Behind Latent Space: A Rare Look Inside the Pod | AIEi Shanghai
- Shawn Wang sees coding agents shifting bargaining power from mid-tier SaaS vendors to their customers. Facing a $40,000 software purchase, he offered $10,000 for an open-source replacement; the competition attracted 69 submissions, with roughly five he would pay for, built in a week. His warning to vendors is blunt: “The moment I have access to your software, I can clone it.”
- For business owners, customization can justify maintaining software themselves. Wang estimates conference software costs around $300,000 per event, potentially approaching $1 million annually across his business. Maintenance “is not zero,” but replacing repetitive workflows and integrating tools can outweigh occasional downtime; he nevertheless concedes that buying established SaaS is sometimes better.
- Wang’s preferred AI architecture puts small decisions throughout software rather than intelligence inside one chat window. Diogo’s System 1 approach restricts models to choices and probability estimates that sit directly in ordinary code. “Models are for us. We don’t work for the model”: reliability matters more than the speed headline.
- Technical differentiation needs distribution and customer understanding built into the product. Gao describes Chinese founders who ship quickly but struggle with market fit abroad; Wang argues that marketing cannot repair positioning the night before launch. His example is Diogo’s Doom demo, created by DevRel: “Show, don’t tell.”
- AIE is turning an engineering identity into conferences, distribution, and reusable technical knowledge. Wang reports more than 7,000 attendees against a 6,000-person target, 1.5 million online viewers, and 30 tracks at its flagship event. Publishing talks, transcripts, datasets, and MCP access extends AIE’s reach beyond attendance, while local partnerships expand the brand internationally.
- China offers a different adoption map from San Francisco’s “B2B SaaS bubble.” Gao describes a million-scale community built around shared learning, while Wang says he wants to see the differences firsthand, noting that ComfyUI has little use in his circles while OpenClaw has traction in China. He sees substantial founder capability, but the conversation also preserves disagreement about working intensity and the difficulty of presenting Chinese technical work to Western audiences.
- Wang pairs enthusiasm for interaction models with explicit commercial and safety fears. Full-duplex communication, continual learning, and world models interest him; an AI winter threatens a conference business he acknowledges “relies on hype.” Beyond that cycle, he says lab friends worry about society adapting too slowly, and argues that the default assumption should be “AI doom”—his stated position, rather than a settled forecast.
1. Latent Space began as a way to earn expertise in public
Wang spent eight years in trading, investment banking, and hedge funds before moving into tech. A frustrating software engineering role at Two Sigma pushed his energy into writing and speaking outside work. Publishing became a way to “choose your own boss”: prospective collaborators could discover his work rather than his career depending entirely on an assigned manager.
His explanation for finding tech more comfortable is economic: “Finance is zero-sum.” Sharing an informational advantage can destroy its value, whereas technical knowledge often becomes more useful when others adopt it. Open source, public explanations, and admitting mistakes therefore made sense to him in a way they had not in finance.
Latent Space started as a blog after Stable Diffusion appeared, before ChatGPT. Wang was working at Airbyte and wanted to enter AI without claiming expertise. His analysis attracted a response from Stability AI’s Emad Mostaque—“This is good”—illustrating how substantive public learning could create access without a new degree or formal appointment.
Wang rejects credentialism without promising that effort guarantees success. “Just because you worked very hard, you’re not entitled to win. You have to work very hard and be good.” His own assessment remains qualified: he is “a bit more qualified” than four years earlier, with persistence and judgment determining whether the work becomes substantial.
2. AI engineering professionalizes the work around models
The defining encounter came at a hackathon: a nontechnical woman asked what she could do after going all in on AI. Wang’s provocative conclusion was that engineers could control “100 AIs at once,” while her immediate opportunity was better prompting. That suggested a distinct profession beyond prestigious model training and research.
Building harnesses, in his framing, does not require advanced mathematics or supercomputing experience; it requires strong software engineering and an understanding of how AI fits into conventional code. Alternatives such as LLM engineer and applied AI engineer circulated, but “AI engineer” offered an identity people wanted and employers could hire. “You give them a job title that pays money.”
Wang says the role has changed less than its fashionable specialties: RAG and context management gave way to coding agents, harnesses, memory, and other disciplines. An AI architect increasingly sets companywide practice; an engineer owns implementation and failure. The practical test is responsibility: “The person who wakes up at 4 a.m. to go and fix it.”
3. AIE combines an industry gathering with a public technical reference
Wang’s model is gaming’s GDC, where secretive AAA studios explain difficult engineering once a year. AI labs similarly concentrate resources and expertise, while their own developer events are selective and customer focused. A shared conference lets engineers learn across labs, including the many beyond OpenAI and Anthropic, without making famous executives the entire program.
His objection to executive-heavy AI events is their interchangeable forecasting: AGI is two years away, jobs disappear, and attendees must prepare for something resembling the speaker’s own role. “Conveniently, whatever is next looks like their job.” AIE instead emphasizes practitioners who can explain implementation, failures, and technical choices to peers.
Wang reports a 6,000-person target becoming more than 7,000 attendees, with 1.5 million watching talks online and 30 tracks at the flagship World’s Fair. Every talk and workshop is published. Transcripts, downloadable datasets, and MCP access at ai.engineer/data let users consume the material through models; researched speaker and organization profiles aim toward “the Wikipedia of AI.”
Latent Space also experiments with measuring model recommendations. At aeo.latent.space, Wang compares answers across roughly 200 questions spanning podcasts, databases, banks, and other categories. He says Latent Space ranks between first and third for engineering podcasts, interpreting answers as “probably a pretty decent proxy” for the preferences reflected in labs’ dataset choices.
4. Shanghai’s organizers own the local execution
Wang immediately corrects the premise that he brought the conference to China: “Shanghai brought AIE to Shanghai.” He credits Han Hua, Lei Du, Meng Ying, and Standard Partners. He wanted to do something in China but says he lacks the citizenship, contacts, and relationships; he sees AIE as a way to surface high-quality conversations between Chinese and US AI communities.
International expansion follows JSConf’s partnership model: local teams borrow the brand, YouTube channel, and distribution. Paris applied first; Wang says its organizers were subsequently acquired by Mistral. Partners agree on brand stewardship and revenue sharing, publication of their content, and a code of conduct. Sharing the name creates reputational risk, although he says a bad local event has not happened so far.
China adds cross-border restrictions to that model. “I’m McDonald’s, and you’re McDonald’s China” is Wang’s shorthand for the franchise relationship. He also describes conferences as places where people can encounter counterparts despite restrictions on formal meetings involving some entity-listed companies. His intended scope is explicit: “We won’t do politics; we’ll only do tech.”
5. Kill My SaaS turns a customer’s expense into a vendor’s threat
Wang puts conference budgets between $3 million and $10 million, with roughly $300,000 per event going to software—about 10% of the budget and margin he says he could otherwise retain. Annual SaaS spending might reach $1 million. The frustration is not just price: employees still face slow interfaces, missing features, and requests placed on a roadmap six months away. “We are an AI conference company,” he says, operating in 2026.
Instead of approving one $40,000 purchase, he offered $10,000 to clone the product as open source and customize it afterward with coding agents. The competition produced 69 submissions; most were usable, and approximately five were good enough that he would pay for them. The striking development window was one week, although evaluation took considerably longer.
