used to improve large, state-of-the-art models. So it really depends on how you use it. We use a lot of synthetic data, but we don’t use it the way other people do. We think the more interesting part isn’t that it’s synthetic, but how we wield this tool with an extreme laser focus on the right task, which is the reliability of intelligence.

Harry Stebbings

You said about the hype and the trough, my friend. There is insane—

Diogo Almeida

It’s pronounced “trough.” That’s good to know.

Harry Stebbings

No, I tried to do it elegantly, okay? I’m British.

Diogo Almeida

Yeah, yeah.

Harry Stebbings

It was your charm.

Diogo Almeida

Oh, yeah.

Harry Stebbings

Are you concerned by the amount of money going in, or do you think it is justified?

Diogo Almeida

I am concerned for the world, I think. Throwing money into the furnace is a bad thing for the world. I really don’t want to be too negative, but I don’t want to be called a neo lab because, from my sense, most neo labs are incredibly, incredibly wasteful. It feels like their business model is to hire a bunch of ex-lab people, get a bunch of GPUs, have fun, and then eventually get acquired.

That seems like a waste of resources to me. I’m not a fan of that, on one hand. On the other hand, people are throwing a lot of money into RL, for example, and I actually think that’s kind of cool. There are massive diminishing returns, but the models are getting smarter. It’s kind of good for the world. I think it’s very cool.

Would I make that investment with my money? No, obviously not, because we have money to make bets, and we’re obviously not going to do that. So I’m not a fan of that in the world either. But maybe the VC math works out where there’s a 1% chance of a moonshot, and it actually has positive expected value. And wait, you had a question like, am I concerned, and there was a—

Harry Stebbings

Are you concerned, or is it justified?

Diogo Almeida

Oh. I would guess, based on my expected value calculations, that it’s not justified, but I’d be very happy to be proven wrong—or at least right now. Yeah.

4. What Jev Gets Right

Harry Stebbings

Did you expect it to work as quickly as it did?

Diogo Almeida

For years now, even before TypeSafe existed, I thought I must be missing something about the field. Obviously, I convinced myself I wasn’t missing something, but this direction seemed so weird to me. It was unbelievably obvious. So obvious that I gave up when I first came to the idea because I thought, surely Anthropic is already 80% done. It’s the most obvious thing I’ve ever heard of.

So to me, it’s really obvious that this would be valuable in the world. I still think people don’t get it, even though the launch was obviously phenomenal. The part I didn’t expect was the ChatGPT-for-developers moment.

Harry Stebbings

Mm.

Diogo Almeida

Even non-developers were talking about it. That I could not have predicted. That was—

Harry Stebbings

What do—

Diogo Almeida

—super weird.

Harry Stebbings

What do people not get? What do people get most wrong about Jev and you?

Diogo Almeida

My team hates when I talk about this, but my rule is that on a podcast, I can be let off the rails. I believe it’s a competitive disadvantage for me to talk about what Jev is really doing, because people keep missing the mark. If we just didn’t tell anyone, we’d forever be king of the hill. I care too much about the field not to yell, in painful insanity, “What the fuck’s going on?”

The thing people really don’t get is that what’s special about the model isn’t the cost or the speed. It’s, again, the reliability of intelligence.

Harry Stebbings

Mm.

Diogo Almeida

People are still thinking in LLM land. So many people, I think, are burning their goodwill. There should be copycat models in the future, but in my opinion, they should give a lot of fucks about reliability. The people shipping copycats right now don’t. They’re trying to do the same thing as in LLMs: maximize benchmarks, get hype, and show little things like numbers that aren’t relevant, such as speed and cost. What they really need to do is hunker down in the bunker and make the models really, really smart.

People don’t realize that’s what we’re all about. When we see benchmarks, even if we’re unbelievably in the lead, we’re still pained when our model is dumb. That’s what we focus on, and I think it’s what everyone in the field should focus on. Ask, “Hey, why is it 97% accurate? What was the missing piece?” To be clear, it’s not always a model thing. Very often, it’s a prompting thing.

But that’s how it is to build a developer tool, right? We need to educate people. We need to figure out how we could prompt this better. We need to figure out whether there’s some nuance in this language that could help people make more robust prompts, or whether we need to make the model smarter. This is the thing people really miss.

I think developers get it, and they got it way faster than I thought, because they already don’t give a shit about the benchmarks. They’re just like, “Your model is clearly the best at new stuff.” But the same people who say, “Oh, this is really boring. It’s just a classifier,” ended up making the copycats. They don’t understand that there’s a reason you can put so much effort into making the model more and more and more robust.

Harry Stebbings

You said “copycats” there. Can I be blunt?

Diogo Almeida

Oh, yeah.

Harry Stebbings

OpenAI and others—

Diogo Almeida

Yeah, please do.

Harry Stebbings

—have moved very quickly to create their own reasoning model products after you launched. Why don’t the big model providers just copy you and compete you out?

Diogo Almeida

I don’t think that’s a blunt question at all. I think it’s a totally fair, reasonable one. If they one day started caring a lot about reliability, that would become an interesting battle. So—

Harry Stebbings

Diogo, how do they not care about reliability? They want enterprises. They want—

Diogo Almeida

You have no fucking clue how little they care.

Harry Stebbings

What? How? They want enterprises. They want revenue. How are you going to get a CIO to sign up if you don’t have reliability?

Diogo Almeida

I think people have a hard time understanding intelligence as a product, which is why they use benchmarks as Band-Aids. Benchmarks are kind of a proxy for understanding the intelligence of the underlying thing. But as far as I can tell, the field over the last nine months has learned that benchmarks don’t mean shit. Anthropic models are just the best, even if the Chinese models match them on benchmarks. Even when you measure benchmark-maxing, you still think the Anthropic models are better than the competition, even though they’re not on the Pareto curve for pretty much any benchmark.

Basically, it looks like the subfield of System 1 or decision models learned this in—I think competition maybe started a week later. It felt like people learned it in two or three weeks. We were like, “Oh yeah, the other stuff is just benchmark-maxing. It’s total crap.” It’s very easy to see that it’s different. I think it takes a different type of organization to, again, really care about reliability.