Gao identifies the competitive reversal: customers themselves become competitors to their software suppliers. Wang extends it to models navigating and reproducing interfaces: “The moment I have access to your software, I can clone it.” He imagines customers abandoning a SaaS vendor within a year, while distinguishing his physical conference business from software that agents can reproduce directly.
Wang’s proposed escape routes remain qualified: “Nobody knows the right answer.” He thinks vendors may need to become leading agents within a vertical, citing Harvey and OpenEvidence, or own specialized data and become data suppliers. Software between those positions faces customers increasingly capable of building what they need; he is more interested in helping those business owners than protecting incumbent vendors.
6. Customization still requires maintenance, skill, and human evaluation
Gao’s maintenance objection gets a direct concession: “It’s not zero.” Wang’s argument concerns mid-tier tools rather than AWS, Google, or systems requiring 99.999% uptime. For a noncritical internal database, he would sometimes accept a day’s outage in exchange for customization. Coding agents help with repairs; they do not make maintenance disappear.
Wang frames the productivity calculation from a business owner’s perspective. A developer might see $200 as a lot, but he says he pays $1.5 million for ten employees: if they work 10% better, that is $150,000 in value; he also posits that they might become five times more productive, or $7.5 million in value. That helps explain why he is willing to spend on agents and customized tools.
His concrete workflow involved 1,000 submissions for 25 spots at AIE Code the following month. Existing software required opening records one by one. A contestant supplied a table where organizers could review and reorder everything together—a view the incumbent lacked. The benefit was removing repetitive interaction rather than merely recreating existing screens.
Wang cautions against hiring an untrained person who looks vaguely technical and expecting excellent software immediately. There was a large quality gap between average and winning submissions, especially in human usability. “Sometimes buying a tried-and-tested SaaS is going to be better.” His own organization chooses to work closer to the edge because it wants to be AI-forward.
The AI judge assessed valid submissions, not the team’s eventual favorite: “We didn’t know what we wanted.” Seeing roughly 70 alternatives clarified preferences. Evaluation took about two months, partly because Wang was busy; testing also required switching among organizer, speaker, attendee, sponsor, and vendor accounts to verify that their interactions worked together.
7. Integration weakens the case for buying disconnected tools
Wang says he now has his own versions of Calendly, Google Drive, GitHub, Dropbox, and Slack. He explicitly conditions that achievement on having money and people he can assign. The broader change is software for an “audience of one, or an audience of one company”: it can be worthwhile without becoming a product sold to everyone else.
His sharpest integration example is an open-source DocuSign alternative connected to his own drive: signed documents automatically reach the correct folder. Gao asks whether DocuSign’s installed network is a moat; Wang rejects that argument—“The network is the email address.” He says electronic signatures are legally enforceable and DocuSign has no patent on them. He describes the company as worth $10 billion, “or maybe seven now,” while arguing that the workflow can be reproduced.
Wang’s formulation is memorable: “Inside every one million copy-pastes, I think there’s a fantastic startup and product to be built.” He also sees room for a “99designs for SaaS,” handling specifications, contests, evaluation, and payment. His logo competition drew 400 ideas; he would pay someone to administer a similar process for software, rather than repeat the substantial work himself.
8. System 1 models put intelligence directly into ordinary code
Wang presents Diogo’s approach as closer to his original vision of AI engineering than a large harness pleading with a model for the right answer. “Models are for us. We don’t work for the model.” Narrowing the model’s job to making choices or estimating probabilities lets developers place intelligence directly in critical application logic.
The architectural distinction is between a conventional application with one chatbot window and software that is “at every point smarter.” System 1 models should become as casual to use as an if statement. Wang says speed is almost secondary to reliability, connecting that priority to the TypeSafe name rather than treating latency as the whole proposition.
He explicitly concedes that Diogo did not invent the argument for smaller models: thousands have tried “faster, cheaper, better.” What differentiated this attempt, in Wang’s account, was OpenAI credibility from work on InstructGPT, a clear System 1 versus System 2 explanation, and technical showmanship. A nontraditional background also gave Diogo sympathy for ordinary developers outside the “big models cult.”
9. Advisory work tests ideas across companies while GTM starts before launch
Wang remains a Cognition advisor after spending six months as an employee following his personal startup vehicle’s acquihire. Running a global AIE organization required him to step back and act as CEO. He describes equity advisory relationships as an unusual career he would not recommend broadly, rather than a standard route every technical communicator should follow.
Working across 36 companies gives him a larger sample for developer marketing advice than a single employer could. He can recognize where an approach has failed elsewhere instead of assuming universality. He says most people’s sample size is five, while getting to 50 could make someone one of the top people in a field. The financial caveat is equally direct: “You cannot eat equity,” and companies may remain private for years even when their paper valuations become very large.
Gao’s community helps Chinese founders bridge fast product development and unfamiliar overseas markets. Wang’s first caveat is that a founder needs a genuinely strong product: if they are the global number one at something, marketing can add the “画龙点睛”—the finishing touch. He warns that some founders mistakenly believe they already have that core.
Wang’s prescription is to involve GTM before specifications harden: decide which customers matter, where they gather, what language they use for problems, and how the product improves their lives. Reviewing a launch post the previous night cannot repair those decisions.
“Show, don’t tell” is the operational rule. Wang says the Doom demo associated with Diogo’s launch came from DevRel rather than Diogo himself, making it an example of GTM contributing to the product’s reception. He wants founders to demonstrate benefits directly instead of submitting a slide deck of claims.
10. Taste narrows possibilities before the right answer looks obvious
Wang defines taste as selecting quickly and cheaply from a large possibility space, reaching an outcome that looks obviously good in retrospect. Brute force can test 1,000 options; a tasteful person eliminates most early, sometimes without being able to explain fully why they are wrong.
His analogy is the touchscreen phone: many form factors were possible, but Steve Jobs committed to a flat screen and rejected alternatives. Wang cites General Magic’s failed attempt and a roughly ten-year R&D journey. The point is sustained commitment to a product shape before hindsight makes it seem inevitable.
For AI products, Wang points to Cursor’s choices and says copying is evidence: “You know you have taste when other people copy what you just shipped.” His legal-software example is Legora, with its massively parallel document-review process. The resulting user reaction—“I don’t want to do my work anywhere else”—is the test he emphasizes.
11. China’s adoption patterns challenge San Francisco’s default assumptions
Gao traces WaytoAGI to Stable Diffusion enthusiasts, shared learning documents on Lark/Feishu, and ComfyUI workflows. He describes the community as having around one million users, then says several million users are learning together; he calls it China’s largest AI community. Wang characterizes a million-person community as medium-sized, and Gao corrects him: a million is already top-tier in China.
Wang wants to see those differences firsthand. San Francisco, he says, has a “B2B SaaS bubble”; his circles favor Canva and Fal.ai rather than ComfyUI. OpenClaw has more traction in China than ComfyUI, but Wang says he has not personally experienced the Chinese adoption energy people describe to him.
Gao pushes back when Wang contrasts China’s working intensity with San Francisco: he has heard American founders and engineers work similarly hard. Wang distinguishes founders with substantial equity from employees and maintains that even highly competitive US startups work less intensely. His admiration is categorical but conditional: leading Chinese founders, if based in America, “would kill Amazon” and Facebook.