Harry Stebbings

If they don’t care—

Diogo Almeida

Go ahead.

Harry Stebbings

—about reliability, what do they care about then, Diogo?

Diogo Almeida

It really depends on their business objectives, right? Back in the “ChatGPT is everything” days, before Claude Code, they cared about monthly active users. That’s a very different metric to optimize for than the best chatbot experience. That’s how sycophancy emerged, and it kept going in a cycle: “Oh, it’s more sycophantic. It lies sometimes, but hey, that’s what people want, so we just keep doing it in some pockets of space.” You get what you optimize for. That was what ChatGPT optimized for.

I think now people are optimizing for the MAU equivalent of coding agents. If they’re not training a dedicated model for this and putting a lot of intense thought into it, I don’t think they’ll be able to make a comparable model anytime soon. I hope they all do, because I think it’d be very good for the world. It’s obviously going to be bad for the business, but right now it feels like there’s no real competition. Surely that will change eventually, right?

I hope we get competition and show that we’re the best because we care the most, and because we have the smartest people working as hard as possible, laser-focused on the North Star. But again, this isn’t competition yet.

5. Jevons and Software’s Future

Harry Stebbings

When you look at the decision model product landscape for models that you’re in, what does that look like in three years? Is it an Uber/Lyft market? Is it a GCP, AWS, Azure market? What does the market composition look like in three years’ time?

Diogo Almeida

I’m a bit of a technological idealist, I think. I don’t really think about what the competitive landscape is going to be like. I care more about the technological one. Our strength is always going to be technological shock and awe, so we’re going to have to keep shipping awesome stuff.

The way I see that landscape is that I’m going to ignore competition. In terms of number of tokens, I truly believe we’re going to have a mega Jevons paradox. It’s not going to be a subtle Jevons. It’s not going to be a fake Jevons, like LLMs. It’s going to be super-duper. There’s going to be tons and tons and tons of usage, and I think it will completely dwarf at least the total token consumption of string LLMs.

I think these things are going to be used everywhere, and we’re finally going to have AI running in the background. I think we’re going to start to see a shift in how people feel about AI, from this thing that slops crap at them to something that actually makes life better, because people truly like it when technology improves their lives.

Harry Stebbings

When you say “mega Jevons paradox,” do we have the energy and infrastructure to power the demand of a mega Jevons paradox that you foresee?

Diogo Almeida

Yep.

Harry Stebbings

And what are the bottlenecks to your tech utopia?

Diogo Almeida

Our models are incredibly efficient in comparison. We’re not doing the standard subsidized playbook.

We actually can serve these things incredibly, incredibly efficiently.

Harry Stebbings

Yeah.

Diogo Almeida

There are many boundaries to push on that efficiency angle. Of course, we’re going to have to do some sick research, but that’s the thing I trust us to do. We’ll have to push efficiency quite a bit in order to have that mega Jevons thing occurring.

I think the largest bottlenecks are going to be… It feels a little bit single-player right now. It’s on us to push the boundaries of the technology from an AI point of view, and from a field point of view, we’re going to need to interact a lot with developers. We’re going to need to help them push the boundaries.

This is what I’ve been talking about, a role I dream of. I’m so sick of the status games, especially in SF, where ML researchers are a higher tier than ML engineers, who are a higher tier than software engineers. I think that’s completely messed up.

My hope is that we can create a new type of role that I call… My current favorite name is “software researcher”: someone who actually pushes the boundaries of software. If you knew you were in the Cambrian explosion of the internet or mobile, you’d think there were so many boundaries to push and so much low-hanging fruit that you’d need trailblazers to go out there and build these lighthouses, saying, “Hey, there’s value here. Let’s have a gold rush there.” I’m so excited about encouraging people to push those boundaries. I wish I had the time to do it myself. There are so many crazy, exciting things to do.

I think those two are going to be the biggest bottlenecks to get to System 1 models everywhere and AI actually automating shit. If I could say one more thing: I don’t know if you’re a software engineer, but I guarantee every single person who’s ever been a software engineer has had a conception in their head of “I can automate this” or “I can’t automate this.” You need to know, right? If you wanted to automate something that moves files around, you can automate this. If you want to make a voice interface for your laptop so you don’t need to type with your hands, you can’t automate this.

My hope is that AI will completely remove the “can’t automate” part. It’s only going to be a matter of effort. I have this entire soapbox about the future of software. I will cook on this undercooked idea.

Harry Stebbings

Yeah.

Diogo Almeida

This is something I find to be such a beautiful analogy for AI. AI is so clearly awesome, yet software itself hasn’t changed. Writing software has very much changed, but the software is the same, and that’s tragic.

I love the analogy of the blue LED. When the red and green LEDs existed, they had very narrow uses. They could barely be used for anything beyond indicator lights because they were very efficient. But when you add the blue LED, it didn’t just expand the market cap of LEDs by 50%. It made all sorts of things possible. We couldn’t have the screens we have today. We couldn’t have white light LEDs that are so much more efficient without this missing composable piece.

My hope is that AI ends up becoming a fourth fundamental logic gate. There are three fundamental logic gates that all software is built on top of. They are so dumb, but it’s the brilliance of software engineers that enables you to use these bitwise comparisons to create everything in software today. What if we could add a fourth one? I think that should expand what we can do by way more than 33%. It should expand what we can do by hundreds of thousands of times, and that is why I’m excited about a Cambrian explosion specifically of software.

6. Enterprise AI and Workflows

Harry Stebbings

Why do you think the current state of software is kind of a tragedy? If you look at enterprise software today, it’s enterprise software with a sprinkling of AI at best. Can you talk about how you feel about that?

Diogo Almeida

Yep.

Harry Stebbings

Can you talk about how you feel about that?

Diogo Almeida

I think it’s actually extremely logical. I don’t think it’s a distribution problem. I think enterprises have so much incentive to adopt AI, and the people who have adopted it the hardest were the ones who were burnt the most.

The argument I always make is that LLMs are unreliable, so they cannot be trusted for decisions with stakes to the company.

Harry Stebbings

Mm-hmm.

Diogo Almeida

But you can always make them make decisions with stakes to the user, because companies tend not to care about their users. Tragic, but that is what happens. So what ends up happening is you don’t have AI deep in the logic of your apps. What you end up having is the same B2B SaaS with AI attached to the side, where it can’t do high-stakes things, but it can shoot slop at the user because people kind of like it because it’s responsive. They make sure that all of the burden of checking whether it’s right or not is on the user’s plate.

This is actually extremely logical for the state of reliability. I think it’s tragic that people wasted so many resources going the roundabout way to get to where we’re at right now. I also think this is the right spot, given how smart and reliable the models are.

Harry Stebbings

Do the biggest benefits accrue to the incumbents who have distribution? When you look at a Salesforce of the world, and then reflect on what you just said about the tragedy of the current state of software, do the benefits fall to those who have distribution, or is there a new generation they accrue to?

Diogo Almeida

Number one, SaaSpocalypse is obviously not happening. Martin Casado coined it as “SaaSapalooza,” and I love that. I don’t necessarily think it means big SaaS wins. I think the people who are best positioned to win are the people who care the most.

SaaS has a massive advantage over AI companies in that they’re actually familiar with the workflows. They have the scale to spend the upfront investment to do the engineering, think about all of the details, and actually solve people’s tasks. That is the really valuable thing here.

I don’t necessarily think big SaaS will always be the winner, but I think they’re best positioned, not at all due to their distribution, but because they’re the ones closest to the problems. They understand: if we’re at—I don’t know anyone over at Intuit, you know—if we’re in TurboTax, where could we change the interface to make it smoother? There’s not a chance anyone at OpenAI or Anthropic knows how to do that, right? They’re still living in their ivory tower and thinking that software just isn’t going to exist as a concept. So—

Harry Stebbings

Right.

Diogo Almeida

—they could be the big winners.

Harry Stebbings

You said there about knowing the processes, knowing the workflows, and the importance of that. I’ve had Lin on the show from Fireworks, and I had Erik on the show from Modal. They both talked about the importance of owning your own intelligence as a company, and the benefits of having this incredible proprietary data—

Diogo Almeida

Mm-hmm.

Harry Stebbings

—and these specialized models on a per-company basis. Diogo, if the importance is in the process and the workflow, surely this is the holy grail, no?

Diogo Almeida

Oh, man. How do I say this elegantly and Britishly? That’s an opinion, I think, that goes along the current rails of how the current world works and thinks about AI, where AI is being posed as a coworker. AI is being presented as a coworker. This is the direction Anthropic is really pushing toward, where it’s kind of the same interface as a human being. I believe this is going to be like the horseless carriages—

Harry Stebbings

Yeah.

Diogo Almeida

—in the AI era. Just like in the Industrial era, people were like, “Let’s just make it look like a carriage, but without the horse,” I believe AI coworkers are a form of that. It just makes sense that if you’re doing automation, automation will have a different strength profile than a human, and you’d want to lean into the strength of automation.

Harry Stebbings

Mm-hmm.

Diogo Almeida

Kind of like in the Industrial Revolution, you don’t need to wait for a whole coworker, like a humanoid robot, to do tasks. You just find the extremely narrow thing, like twisting a bottle cap onto a bottle, and then you automate it, and it just gets solved forever. So I think that’s how work will generally go.

If you assume that AI coworkers are the way to go, which I absolutely do not, and you assume that most companies have the willingness, capabilities, and desire to train them because they need to be trained in distribution for your company, then I think that owning your own intelligence direction can kind of make sense.

I do believe that the labs’ approach of an AI coworker with general intelligence makes a lot more sense to me. Pretty much any company, if you gave them infinite interns that cost 1,000 times less, could probably find a lot of productive use for them. The question we as a field should answer is: Why are we not getting that effect right now? The people who say you need to own your intelligence, their excuse is, “Oh, you need something like continual learning,” or, “You need something more narrow,” and I just don’t think that jibes with reality.

Harry Stebbings

So Fireworks have helped several very top-tier brand-name firms in the Valley.

I think Harvey, Ramp, and Cognition might have gotten some things wrong in building their own models. Are you saying it’s a mistake to pursue that route, that horizontal or alternative models will be just as good, and that it’s basically a futile exercise?

Diogo Almeida

Oof. It’s complicated. I would 100% believe that narrow models could work. Number one, benchmarks are incredibly misleading. The people making them have a lot of perverse incentives to make their models look better internally and, from an external narrative perspective, to avoid seeming like they’ll be completely boned by the large models. I don’t know how good they are. I would imagine they could get cheaper and more narrow. I think this is a sign of maturity, when we’re plateauing at a certain capability level.

Harry Stebbings

Uh-huh.

Diogo Almeida

I think AI tends to improve in more discrete steps. Those steps happen when you start getting data in a particular distribution. I think it’d be hard to deny that the largest jump in coding agent performance happened around November last year.

Harry Stebbings

Sure.

Diogo Almeida

I can’t remember which Opus that was. That was when they put it in distribution. Everything since then has been, “We love Fable 5.2 or Opus 5.5.” It seems like there have been diminishing returns since then. When you’re seeing diminishing returns, you can start doing micro-optimizations, like cost per token.

My view is that these are all symptoms of not reliably solving problems. Software engineers I know, and I’m a software engineer too, hate updating dependencies. We hate moving to the latest and greatest stuff. The only reason we update is because it didn’t work sufficiently before. Otherwise, you’d just lock it away in the basement and only update it when it gets way cheaper.

I believe we have a lot of insufficient automation right now, and that’s what I think the data most suggests. I’m biased, of course. My view is that because of Jevons and Jevons-like models, we’ll start locking in real automation, and people won’t want to update it.