Distribution also filters what Western audiences notice. Wang says Alibaba’s Qwen roadmap was buried in a broader corporate presentation that his audience did not see, leaving him to translate it. AIE could present such work directly in a familiar format. Gao hopes Shanghai also surfaces agent infrastructure—particularly safety and payments—that he finds less discussed in China.
12. Interaction models excite Wang while winter and safety remain material fears
Beyond System 1 models, Wang highlights interaction models associated with Thinking Machines Lab: full-duplex communication in which both sides listen while speaking, interrupt, and adjust immediately. He connects that to continual learning during use rather than a finished pretraining/post-training cycle. World models complete his list of areas that currently interest him.
His immediate business fear is an AI winter. He calls the present period possibly the longest AI summer in human history, while noting that AI has been around for about 50 years and that optimism has repeatedly reversed. For a conference company that “relies on hype,” the question becomes personal: “When is it going to happen, and am I going to go bankrupt?”
The broader concern is society failing to adjust fast enough. Wang says friends at labs are really worried about risks beyond misinformation and lost employment. Cybersecurity is today’s visible expression; bioweapons are an extreme example, framed as accidentally inventing “COVID-2.” He worries there may be something worse that people do not yet understand.
Wang’s strongest position is that “the default assumption for most people should be AI doom.” He invokes encounters between more and less advanced civilizations and says AI is evolving roughly 1,000 times faster than humans. That is the analogy underlying his concern: capabilities could advance beyond society’s ability to determine how to respond.
Full transcript
This is a special podcast because it’s about AIE Shanghai. AIE Shanghai is the first AIE in China. It’s brought to us by Standard Partners, one of our local partners. We have a few AIEs worldwide that my team doesn’t run, and this is the first one in Shanghai. I’ve always wanted to bring AIE to China. I think there’s a lot of East meets West, and a lot of progress in China that needs to be shared and distributed among the global community. China’s super interested in us. There have been many Chinese fans of Latent Space, and I’ll flash a few of them up on screen here on YouTube. I primarily want to recognize Meng Shao, who is our primary fan, always translating basically every episode into Chinese, which I’m super appreciative of. In general, I also want to congratulate and celebrate the progress that China has made, not just in open-source models, but in robotics and a lot of other startups. Who can remember Manus, which started this time last year and then moved to Singapore? Obviously OpenClaw had a big run. We talk about what that’s like in terms of the overall mood on the ground, how the scale of the AI community is so different in China, the types of things that get adopted, and startup culture. People hear a lot about 996; we talk about the reality of that versus a place like Cognition. Finally, we talk a little about the origin story of Latent Space, when and why it was started, how we think about running multiple shows, and a little about safety and alignment. There’s a little here for everyone. It’s basically for fans who’ve been following Latent Space for a while. If you’re interested in AIE Shanghai, it’s not being livestreamed, but we’re going to post it on YouTube at some point. It starts tomorrow, so anyone here in China watching this, I’m really looking forward to having it. Thanks to Meng Ying and the team for organizing all this. I also have to shout out Han Hua and Leidu, who basically got this thing started. It nearly didn’t happen a few times, but it’s so exciting to have AIE in Shanghai and bring Latent Space to China in general. So thanks for watching, and I look forward to the interview.
The models are very good at clicking through every screen of your software and then replicating it independently.
Oh, yeah.
Right? The moment I have access to your software, I can clone it, so good luck. There are a lot of talking heads who have never coded a single day in their lives. They’re all like, “My timeline to AGI is two years away. All of you are not going to have jobs in the future. You need to prepare for what’s next.” Conveniently, whatever is next looks like their job.
I have my own Calendly, my own Google Drive, and my own GitHub. I’m serious. They’re such good founders.
Yes.
They live in the wrong country. If they lived in the US, they would kill Amazon. They would kill Facebook. The only reason they’re not richer is that they live in China.
AI doom. In all of human history, when a more advanced civilization comes along and meets a less advanced civilization, the less advanced civilization goes extinct. This thing is evolving roughly 1,000 times faster than us.
Hi, Shawn. Thank you for joining me for this show. I’ve been a longtime listener of the Latent Space podcast, and I know many founders and AI engineers in China have already learned a lot from your work and conversations on the podcast. I’m also grateful that you’re bringing AIE, the AI Engineer conference, to Shanghai for the first time this year.
Several months ago, I read your piece on Drive Forth about your relationship with Silicon Valley. You move back and forth between Singapore and here. I guess that’s how you coined or invented the role of AI engineer the last time you moved to Silicon Valley. It feels special that you’re bringing AIE to Shanghai for the first time. Tell us a little bit about your background and your relationship with Silicon Valley.
I should correct you on one thing—
Okay.
—that’s wrong. I did not bring AIE to Shanghai. Shanghai brought AIE to Shanghai.
AIE to Shanghai. Okay.
Right? It was Han Hua, Lei Du, Meng Ying, and the folks at Standard Partners. I’m very grateful to them because I always wanted to do something in China, but I’m not a Chinese citizen, and I have no contacts or relationships there. I think there’s a lot of interest on both sides of the world in the US and in Chinese AI in general. Whatever we can do to surface high-quality conversations, I think AIE in China can help with that.
1. From Finance to AI
My TL;DR is that I’m from Singapore, a fifth-generation Singaporean. Oh, sorry, fifth-generation Chinese, right? Most Singaporeans are Chinese. People ask me, “你的家族是在哪里？”
Oh. Nice Chinese.
“我曾经留学在北京。”
Oh, cool.
That’s how it feels to be part of the Chinese diaspora, right? We don’t know what our origins are because it’s all mixed, and nobody kept any records. Sometimes nobody even knows our birthdays. I’m fifth-generation Singaporean. I grew up in Singapore and went to the US for university.
My first career was in finance. I was in trading, investment banking, and hedge funds for 8 years, then switched to tech. I worked at Amazon and a few startups, and then, 4 years ago, I became independent and wrote Latent Space. I figured out that there was going to be this AI engineer thing, and then I started AIE.
Cool. I know you have different roles right now. You’re doing the Latent Space podcast, writing newsletters, and you also have different roles in several startups, like Cognition. I’m wondering what your current titles or roles are in the different—
Yeah.
—different roles.
It’s very funny because I recently had to go over my taxes and figure out how much money I have. Cognition is the biggest one. I’m still an advisor to Cognition. I was an employee for a while.
The idea was that I have 3 companies: AIE, Latent Space, and miscellaneous. Miscellaneous used to be Small AI, which is my personal startup vehicle.
Right. Yeah.
I basically got acquihired into Cognition.
I see.
I worked there for 6 months, and then stepped back to be an advisor because AIE is now a global conference organization, and I had to be CEO. I couldn’t be a part-time employee and do this part-time. Cognition is very big. Fireworks is very big. Modal is very big. Supabase is, too.
I think it’s a very common thing if you’re a Silicon Valley influencer or you’ve been in developer relations circles for a long time. That’s my specialty: I’m technical, but I also write and speak. That’s a good combination of skills to have, because any startup will want to pay you to talk about their stuff. Then you start working with a bunch of things. This is not a normal career.
I don’t even recommend it for most people, but it is my role in life. I can advise a few different companies and work with them in exchange for stock. That’s the rough idea.
I just want to deep dive a bit more into the early days of Latent Space, because I read your essay about how you started learning in public, and then—
Yeah.
—telling everyone what you learned in the early days.
Yeah.