The way companies get their own proprietary intelligence isn’t by having their own trained models, because that’s hard to do. It’s by having software around them. The reason it’s so important for companies like Harvey and Cognition to say they own their intelligence is that otherwise they’re a very light wrapper on top of the LLMs, and it looks like they have no moat.

But software and engineering are phenomenal, even if they aren’t a moat. Investing up front to understand real problems and automate them so they never have to be solved again, so humans don’t have to do that roadwork, is so good for the world. I think that’s a much more plausible way for the companies of the future to progress. Their proprietary intelligence happens at the code level. It’s kind of like how there may be some moat in which database you use, but the real meat is how you use those databases.

Harry Stebbings

Diogo, I’m fascinated. I know the Harvey legal market pretty well. Do you think they’re a thin wrapper? To me, it’s a very specific domain that requires deep product functionality that can only be done vertically, and that’s why they exist and are very successful. I’d push back strongly on calling them a thin wrapper.

Diogo Almeida

Oof. “Thin” is all relative. They’re much thinner than I—oh, that sounds politically correct right now. It depends on how deep I’d want it to be. My understanding, the last time I checked with Harvey—and I obviously can’t even afford a license, so I don’t fully know—is that back when Harvey was working with OpenAI, they had this dream of automating larger and larger projects autonomously. That kind of thing didn’t end up working. They ended up having to pivot to copilots for lawyers, which, as I understand it, they still are.

You can do a lot of domain-specific things as a copilot for lawyers, but there’s only so much engineering you can do around a copilot. Is it domain-specific MCP tools to pull up RAG based on real law? The reality is that people want to automate large chunks of work, not just make people double-digit percentages more efficient, which is where a lot of AI tends to land.

I think it’s awesome that they’re doing a lot of engineering around it. I think there’s going to be a lot more engineering. Even system messages are a really ghetto form of engineering, where you’re programming the LLM to behave in certain ways, but you have no guarantees, no confidence, and no real way to test how effective it is, other than through the very weird evals people do these days.

I’d like more engineering absolutely everywhere, especially in things that are AI-adjacent. I hope we can one day move past copilots and just automate work that takes weeks to do in the background. Maybe that will be shaped very differently from the work current humans would do, because automation tends to require adapting what’s available to the technology.

7. How Model Progress Is Measured

Harry Stebbings

You said that earlier, about the cost of tokens, at the beginning of that two- to three-minute talk.

Diogo Almeida

Oh, shit.

Harry Stebbings

You said—

Diogo Almeida

Am I that bad? Oh, shit.

Harry Stebbings

No, dude. I love it.

Diogo Almeida

Okay.

Harry Stebbings

It’s fantastic. But the trouble is, I have to remember parts of what you say. My question was going to be: when you look at raw intelligence today, is raw intelligence per token actually improving like we’re told it is?

Diogo Almeida

Ooh. I’m curious about your take on that. I actually want to write a blog post on this.

Harry Stebbings

No. I don’t think it is anymore. I think it improved dramatically. I think raw intelligence is still improving, but minutely, almost to the point where you wouldn’t know. It’s kind of like meeting someone with a 180 IQ versus someone with a 190 IQ. You’re not going to tell the difference. So yes, but not significantly.

Diogo Almeida

I actually think even the 180-to-190 IQ analogy is extremely generous for what LLMs are doing right now, because I wouldn’t call them 180 or 190 IQ people. Sometimes they might be 180; I don’t really know. But then sometimes you get an 80, or even worse.

My take is that it’s not a single dimension. Number one, benchmarks don’t matter, so we should toss them out. If you try to look at dollars per benchmark over time, that’s complete bullshit. Number two, I think there’s a general notion of intelligence, which is pre-trained intelligence per dollar of compute, and, to my understanding, that’s barely moving right now.

I do think that as you get more data and focus on different subsets of tasks, you can make models more in-distribution. Back to the Bitter Lesson: you get way better performance within that data distribution. Are we massively moving the needle on coding from an intelligence-per-dollar perspective? Super-duper. But that’s not because the base models are intrinsically more general; it’s because we’re putting coding more in distribution, so they’re more reliable at it.

Harry Stebbings

Mm.

Diogo Almeida

But has writing improved in the last nine months? I don’t think so. Maybe I’m writing with them wrong. There are many other capabilities where there’s absolutely no improvement at all. If you keep scaling up data in coding agent land, it’s not going to improve lots of other forms of data. There are diminishing returns to it. I think it’s complicated and nuanced.

Harry Stebbings

You mentioned the variance in model capabilities: sometimes 180, sometimes 80. When we look at the different models, open source is now getting as close as possible to the frontier in many respects. How do you see the open model landscape?

Diogo Almeida

Oh, boy.

Harry Stebbings

And what do not many people know that they should know?

Diogo Almeida

That last question’s a hard one. Oh, boy.

Harry Stebbings

You can say anything. It’s just me and you.

Diogo Almeida

I think the thing people should know is that benchmarks are total bullshit. All benchmarks are bad. There are perverse incentives, and ironically, the people they hurt most are those who don’t have time to actually play with the models. They’re the ones who get manipulated by the benchmarks, with higher stakes, including for everything open-weight and open-research.

There are ways of measuring intelligence. It’s very hard. Labs don’t talk publicly about it. They have somewhat-proxy measures internally, of course. It’s an unfortunate public state of affairs that there’s so much misinformation and people aren’t incentivized to release their own best measures of intelligence. I think that’s tragic.

Harry Stebbings

Diogo, who bullshits the most?

Diogo Almeida

Who bullshits the most?

Harry Stebbings

You said something spicy on X.

I’ll give you spice.

Diogo Almeida

Oh, okay, okay, okay.

Harry Stebbings

It—

Diogo Almeida

The top-tier labs are incentivized to be truthful, so I believe that OpenAI and Anthropic are mostly truthful. The second-tier labs are incentivized to lie to themselves. They want to think that they can become a first-tier lab. They have a lot of internal pressure to benchmark-max, but they cannot admit it to themselves. Otherwise, their org chart will go ballistic.

This is what happened with Llama 4, where they were actually too on the nose about benchmark-maxing, even though there was so much internal incentive to do that. Same thing for Google right now, and same thing for a lot of the other Chinese labs. I think they’re trying to aim for the top, and therefore they’re not trying to explicitly benchmark-max.

I think a lot of the people fighting for consolation prizes are just doing completely insane levels of benchmark-maxing. One domain, if I were to be a little spicy here, would be the rough domain of American open-source leader. It’s enough of a side consolation prize that they’re not aiming for the top tier, and that is pure bullshit and optics because no one actually uses those models. So who’s going to call them out?

The reasoning-model landscape is so much worse, in my opinion, too, because no one even understands what good performance is, so they’re all benchmark-maxing to the absolute gills. It’s really good for us because no one will believe those benchmarks. It gives us a lot of power over the benchmarks themselves, which I think is optimal for the field. In the ideal world, no one ever cares about benchmarks on any machine-learning model, System 1 or System 2. But there’s so much of that right now.

Harry Stebbings

Let’s unpack this, dude. Do top-tier labs have a duty to be truthful, or is it in their interest to cheat?

Diogo Almeida

No, sorry. It’s not a duty. They don’t need to cheat in order to be high up there, and they want to truly believe in it. Morale is very important. They truly want to actually be top tier, and they can be. So they don’t need to cheat, is the answer.

Harry Stebbings

I think there’s this feeling in the community that we need to have an American open-source leader, and we need to propel this company forward, whether it’s Reflection, Poolside, or Thinking Machines. Do you feel that we need to propel an American open-source model?

Diogo Almeida

Who is “we,” and what does “propel” mean?

Harry Stebbings

I think there’s a consensus that Chinese open-source models are very efficient, and unless we have an alternative to them, Chinese open-source models will be used considerably and threaten some elements of American security.

Diogo Almeida

It’s actually interesting to dig into the threat model. There’s definitely a threat if you run them vanilla, as they’ve been post-trained. But if you’re actually using—I can’t remember how you described it—the private intelligence of companies, in theory—

Harry Stebbings

Mm.

Diogo Almeida

If you’re training the models, you can remove a lot of bias to different extents. In theory, this type of thing is totally doable. So it actually depends on whether the threat model is literally plug-and-playing these models.

Harry Stebbings

So you don’t think there are threats from Chinese open-source models in terms of dangerous prompt injections that you don’t know about, or backdoors?

Diogo Almeida

I would not use a Chinese model in my coding agent vanilla, just to be clear. But that, again, is one threat model. If you live in a world where a coding agent is the only shape of AI, then sure, that is very much a threat.

But the alternative we need to take into account is: how good are the American open-source models, and can they catch up to the Chinese models? My understanding is that they hold themselves to a higher ethical and legal baseline, and it might be very hard to catch up to those models because of that. If the models are actually not useful, it really depends on how much of a difference it makes to have a slightly smarter model, which is going to be hard to forecast, especially as things are plateauing. But it might be a fool’s errand if they cannot be competitive.

That is going to have to be an open question for the field if you just assume the extremely narrow view of plugging them into coding agents literally as is. If you imagine a world where there are many other alternatives—like, we can measure bias, have threat models for bias, train the bias out, and put guardrails on it—there are so many other ways to get security. This is actually a huge use case for us right now, which is Jev [?] classifying whether the tool calls are safe, or whether they’re phoning home to China or whatever else. Everyone has been wearing mono-model horse blinders for so long, and I just reject that premise.

Harry Stebbings

Do you think American open-source models will catch up?

Diogo Almeida

In terms of true intelligence, my guess is no, unless China decides to stop trying—not due to capabilities, but due to ethics. But it might not make a difference. If the field plateaus hard enough, maybe that X percent better doesn’t matter.

Harry Stebbings

Do you think we’re right to have the export ban with China?

Diogo Almeida

Oh boy. My understanding is that there are different parts of the—I don’t know if you mean the GPU export ban, or—

Harry Stebbings

Yeah, the GPU export ban specifically, because they’re showing—

Diogo Almeida

Oh—

Harry Stebbings

—on some of the latest DeepSeek models and some of the latest Chinese models, they were in fact trained with the latest Blackwells.

Diogo Almeida

Oh, yes. That was Will Depue’s result. I’ve been talking about that with my friends. There’s some spicy stuff online. What I heard online was that, using the same technique, you could tell DeepSeek 4.1 Flash wasn’t trained in that way.

Harry Stebbings

Hmm.

Diogo Almeida

Maybe that is consistent with the field. I don’t know. My job is to navigate all of the insane hardware-bubble neocloud craziness in a way that we end up getting the future we want. Because we’re so efficient, we can be on the sidelines and not be a main player in the GPU stuff, and hopefully ride out whatever crazy stuff is happening right now.

8. The AI Bubble Risk

Harry Stebbings

The crazy stuff that’s happening—you mentioned that, and neoclouds. Dude, it’s a wild space in the neocloud layer.

Diogo Almeida

Hmm.

Harry Stebbings

What happens in the next 12 months in neoclouds that people are not contemplating or seeing?

Diogo Almeida

Well, I believe AI has been a bubble for a while. So I’d like to keep planning for the eventuality of a bubble burst.

Harry Stebbings

What will happen in the bubble burst? To me, it’s not going to be a supply-side issue. It’s just going to be a demand-side issue. We’re seeing no demand-side shock, so what will cause a bubble burst?

Diogo Almeida

Oh man, okay, we can get into that. That’s a spicy one. Coding agents have definitely changed the game for AI, but this argument has been made for many years now.

I think it’s very interesting when you combine many different exponential-variance-type things. I believe some amount of demand is caused by FOMO. FOMO is caused by people thinking that they have good results and that they have to go all in on this type of thing. I do believe demand can burst quite quickly when people start facing the realities of what AI can do.