And then you felt the need to transfer the conversations you were having with different researchers online into some of the early conversations on Latent Space. Probably just—
There are two different things you’re mixing together, right? The first thing is the principle of learning in public, which is something I realized when I left finance and worked in tech. Why was tech going so much better for me? It felt so much more comfortable.
For people who are interested in both fields, the very core reason is that both are high-paying, prestigious jobs. The difference is that finance is zero-sum, right? If I know something and I tell you—
Oh.
That truth loses value.
Right.
But in tech, most things are positive-sum. If I know something and tell you, it increases value. That’s a very different paradigm for people crossing over from finance to tech, and it’s very liberating. People will get on stage and open-source their work.
Right.
They share their code. They screw up and share that they screwed up. In finance, no. You want to pretend you never screwed up and that everything is top secret. I really took that to heart.
Another thing that happened to me was that my first job was at Two Sigma, which is still in finance, but I was on the software engineering side. I realized I had a bad boss. I got really bored, so I started doing a lot of writing and speaking because I wasn’t able to do the things I wanted to do at work. Since I was so bored at work, I put all my energy into the stuff outside of work. That took off because no one else was doing it.
Mm-hmm.
By comparison, that lets you choose your own boss. Most people join a company, get assigned a boss and a team, and often don’t know whether the team is good. Then they find out it isn’t, and they’re stuck. When you learn in public, you can actually choose your own boss.
Right.
You work for yourself and choose your own boss. All those things are very good. Latent Space was created when Stable Diffusion happened, before ChatGPT.
Yeah.
When Stable Diffusion happened, I was at a startup that wasn’t going very well. It was called Airbyte. It’s still doing okay today, but nobody would call it a rocket ship. That’s objectively true, and I’ll say that to the founders. I’d taken some machine learning courses, but I wasn’t an AI expert. I was so interested in the field, and there was only one way for me to get into it: learn in public.
So I started a blog and began analyzing and learning, then writing up what I learned. At first, nobody would notice, but eventually, if you start talking about people’s work, they notice. Emad Mostaque, the CEO of Stability AI, which created Stable Diffusion, read my blog post analyzing Stable Diffusion and said, “This is good.” That work was basically me learning AI over the past 5 years, grinding and writing about it.
You’re not qualified, and I’m still probably not that qualified today, but I’m a bit more qualified than I was 4 years ago. That’s good, and I think it’s the only way you can learn. Especially in Asian cultures, there’s a lot of emphasis on credentialism.
Yeah.
You need a piece of paper. You need some qualification. If it’s not a degree, it’s a certification.
Certification.
If it’s not a certification, you must have had a job to do it.
Yeah.
How do you get the job? You must have a degree. It’s an infinite loop of things that basically mean that when you’re 16, you need to decide what you want to do in life. But life is a lot more flexible than that, and you can choose to pivot if you want to. You just have to be persistent and realize that just because you worked very hard, you’re not entitled to win. You have to work very hard and be good.
A lot of people try and fail, and they’re like, “Why did I fail?” Well, you’re not good. Too bad. But some people try, work hard, and have judgment. They work on substantial things rather than insubstantial things. That’s generally what happens with learning in public.
2. The AI Engineer Emerges
So, yeah, Latent Space started as just a blog.
It started as just a blog.
Yes, just a blog. I had no other idea apart from, “This is Swyx writing about AI.” That’s it. But we happened to host some events and gatherings. ChatGPT came out, obviously the most exciting thing of all time, and we wanted to host some hackathons. At a hackathon, the idea for AI engineer came about.
A woman came up to me and said, “I’m all in on AI. I’m non-technical. What can I do?” I realized the simple answer is that, compared to an engineer, she can do nothing. Engineers are about to be so much more powerful than non-engineers because engineers can control 100 AIs at once and program them.
Yeah.
Whereas the most a non-technical person can do is prompt better. I thought, “I’ve seen this before.” Throughout my career, I’ve seen the professionalization of roles that aren’t prestigious. What do I mean by non-prestigious? It’s not prestigious to be a prompt engineer. The prestigious thing is training large language models. You do pre-training, you do post-training. You have to get a PhD. You make fancy algorithms like GRPO, and you do manifold hyperconnections, which is what DeepSeek does.
Yeah.
That’s prestigious, and it continues to be prestigious. But people overcorrect by looking at everything else that isn’t ML research and thinking it’s worthless. That’s untrue. You do want to build harnesses, and harnesses don’t require you to know advanced math or have a lot of experience in supercomputing. You just need to be a very good software engineer who’s open-minded enough to understand how to put AI into traditional software code. That’s a harness.
Mm-hmm.
We didn’t have the word “harness” at the time, but it was clear that engineers would be able to do things like this that normal people couldn’t.
Right.
They’d be in demand. Companies are trying to hire these people, and developers want to become them. They want a term to call themselves. All you have to do is figure out the right term. The simplest one is “AI engineer.” Then you make it an industry, make it a discipline, and that’s what I did.
Actually, I can share a side story from China. At the time you just mentioned, when Stable Diffusion was open-sourced, the largest AI community in China, called WaytoAGI, also started from a bunch of people who were ambitious and loved Stable Diffusion technology. They wanted to transform it into different image-generation applications, and it also became the foundation of the ComfyUI community. You’re definitely familiar with the ComfyUI community.
Yes, Comfy Anonymous. We’ve had him on the podcast.
Yeah, in China as well. They gradually invited more people to learn in public and contribute to online documents on Lark, which is Feishu by ByteDance. Everybody contributed their knowledge and shared their use cases about Stable Diffusion and ComfyUI, and the community became larger and larger. I think right now the online community has around 1 million users learning together, and several million users learning together, so it’s become the largest AI community in China as well. I think that happened at exactly the same time. I think you—
Yeah.
—you started to feel like we needed a new word for this type of engineer, and then you published the essay, “The Rise of the AI Engineer.”
Yeah. Here are some of the alternatives people were trying at the time. There was AI Tinkerers, another significant community. I know Joe and Alex well. People were also saying “LLM engineer,” “applied AI engineer,” and all these other terms. Basically, you have to focus on the job title that people want to have.
Right.
I think I picked this up from James Clear. Is James Clear the author of Atomic Habits? I think so. I forget the authorship relationship, but I think it’s James Clear. Basically, if you want to deeply understand or connect with somebody, you have to connect with them not on the level of their job—
Mm-hmm.
—not on the level of their activities or hobbies or whatever.
It is their identity. Who are you?
Right.
What do you aspire to be? If you can give them a path toward becoming a better version of themselves, then you can build a community around that. A lot of people who weren’t in AI wanted to be in AI.
Right.
And what’s the path for them? You give them a job title that pays money.
Right.
Then they’ll be loyal to you.
Yeah, yeah. Right. That’s exactly—
And so, now there are a lot more AI engineers in the world.
Yeah. And to follow up, just out of my interest as well: I know there’s a lot of work to do as an AI engineer right now. You mentioned harnesses, and a lot of developments are happening in AI. Do you think the role, or the term “AI engineer,” has evolved? Is it different from what you mentioned three years ago, when you started this work?
Honestly, it hasn’t really changed that much. What’s hot within AI engineering has changed, but the job title is pretty much correct. In the early days, it used to be RAG.
Yeah.
Managing context and all that. More recently, it’s been coding agents, harness engineering, maybe loop engineering or whatever.
Yeah.