My experience of a lot of the world of people spending money on AI is that they don’t have a full understanding of AI’s capabilities. They’re quite miscalibrated and actually overestimate what AI can do. I think that’s the kind of thing that can cause a demand shock. There has just been so much overpromise and under-deliver in the AI space.

Harry Stebbings

That goes against everything people say, which is that if model progression did not improve for a significant amount of time, there is still so much value left for humans to extract that they haven’t extracted yet. That kind of goes against what you’re saying.

Diogo Almeida

I believe the people who say that are the people who try to extract value out of existing models, so there’s a little bit of bias in that. I don’t disagree with the spirit of what they’re saying, which is, “Oh my God, the models are so fucking smart.” I agree that they’re so fucking smart. I just think their interface and how that intelligence is unearthed are suboptimal.

Again, we need to look at the flip side: the people who went way too all in on AI and got burned by it. There are a lot of companies that said, “Let’s fire a lot of our workforce because AI is solving all of these problems,” and they need to backtrack on that. That has a lot of long-term ramifications for the future of their companies.

Those people tried not just to extract the maximum or overplay their hand; they tried to bet logically based on the reality that AI can do a lot more. So I think that reality—or that story—is very nuanced when it comes to whether people have the capability to extract a lot of that value, even if capabilities pause.

Harry Stebbings

I think a lot of people look at the margin structures of a lot of these companies and say that they’re bubble-like. Do you have good margins?

Diogo Almeida

We do, yeah.

Harry Stebbings

How do you get to scaled revenue, dude? You’re so cheap, which is brilliant.

Diogo Almeida

Mm-hmm.

Harry Stebbings

And the cost is a huge benefit of using Jev. But the scale required to get to a $10 billion revenue business is just enormous.

Diogo Almeida

Yep.

Harry Stebbings

How do you help me get there?

Diogo Almeida

Jevons paradox, right? The way Jevons paradox tends to work, it’s not like magic: you make it cheaper and it gets better. That’s an insane way for the world to work. The way it should work is, “Hey, I have this budget for AI. Jev makes it that much cheaper, so now I can make this project work. And ideally, it works so reliably, because Jev’s also smarter on the axes that matter for automation, that now I can do these other projects, or throw intelligence at things I would never throw intelligence at.”

Even then, if you’re starting to throw intelligence at things you haven’t thrown intelligence at before, that wouldn’t make you increase your budget. Jevons paradox means increasing the budget. What you would need to see is so much value from those projects finally succeeding—which AI tends to fail at—that you’re like, “Hey, this actually works. Let’s keep betting until the budget matches what it was before, and even increases beyond that.”

This is why reliability of intelligence is so important. What I like to say is that, in the grand metrics that truly matter, which is real automation, Jev, Fable, and Astra are all tied at roughly 0%. Yes, there are a couple of little bits here and there, but we have so much automation to do. We’re trying to make an industry that currently doesn’t exist, which is run-in-the-background AI. It would clearly be much larger than the whole…

Harry Stebbings

Larger.

Diogo Almeida

…larger than a coding agent assisting software engineers by making software, say, 20% more efficient to write.

Harry Stebbings

How do you analyze the coding agent market today? You see Cognition, fourth place so far, be a $1.6 billion revenue business. The depth of this market is so large. How do you analyze this market today, and what does that look like in 3 years’ time?

Diogo Almeida

Who’s third place?

Harry Stebbings

Well, you’ve got Claw Code, you’ve got Codex, you’ve got Cursor—

Diogo Almeida

Ah.

Harry Stebbings

—you’ve got—

Diogo Almeida

Okay, cool, cool.

Harry Stebbings

—Cognition. And you’ve got Factory.

Diogo Almeida

If I’m totally frank, I don’t analyze the coding agent market. I personally want Jev to be used for coding agents, and I think there’s very cool sci-fi that could be done. I actually think that the further away we stay from the coding agent market, the more they keep doing their little fight, and we can keep focusing on automation.

From a company perspective, it might even be better in the short term if Jev had absolutely no coding agent users, to make it a little bit easier. It’s a little bit too late for that, but I can’t tell. I don’t analyze the size of that market. I think it will be a very large market. Will the market be much larger than it is today? I don’t know, because FOMO really moves the needle on these things.

I’ve been around in AI long enough, and surely you have too. You must have seen AI look like this exponential upward trend, but it’s actually been many ups and downs.

Harry Stebbings

Sure.

Diogo Almeida

Each FOMO of the previous upswing adds to the next one. But a lot of them have just crashed down. I think it would be kind of arrogant to assume this is the one thing that will just be up, up, up, up.

9. Raising and Building Jev

Harry Stebbings

I remember 2015, 2016. There was this real hype for about a year, and there were Bloomsbury AI, Message.ai, and a load of these companies. It was exactly that. So yes, I completely get you there.

We talked about the bubble-like moment. With the greatest of respect, you raised a shit ton of money.

Diogo Almeida

Sure.

Harry Stebbings

Tell me about that round. Was it just like, you know, stop off at one VC, collect a term sheet, stop off at another VC, collect a term sheet? “Ah, I’ve got 28 term sheets.” How did that round go?

Diogo Almeida

It was a very, very confusing time for us. It was kind of an insane time. Oh my God, did we learn so much about VCs themselves.

Harry Stebbings

What did you learn?

Diogo Almeida

How technologically advanced the VC’s portfolio companies were. Basically, it was all inbound. We didn’t have a deck. We had to tell people, “Hey, we don’t have time for this deck, and it doesn’t seem anyone else needs this deck. You tell us about the market signals you’re seeing, because everything’s on fire right now, and we don’t really have the time to pull these things up for you.”

It was so sick to see how things evolved for different VCs over the course of the first days and weeks: how their excitement changed, and how they told us how their portfolio companies were using us. We don’t snoop on their data, so we actually don’t know what they’re using it for. It was so cool to hear them say, “Holy shit, you guys are the real deal. You’re crucial. This is the future.”

That was a really fascinating additional signal. Honestly, I miss that signal on the state of the field. It was so fucking cool to hear about who was using us, why, and how. That’s what enabled the round to work out.

Harry Stebbings

Did you start with 750 as the ask at 7.5, or did it move up?

Diogo Almeida

It moved around a lot. When we started the conversations, we didn’t know what we were really looking for in the round. What we ended up deciding was, “Hey, we don’t actually…” This is maybe kind of crazy because we did raise a lot of money, but we realized we don’t need to valuation-max. We don’t need to do a cuck tranche or anything like that. We don’t care about that. We don’t need a fuck ton of money either.

What we need is allies to fight the largest companies in the world. Let’s find the best people to be part of the team. That’s what we really optimized for. The rest was more in the micro-optimization territory, versus discovering, “Hey, we want the smartest possible money, the best allies, the people who can help us navigate the rocky waters ahead—the rough waters ahead.”

Harry Stebbings

Yeah.

Diogo Almeida

I think we were optimizing for very different things from other companies because we are profitable. Money in the bank is going up right now. Quite frankly, a company that’s raised as little as we’d raised before shouldn’t have a plan to spend whatever we raised right away.

Harry Stebbings

I can’t believe “cuck tranche” is actually a word these days, but let’s roll with it. I do. It’s completely fucked, and it makes me just—

Diogo Almeida

Wait, you know about that, surely.

Harry Stebbings

Yeah, absolutely. I have a very big ego.

Diogo Almeida

Yeah, and that came up for us. We had so much demand for that. People were even saying, “If you’re going to be raising at 3x the valuation, count me in.” The thing is, we didn’t need the money.

Harry Stebbings

But that’s a bubble. That’s bullshit. I’m sorry. I really like you, dude. Someone’s like, “YOLO, 21 billion, I don’t care.”

Diogo Almeida

Mm.

Harry Stebbings

You know what? You just need to get some self-esteem, because that is like sitting and watching.

Diogo Almeida

We don’t need the money or the valuation.

Harry Stebbings

Why did you choose Sequoia and Andreessen? I love them, but why did you choose them?

Diogo Almeida

Oh, man. It’s going to seem like I’m sucking up, but I love them a lot. Andreessen is—I’m kind of insane about technology and techno-optimism. I think a16z is just the embodiment of that. They’re as philosophically aligned as we can be.

I feel like most VCs are lovers of technology from a strategic perspective, but they’re financial institutions that care a lot about how to play their financial cards. Andreessen, deep into the firm, cares so, so, so much. That podcast I did with Martin Casado and Ben Horowitz was so much fun. It felt like we were old friends talking from our hearts. Maybe they’re so good that they can convince me and they’ve sold me on some fake-ass thing. But I don’t believe so. I think they’re just so philosophically aligned.

What led us to where we’re at right now, and I hope we can keep being this way, is wearing our hearts on our sleeves about what we care about and blazing our own path in the AI field. I would love to act in accordance with that and have them on board. Obviously, Martin is brilliant.

And actually, Sequoia—if I’m going to be totally frank, I didn’t want their money. I wanted Sonya Huang. That is why Sequoia’s around. I didn’t get to know Sequoia super well as a fund, but I love Sonya as well. And I just love that we have—

Harry Stebbings

Now help me understand. I want to be the best ambassador. What did you love so much about Sonya that made you choose her over everyone else?

Diogo Almeida

Oh, man. I would say that she’s not just brilliant, but a very unique type of brilliant that I think we would extremely want on our side. It’s that particular shape of brilliance that I find no matter how many advisors I would talk to—

If I have a conversation with Sonya, I get a different, very interesting perspective that should be part of our strategic cards. I care a lot about strategic thinking for myself. One of my spicy type-safe values is “Make great decisions slowly.” It’s very non-startup, but I think that’s how it should be done. She’s an incredibly valuable thought partner there.

Harry Stebbings

“Make great decisions slowly.” What bad decision did you make fast?

Diogo Almeida

I’d say I probably make decisions more slowly than most other people. My sense is that a lot of the business world is on autopilot.

Harry Stebbings

Hmm.

Diogo Almeida

I hate working with traditional business people. They just do things. I’m okay with doing things if it doesn’t matter how well they’re done, but for important things, I care a lot about the why. What’s the philosophy? What’s the framework? How is this done? Whenever I work with traditional business people, I feel like I’m being driven crazy because they never do things so well that I don’t need to worry about it. When I do worry about it, I need to understand the why, and I’m like, “Oh, this person’s just doing random shit.”

I think having a plan matters quite a bit. There are so many things I’d spend way more time thinking about. I wish I could spend more time thinking about our data and our data roadmap. I want to think more about product, how to be more trustworthy, and every single role we hire for, to make sure we get the best possible people. I should acknowledge that I don’t have time for all these things, and it’s so fucking important to think, think, think.

Harry Stebbings

How hard is the hiring market today? You have to get the best people on your team. How hard is it, actually?

Diogo Almeida

The hardest part is hiring for our data domain team. Back to the Bitter Lesson: when you have a new task, you have a new type of data. You need a new shape of data to make it work. We basically have our own little subfield. Imagine the best lab in the world, but internal to a company, to make data the way we make data. I’d think this should be the easiest thing in the world to hire for because it’s so unbelievably cool. And we have such a hard time hiring for it because no one knows how to get into it. We keep getting the wrong type of people applying.

That, to me, is the thing that keeps me up at night: how can we convince brilliant people who don’t think they can push the boundaries on ML because they’re not ML researchers to work in this role, where we have a subfield and they’ll be tutored by the world’s experts? It’s going to be so valuable because you actually make the model smarter. I don’t know how to communicate that to the world, and that sucks. Everything else is good—

Harry Stebbings

Is the comp wild? We had Eric at Modal say that he has people he has to fight $20 million packages for, some grads who are utterly brilliant. Are the comp packages just unprecedented?