There are also other disciplines, like memory. For those interested in all the fields of AI engineering, AI Engineer World’s Fair—the flagship, biggest conference—we have 30 different tracks on what it means to be an AI engineer. Evals are part of that, multimodal is part of it, and all that stuff. So I think it has shifted over time. “Agent engineer” might be one of the terms, or—
AI engineer.
AI architect. AI engineer is the individual-contributor version of this, but who is the company-level, company-wide boss? AI architect seems to be the emerging term for the person who decides, “This is how everybody in the company does AI.” There are also AI designers, AI PMs, and all these things. It’s about who does the work, who writes code, and who is responsible for checking that code. If something fails, are you responsible? If not, you’re not an AI engineer.
Right.
The person who wakes up at 4 a.m. to go and fix it—that’s the engineer.
Yeah.
Right? We want to serve those people, and that’s the real core of it. The other problem with AI is that there’s a lot of bullshit. There are a lot of talking heads who have never coded a single day in their lives, but they’re very big-picture and very senior. They’re like, “Oh, I’m a CTO,” and so on, but they don’t touch code.
Mm-hmm.
A lot of AI conferences and communities focus on those guys because they have big, fancy titles, but they all sound the same. They’re all like, “Oh, my timeline to AGI is two years away. All of you are not going to have jobs in the future. You need to prepare for what’s next.” Conveniently, whatever comes next looks like their job, so it’s very self-serving. One of the reasons for AI Engineer World’s Fair is that we try to focus on the lower-level engineer who usually doesn’t get the spotlight, but actually did the work and can talk about their war stories and share what they’re doing with their fellow engineers. I borrowed this idea from other conferences I’ve seen in my career, like GDC, KubeCon, JSConf, and PyCon, where it’s engineer to engineer. They’re the decision-makers for the tech stack, and they’re the people who create and innovate new things. They’re not necessarily the most famous people, but they’re definitely right when it comes to engineering. There’s no bullshit, and I think there’s a lot of bullshit in other things.
Yeah.
Yeah.
And that brings us to the first AIE summit you organized about three years ago. Now it’s a three-day, full summit with over 7,000 people joining this year. For the audience who aren’t familiar with the event, could you talk about the different kinds of content, workshops, and topics that AI engineers care about? And how did you bring this small event to such a large scale this year?
3. Building a Global AI Conference
AI Engineer World’s Fair is an industry flagship conference, like the ones every other industry already has. Gaming has GDC, with 30,000 people in San Francisco. Security has two: RSA and Black Hat. If you’re in cybersecurity, you go to one of those things.
Yeah.
That’s where you learn. I really like the gaming one because gaming is very similar to labs. You have AAA studios working on the hardest games in the world. It’s extreme engineering, and everything is top secret. They don’t do anything else, but once a year they come to GDC and share, “This is how I made Game of the Year.” If you’re trying to be a game developer and want to work at a AAA game studio, it’s the only way you learn.
Yeah.
I think model labs are similar to game studios now. They have a lot of money and are working on very big, major projects. There are a lot of smaller companies trying to compete with them, but the labs are the important ones. You want to learn from the labs, so where are you going to learn from them? They have very small conferences, like OpenAI DevDay or Code with Claude for Anthropic.
Mm-hmm.
But those are super small, very selective, mostly for customers, and not for engineers—
Mm-hmm.
—talking to other engineers. And there are 50 other labs that aren’t OpenAI or Anthropic that also have something interesting—
Yeah.
—right? Usually, every industry needs one big event where everyone comes, and you can walk around the room and talk to everybody.
Mm-hmm.
Then they can talk to each other.
Yeah.
That’s very interesting, especially for Chinese labs, some of which are on the Entity List. Is this something you guys talk about?
Yes.
They cannot officially take meetings with you.
Yes.
A lot of American companies cannot officially take meetings with you.
Yes.
They’re not allowed. They don’t even respond to you. It’s not because they don’t want to or because they’re assholes.
Mm-hmm.
They’re not allowed. But if you’re in the same location, you run into each other.
Right. Right.
I think that’s kind of interesting, because I can create opportunities for magic to happen. People get hired, people find a co-founder. I know one couple that started dating, and now they’re basically married.
Wow.
Yeah, that’s my number-one KPI: the number of marriages and babies that I’ve created. I really like that. I think we can create that kind of high-quality community. We were aiming for 6,000 people this year, and we managed to hit 7,000-plus. One and a half million people watched our talks online. People who can’t afford to come, don’t live in San Francisco, or don’t live in the U.S. can still watch—
Right.
—our videos and catch up on all the content. We publish every single talk and every single workshop. As I said, there are 30 different tracks. There are too many to name, but you can see them on the website. I try to focus on things that are working, that you can make money with, and that have clear role models for you. If you’re trying to learn this field or you’re a startup trying to build something, somebody has probably talked about your thing at AI Engineer. That becomes a very valuable resource for anyone looking up anything. Another thing I’ve recently done, which I haven’t been super public about, is transcribe and catalog everything. You don’t even have to watch the talks. You can take our whole website and feed it into an LLM, and then it can teach you whatever.
Oh, are you actually creating a CLI based on top of that?
Everything is MCP.
MCP.
But yeah.
Oh, nice.
Everything has been transcribed. It’s available as a dataset. If you want to train—
Oh, nice.
—your LLM on it, go ahead.
Have you published it yet?
Oh yeah, it’s on—
It’s on—
—ai.engineer/data—
Oh, nice. I’ll check it out.
—which also shows the MCP and the datasets that you can download. It just depends how you want to consume it.
You know, we transcribed everything. We also did deep research on every speaker and organization. Whoever has ever spoken at AI Engineer will have a profile, and we’ll basically try to make the Wikipedia of AI.
Nice. One last question about AI Engineer: when did you actually bring it to different markets around the world? I know this year we have Singapore, but a few years ago you had Paris.
Yeah, Paris was the first one.
Paris was the first one. How was that journey?
It mostly follows the JSConf model. JSConf is the number one JavaScript conference.
Yeah.
It’s a very simple model. I’ll never reach every country in the world, but every country in the world will probably want to do something with AI. Their country deserves to have a leading conference. They can either make their own or borrow our brand, our YouTube channel, and our social media. So we set up a partnership program, and Paris was the first to apply.
Okay.
A fun fact: the organizers of AI Paris got acquired by Mistral.
Oh.
So now it’s run by—
Nice. Yeah.
—France’s and Europe’s leading AI lab, which runs AI Paris. That’s very cool. I want that for every country. You have AI India, you have AI Japan.
Right.
I’m working with all those companies now, and it’s very cool. I’ll never know those regions as well as they do, but the people there deserve a good conference. Absolutely. Then they can do business, hire people, and all that stuff.
Mm-hmm.
The world is very big, so we follow that model. There are 3 things. First, you agree not to abuse the brand name, and we have some kind of revenue-sharing arrangement. Second, you publish everything you’re doing. Third, you have a code of conduct, so if someone is abused or mistreated, there’s a way to resolve it. I think that’s mostly an ethical thing. We take a lot of risk by sharing our brand with people we don’t work with. What if they put on a bad AI conference?
Yeah.
So far, it hasn’t happened. I think we’re careful about who we work with, and the people drawn to us are pretty enthusiastic and organically interested in AI. Especially in Singapore, with AI Singapore.
Yeah.