Diogo Almeida

Not for us. I’m so happy we got clarity on this. We got real clarity that we don’t want mercenaries. We want to be good, but not like the insane labs. I think that feels great. Anthropic, I think, is suffering a little bit from their over-mercenaries lately. They want missionaries, obviously, but they also pay more than anyone else. So there’s a lot of incentive not just for mercenaries to join, but for people to fake their culture interviews. There’s an entire market now to prepare people for Anthropic culture interviews.

I’m so glad we decided not to do that. We want to pay people good amounts, but not the highest ever. We’re super happy with where we are algorithm-wise. It’s the kind of thing that takes care of people and also filters out the hired guns.

Harry Stebbings

Final one for you. If you could bake—

Diogo Almeida

Oh, yeah.

Harry Stebbings

—one character trait into a potential hire, what would it be? For my company, it would be obsessive.

I need someone to be obsessive about their craft, addicted to making it better. They can’t sleep at night. It compels them in every way. The ability to become obsessed with something.

Diogo Almeida

I don’t know if that’s the same for every role, but if I had to pick, I don’t know a better word for this. It sounds a little sexual, but we don’t mean it in a sexual way: technophilia. Nonsexual, just to be clear. Not sexual, okay, relax. We truly love technology. That’s the kind of thing that’s so beautiful in a bubble, and it’s what I want in every single role, so we can talk about technology at lunchtime or when we’re hanging out. I want people around me that I like, and I want us to be one big team working toward constantly pushing the boundaries of sci-fi, because that’s how we actually ship technological shock and awe. Again, nonsexual. I don’t know a better word. Technophile, techno—yeah, yeah. Just to be clear, if you guys edit this out, that’ll be very embarrassing.

Harry Stebbings

Don’t worry. We’ll keep it in. No sexual—

Diogo Almeida

Yeah.

Harry Stebbings

We’re going to do a quick-fire round now. I’ll say a short statement—

Diogo Almeida

Oh.

Harry Stebbings

—and you give me your immediate thoughts.

Diogo Almeida

Okay. I tend not to do quick-fire, as a warning. I’ll try.

Harry Stebbings

What have you changed your mind on most in the last 12 months?

Diogo Almeida

How much hype matters.

Harry Stebbings

How much hype matters?

Diogo Almeida

Yes. I think it matters much less than I thought. When Jev launched, it was like, “Holy shit, hype matters a lot.” Then other stuff launched, and they were benchmark-maxing to the gills. When we talk to real people, everyone says, “Yeah, we tried them. They suck. No one cares. We can already tell you’re the best and you care the most.”

I was like, “Wow, this is so convenient.” We can just do what we love most, which is underpromise and overdeliver. This is the best world ever. I wouldn’t have expected this at launch time, and this is so awesome for us.

Harry Stebbings

What’s the biggest belief you hold that the smartest people at frontier labs would disagree with you on?

Diogo Almeida

How much AI can do today, and how important reliability is.

Harry Stebbings

Why do you not hire from OpenAI?

Diogo Almeida

My sense is there’s a certain selection bias you need to get into OpenAI that’s kind of an anti-signal. I feel like you need to be in a particular Kool-Aid, number one. Number two, when I say this, I mostly mean I don’t hire researchers from OpenAI.

Harry Stebbings

Mm.

Diogo Almeida

The researchers at OpenAI, by construction, love to do algorithms and then scaling. They tend to hate data. Data is a very low-status thing at most labs. Obviously, they don’t think about the right task, because that’s quite rare. There are maybe 3 people at OpenAI who think about the right task. I wouldn’t hire them either, not because they aren’t fantastic, but because if you’re obsessed with the right task, why would you want to do my right task? My right task is machine-native AI.

If you’re obsessed with making text-to-image a thing, which Aditya did at OpenAI—holy shit, he invented an entire field. That is so fucking cool. He is so unbelievably good. You can’t transplant him into another task. He should just be cooking on his sci-fi, because that’s what you do when you’re so good at that.

By construction, all that’s left of the ML research that could be hireable are the ones who are obsessed with algorithms. That, to me, is the kind of thing you do when you want to throw money in the furnace.

Harry Stebbings

What is the most overhyped category today?

Diogo Almeida

My instincts tell me robotics. Not because robotics isn’t cool, but if you look at the grand scheme of reliability, it’s very far away. If you imagine what needs to be nailed first, there are tons of prerequisites that still haven’t been nailed. I think robotics is the furthest away from that.

Harry Stebbings

What is the most underhyped?

Diogo Almeida

This direction of software research: people playing with AI plus software to make new things, especially new interfaces. That’s so unbelievably underhyped. It’s tragic.

Harry Stebbings

Final one for you. What do you believe today that looks crazy, but will be incredibly common in 5 years?

Diogo Almeida

I think principal-agent problems are abound in the world. When you have a principal, like an electorate, represented by an agent that’s sometimes misaligned with it, like a representative, why can’t the electorate directly represent their desires and have it behave more like a true democracy? This is just one contrived example, but it happens all over the place. People reporting to you, stockholders to companies. There are so many principal-agent problems abound.

My view is that we’ll find way better solutions to all of these, and it’s going to be crazy that we allowed the level of misalignment we have today. That will be such a sweet world, where we can start enforcing the spirit of the law instead of the letter.

AI is basically that. The law is like writing software: a detailed algorithm to follow to make it fair and reproducible. It was just missing the primitives to specify the spirit. I think this will really unite society, because even on our most controversial topics, a lot of people can unite around the spirit of things. The problem is that the spirit can’t be enforced, so you need to put it into the letter, and the letter then has a ton of disagreements.

Harry Stebbings

Do you know my problem with you, Diogo? It’s that you just don’t have much energy.

Diogo Almeida

My bad. My bad. There’s a reason they call me Cocaine Labubu on Discord.

Harry Stebbings

Dude, this has been so much fun.

Diogo Almeida

Oh, yeah.

Harry Stebbings

Thank you so much for bringing the energy.

Diogo Almeida

Hell, yeah.

Harry Stebbings

And for being so willing to be so open.

Diogo Almeida

Hell, yeah. It was tons and tons of fun. We should stay in touch. This definitely was very spicy. Spicier than I intended.