First of all, I’m Singaporean, but that actually didn’t matter at all. The top AI conference in Singapore is actually a crypto conference, right?
Yeah, which turned out to be SuperAI.
A lot of crypto people want to be in AI.
Yeah, yeah, yeah.
And yes, it is SuperAI. I think they know I don’t like them. They’re like, “This is embarrassing for my country. You’re telling me this is the number one thing in my country? We need something better.” So now I think AI Singapore is up there. It’s at least competitive, and maybe better than SuperAI. I’m very proud of the team for doing that, and I didn’t organize any of it. I think it’s important for every country in the world to have a conference like this, and we can support that.
The funny thing is, I saw you went to Singapore for the first AI event. I knew Le very well, and I asked him, “Hey, AI Engineer came to Singapore for the first time. When—”
With China.
“—when are you going to Shanghai or China?”
Yeah, sorry.
Then he told me he was working with you.
Yeah. China is tricky because I’m a Singapore citizen, not Chinese, and obviously there are a lot of restrictions between the U.S. and China. That’s the main tricky part, which is why I needed Le and his team to really own the whole thing. I can’t even set up a local entity in China. There are a lot of controls, and I don’t want to get in trouble because my main business is in the U.S.
Mm-hmm. Mm-hmm.
I can’t get into trouble with the White House. But if you want to share the brand and run it as your own thing, as a franchise, that’s fine. There are lots of franchises. As far as we’re concerned, I’m McDonald’s, and you’re McDonald’s China.
Yeah. Yeah.
Okay.
Yeah, happy we can bring that.
There’s KFC.
Yeah.
KFC China.
Yes.
All right, great.
And next, I guess, is the contest you shared a few weeks ago about Kill My SaaS. I really like that show. Tell us something about the idea behind Kill My SaaS.
4. When SaaS Gets Cloned
Okay. I’m wondering how many of your listeners will care about this. You’ll only care if you own a business. Most developers don’t own businesses, so they’re not that interested. But if you aspire to own a business someday, you’ll care more.
Yeah. SaaS was a very hot topic several years ago, and it’s dying and being discussed a lot this year. I found this show very interesting.
Okay, let me introduce the overall context.
Yeah.
My budget for running a conference is somewhere between $3 million and $10 million, depending on the size. Roughly $300,000 of that per conference is software. That’s about 10% of my budget and 10% of my margin that I could be getting back. That’s going to somebody else’s software, which is shitty, slow, ugly, and not exactly what we want. I might be paying $1 million a year just for SaaS. You want this, you want that, so you can do your work.
Yeah.
As the boss, when my employee comes to me and wants to use something, they say it’s important for their work, so I say yes. But when it all adds up, it’s a lot of money.
It’s costly, yeah.
Then they use it and say, “It doesn’t give me what I want.” It’s slow. As a customer of a SaaS, you’re in a queue for support. You want a feature, and they say, “Six months from now, we’ll put you on the Q3 roadmap.”
Mm.
“You are very important to us.” But we’re an AI conference company, it’s 2026, and there are coding agents. We can do this. So the idea was that I took a $40,000 SaaS that my employee was proposing to pay for, and I said, “I’ll pay somebody $10,000 to clone this whole thing.” It’s open source, and it’s exactly what we want. From there, we can use whatever coding agent—Codex, Claude Code, Devin—to customize it however we want.
We did it. We announced the competition and had 69 submissions. When you have a competition like that and offer free money, and I’ve offered to cover everybody’s coding agent subscription, you get a lot of bad submissions. I didn’t give money for the ones that were obviously bad. But we had 69 good ones, and most of them were usable. About 5 were good enough that we would pay for them. They were done in 1 week.
Yeah.
As an owner of a SaaS business, maybe you have 50 customers, all of them paying you $40,000. Imagine if they all realized this and just turned away.
They just turn away.
They just turn away, and in 1 year your business is gone. You went from earning $20 million a year to nothing.
Yeah, yeah.
It’s very scary for a business owner. That’s the tricky thing about owning a SaaS. I own a physical business, so that’s not going away anytime soon. Robots are cool, but they’re not organizing conferences.
Mm.
But people who own mid-tier B2B software businesses—Salesforce is fine. Nobody’s worried about Salesforce.
But the mid-tier SaaS companies serving a lot of small and midsize businesses—do you call that SMB?
Yeah, SMB.
SMBs are going to really struggle.
Yeah.
Or they’re going to pivot to something better. They’d better figure it out now, because all of us business owners who are their customers can code.
Yeah.
And it doesn’t take that much time.
We are the actual competitors to their own software right now.
And by the way, the models are very good at clicking through every screen of your software and then replicating it independently.
Oh, yeah.
Right? The moment I have access to your software, I can clone it, so good luck. That’s the reality. If you’re a SaaS owner, nobody knows the right answer.
Mm-hmm.
I think you basically end up having to pivot to being an agent builder. You have to become the U.S. equivalent of Harvey or OpenEvidence. You own a vertical and you’re the top agent in that vertical, or you own specialized data and you’re basically a data vendor. Anything in between, someone else is just going to code their own thing.
Can I ask a follow-up? You already answered my first question by giving advice to small SaaS owners.
5. Software for an Audience of One
I’m more interested in advising business owners. I’m not that sympathetic to SaaS owners. They’re big boys; they can figure it out. I run a small business.
Yeah.
I’m an SMB.
Yeah.
There are a million people like me. They’re all just trying to make an honest living. I pay my employees, I pay my taxes, and it’s a good life.
Mm-hmm.
It can be better if you vibe code your own SaaS.
Oh, yes. Right.
My customers are happy, my employees are happy, and I’m happy. It’s a cool story to tell. SMBs are in a golden age. If you do anything in the real world and you’re just trying to do your job, and there’s a lot of shitty software between you and your intended outcome, you can suddenly shorten that timeline.
Buy your Codex, buy your Claude, and think carefully.
Yeah. I mean, $40,000 worth of Claude is a lot of Claude.
Yeah, it’s a lot of Claude.
I can spend $10,000 and still have plenty of money to do whatever I want.
Yes.
That’s why I call this a business-owner problem, not a developer problem. A developer might say, “Oh, $200 is a lot.”
Mm.
I’m a business owner. I understand the productivity benefits of having 10 of my employees work together well. I’m paying $1.5 million for those 10 employees. If they work together 10% better, that’s $150,000. Potentially, it’s not 10% better; it’s 5 times better.
Mm-hmm.
That’s $7.5 million. Great. That’s all the agent work being done on my behalf, when people used to say, “Well, okay, I’ll come back next week with a proposal.” No, I want it this afternoon.
My follow-up is: once you’ve created so much software yourself, will the maintenance work be difficult?
Yeah, it’s real. It’s not zero.
You still have to use it, like—
It’s not zero, but we have coding agents.
Oh, yeah. So, it’s self-maintaining?
Yeah. A lot of it is—this is why I’m talking about mid-tier SaaS. You’re not Salesforce, Google, or AWS. You’re not providing a mission-critical, 99.999% uptime service. You’re just a database where people type in their information and get something out.
Yes.
So, okay, it goes down for a day. Whatever. You can fix it in a day.
Right.
I’d take that downtime in exchange for being able to customize it. Being able to customize anything is amazing, and my team is so much more productive. I actually talked about this in the Kill My SaaS episode. One of my favorite features: for AI Engineer Code, my next conference next month, I have 1,000 submissions and 25 seats. I have to go through 1,000 people to pick 25. In the SaaS we currently use for that, we have to click through every one, one by one. Every screen: click, click, click. In the Kill My SaaS competition, one of our contestants changed everything so you can reorder it all in one table. That view doesn’t exist in the SaaS.
Okay.
To get that view, we’d have to request it six months in advance. But you can just request whatever you want.
And they just clone it and customize it.
Now I can do it, and I can see everything in one go.
That’s so quick.
If you think this is a stupid example, it is. I’m simplifying for audio.
I found out that the top winner is someone who can manage all the speakers and lineups. That’s really useful.
Yeah. And it’s good UX. They’re testing it from the point of view of humans. There’s a big difference between the average submission and the top submission.
Yeah.
Today, in 2026, if you’re a business owner and you hire the youngest kid who looks vaguely technical and say, “Hey, vibe code this SaaS for me, please,” maybe they won’t do so well because they don’t have any training. But give it a year. Watch a lot of AI Engineer videos.
And then you’ll be kind of—
There’s a skill level, right? A lot of it is slop, and you don’t want to use slop. Your employees deserve better. You deserve better. Sometimes buying a tried-and-tested SaaS is going to be better.
Yeah.
It is going to be better.
Yeah, of course.
But we want to be AI-forward as a company. We’ll be more cutting-edge.
One last follow-up. You said in the podcast that you had an AI judge, and the AI judge’s picks were quite different from the winners your team chose.
Yes. Wow, you really listened to the podcast.
I’m interested in the whole process. I just wonder why that was.
That was partially on purpose, and partially because that wasn’t the goal of the judge. The judge was only checking whether you submitted a valid submission and qualified.
Uh-huh.
But the judge wasn’t meant to determine which submission was the best out of the whole cohort.
Oh, okay.
Obviously, there’s a difference.
Okay. So it wasn’t for your team’s conference purposes, to pick what was most useful, but just to qualify the submission?
Because we didn’t know what we wanted. I can’t make an AI judge for things I don’t know what I want.
Okay.
But once I see 70 different examples, obviously I know what I want.
And then you know which one is the one.
Yeah.
Yeah, that is the one.
By the way, fun fact: the Latent Space logo, which you can see on the podcast and on the sign outside this room, came through a website called 99designs. Are you familiar with it?
No.
When you want a new logo or any kind of design, maybe you ask a friend, “Who’s good at design?” They give you 3 names, and you pick one. Then they give you a logo, and that’s it.
Yeah.
With 99designs, you put up a contest and get 400 different ideas.
For a logo.
Wow.
And then I award the prize to the logo we got, which is a very good logo.
Oh, okay.
It’s a very good logo.
Yeah, yeah.
It’s so much better, right? You want diversity in ideas. So this idea of doing contests—
Oh, have contests.
Yeah, yeah.
Contests, yeah.
Which I can do because I’m an influencer.
Yeah.
Most people can’t do that because they don’t have the audience.
Okay.
But I think there’s something there. 99designs is a very good business because there are so many designers. It doesn’t take that much to create a design, right? It’s more about the quality of the idea than the amount of work. You just need high-quality ideas, and you need to be able to—
Create it.
—to spread it—
In code, yeah.
—over a lot of different contests. 99designs is like that, and I think there could be a lot of 99designs for SaaS, which is something I’m thinking about. You know what I mean?
Yeah.
I basically had to create a 99designs for myself: “Here’s the $40,000 SaaS. Here are the specs. Here are the judging criteria. Here’s the rest of it,” right down to, “Here’s how we pay you when you win.” I had to work all of that out myself. But would I pay someone else to do all this for me, so I just tell them what I want—
Mm-hmm.
—and get 70 different versions, and pick one?
Mm-hmm.
Oh, yeah.
That’s very cool.
This is all the admin for me, right?
Yeah, yeah.
So there should be a 99designs for SaaS, and for cloning all these things. As a business owner, honestly, I’m only doing this because it’s good content. It’s a podcast.
Yeah.
You’re interested in all these things. But if I didn’t have a podcast, I probably wouldn’t do it.
Yeah.
It’s a lot of work. It took us about 2 months to evaluate everything, partially because I’m busy, but also because there’s a lot to evaluate.
Wow.
There are 5 roles for this particular SaaS: organizer, speaker, attendee, sponsor—
Yeah.
—and vendor.
Mm-hmm.
They all have to agree on what the conference organization looks like.
Yeah.
To test it, you have to create an account for each role, do some data entry here, change accounts, do some data entry there, and change accounts repeatedly. That’s why it’s complicated.
Yeah.
That’s what a mid-tier SaaS is.
Yeah.
That’s why I pay $40,000 for this, because I don’t want to do it. But now that you can, it’s a very fun job and problem to work on.
Yeah, sure. It’s a good experiment. We also run conferences in China, with over 1,000 attendees before, so I can relate to what you’re saying.
Yeah.
About—
But then think about the rest, right?
Yeah.
I used to pay for Calendly. Do you guys have Calendly?
I don’t, but—
You know what I mean, the booking link. That’s how we got to this meeting. Calendly is like a $13 billion business.
Yeah.
And, yes, you can acquire Calendly.
Yes, I know.
Of course. I have my own. I have two versions.
You already made your own Calendly?
Yes, yes, yes.
Okay.
I have my own Calendly. I have my own Google Drive. I have my own GitHub. I’m serious: I have my own Dropbox.
My God.
I have my own Slack. All of this is because I have money and people, and I can assign them to do things. But I do think you can build software for yourself for an audience of one, or for an audience of one company—
Mm-hmm.
—and that’s okay.
Mm-hmm.
The fact that you can do that now is something that has never existed before for any business owner, which is amazing.
Yeah, yeah. Yes, great.
Yeah.
Yeah.
Okay, you do some investing, right? And you’ve raised money.
Uh-huh.
So you know DocuSign?
Yes.
A $10 billion company.
Yeah.
Or maybe $7 billion now.
Yeah.
What is DocuSign? You send a PDF, and they say, “Oh, yeah, sign.” It gives you a certificate. That’s it.
Yeah.
They charge you something like $10 per PDF, per signature, or per envelope.
Per PDF.
They call it an envelope. There’s an open-source DocuSign, and I host it for myself. You know what the difference is? Because I have my own Drive and my own DocuSign, the moment the DocuSign is signed, it goes straight into the correct folder on the Drive.
Oh, yeah, because you integrate them together.
Yes.
You integrate them together.
Yes. And that doesn’t exist in DocuSign.
Right, right. Yes.
My data room has the DocuSigns and everything in it, without me having to touch it, because it already has—
Right, yeah.
—everything.
The one thing special about DocuSign is that, because it was built a long time ago, it creates a network effect: everyone joins the network, everyone uses it. Do you still think that’s a moat for this kind of software?
Fuck no.
No?
No. Electronic signatures are legally enforceable everywhere.
Yeah. Oh.
DocuSign has no patent on it. I can do my own electronic signature. The network is the email address. It’s like—
Yeah.
It’s so stupid. I can keep going about why I made each of those things. It’s not like I’m encouraging everyone to do this. Whatever problem you face in your business or your life, you can find an open-source version of it, host it for yourself, customize it, and integrate it: “Okay, I want this plus this.” Boom, they just work, because I fucking use them all the time. I copy here, paste there, copy here, paste there. Inside every 1 million copy-pastes, I think there’s a fantastic startup and product to be built.
Great. For my next question, I’m quite interested in your conversation with Diogo from Jev, because it was a really hotly discussed topic and product.
Yeah, yeah.
It creates a new way of thinking about using LLMs from a totally different angle. I even heard about a podcaster from China who didn’t know anything about coding, but used Jev to judge whether clips from her podcast would go viral. Jev gives her an answer, and then she posts it on the network. I found so many great use cases based on Jev, so I wonder what your take is and what might happen.
For those who didn’t listen to the podcast, we were the first podcast with Jev.
You were the first one.
I forced him because I’ve been friends with him for 4 years. I said, “If you don’t give me the first one, I’m going to kill you.” He’s obviously very accomplished. He was a co-author of InstructGPT, which became ChatGPT. I think the interesting thing about Jev is that it’s the original vision of AI engineering for me. Right now, AI engineering is what we talked about: a big harness around a model. I’ll do a bunch of work, hand it off to a model, and say, “Please, God, model, give me the right answer. You can take 10 days to think about it.” Jev is like, “Well, no. Models are for us. We don’t work for the model.”
Right.
I’ll reduce the scope of the model to just one thing: making choices or estimating probabilities. Then I’ll put it in the path of my critical work. If you think about the failure mode of the current LLM paradigm, you have the big traditional SaaS, and then your LLM is this window that’s a chatbot.
Mm-hmm.
Right?
Mm-hmm.
Or you look like Codex, ChatGPT, or Claude.
Large agents.
And it’s like a chat window here—
Yeah.
—menu here. That’s it. What Jev represents, and what System 1 models represent, is why Diogo doesn’t call it a decision model; he calls them System 1 models. They’re in the path of code. You write them as casually as you write an if statement in a normal programming language.
I think that’s what he means by smarter software, which I really endorse: your software should be smarter at every point, instead of being smart at one point and dumb everywhere else.
Mm-hmm.
That’s really dumb.
Mm-hmm.
Yeah. I think the speed is almost secondary to the reliability, which is what he’s really trying to say. That’s why the company is called TypeSafe. But the speed is nice.
I just found out that Diogo himself is also a very social person, and he’s really passionate about what he’s been doing. I wonder: you’ve talked to a lot of researchers and developers, or AI engineers. Why do you think it was Diogo who made Jev? Was there something special or different about him compared with other researchers or people you’ve talked with? I find it quite special how he talks about what he’s building and what he believes in.
He has a nontraditional background. He didn’t go to Stanford for a PhD or follow that kind of path, so he has sympathy for the regular developer. But he also worked on ChatGPT.
I think a lot of people who went through the traditional machine learning researcher path—going to a top university, working on an interesting paper, getting hired into a lab, and working there—are in the big-models cult. Everything will be solved by making our model bigger, by having more FLOPs in our training run, more parameters, more, more, more, more.
Right.
It takes someone who doesn’t come from that mold to say, “Well, actually, there’s another way to do this: smaller things sprinkled everywhere instead of one big one.”
And Diogo wasn’t the first to do that. Thousands of people have tried to say smaller models are good: faster, cheaper, better, whatever. But somehow Diogo’s approach worked.
Mm-hmm.
First, he has the credibility of OpenAI. Second, he made it a System 1 versus System 2 thing, which I think makes the communication very clear. You can play Doom—or whatever game demo he showed in the launch post—and it’s obvious there’s no way you’d ever use an LLM to do that.
I think that showmanship—the pink jacket, how passionate he is, how he talks—combined with the technical capability is similar to learning in public. I’m a small version of that: I’m technically capable, and I can talk. But he’s 10x on both dimensions, and I think that really comes through when you talk to him.
I guess my next question is about your role advising other startups, since you just mentioned that. I’m curious: in an advising role, exactly what kind of advice do you give those startups, and how is that—
I’m also curious: for your listeners and for me, what do you do for other companies?
We’re a community that started 3 years ago. We found that a lot of Chinese-based founders want to go global. They were pretty good at building and shipping products quickly, but they found a huge gap in go-to-market, marketing, and understanding the market.
You’re a Chinese founder trying to break into the US. You’re very good at your tech, but you suck at marketing. You’d better talk to someone who’s good at it, because you can’t do it yourself. That’s it.
Yeah, that’s how I truly found out there’s a huge gap between understanding the local market and shipping the product correctly or finding product-market fit.
We hold a lot of meetups and work with other advisors here. We try to provide different workshops and resources to help those founders get on the right track toward product-market fit, since they still lack some talent and help on the go-to-market side.
Yeah. It’s interesting; we also host meetups. Before Diogo launched Jev, he was here in this room. We hold a sort of writers’ meetup where people can show their work, and I give them feedback. The TypeSafe manifesto came out of that.
It’s similar to what you do, except that for some people I’ll do it for equity. Most of it I do for free. Some want a recurring call, and then they offer some equity. It’s just easier to do that, and it turns out to be worth a lot of money if the company does very well.
I think it’s interesting as a job, because you can’t eat equity. Maybe they never sell. Databricks is 15 years old and worth $200 billion, and I have my equity, and it’s like, what is it? But I think it’s useful to be able to test your theories on a number of different companies.
When you work as an employee at one company, you only see one example, and your advice only works in that company. But now that I’ve worked with 36 different companies, I can see the universal truths among them, and I’m better as an advisor because I’ve tried everything.
Right.
Yeah. I’ve talked to everyone. I understand the different scenarios, like when my advice isn’t going to work for you because I tried it at another company and your situation is clearly very different.
I feel like almost all domain subject-matter experts have to do this. I happen to own developer marketing; that’s my thing. But whether you work in strategic finance or GPU compute acquisition, if you’re trying to be the global number one in that area, you have to go through a tour of duty advising everyone so you can just fucking test everything.
Yeah.
Some of it doesn’t work, but some of it is clearly going to work.
Mm-hmm.
And then you learn. Your sample size has to be more than one, and most people’s sample size is five. But if you can get it up to 50, you’re probably one of the top people in the field anyway.
I can share something I’ve found that’s quite similar among Chinese founders going global. More and more of them come from engineering backgrounds, even research backgrounds.
They think they’re building the best products because they have the best harness skills, technical skills, and knowledge. But sometimes there’s still a huge gap between what the customer actually wants, their skills, and the product they offer. Different local markets can also have different demands, even for similar things.
All those technical and research founders—I found out that—
Um, okay.
—there are also a lot of founders from the same background, yeah.
There are a bunch of ways to survey this or tackle it. The most important thing to recognize is that if you have something smart and you think it’s valuable, and you’re really the global number one—you’re the best in the world at one thing—that’s the most important thing.
Everything else is what I call “画龙点睛.”
点睛.
Right? You have the long; I’ll help you with the jing.
And most people don’t have the long.
Mm.
Right, that’s good.
Mm-hmm.
I think there’s a percentage of people who don’t even have the long, but think they do. That’s an arrogance-and-ego thing, right? But let’s assume you generally have a very good product.
Yeah.
The important thing is that you shouldn’t treat marketing as the last mile: “Hey, I did everything. It’s the last night before the blog post. Can you review it and make some changes?” There’s nothing I can do at that point. You’ve already done everything.
If you involve marketing earlier in the process, before the product is really built or specced out, you can ask: What’s the way we’re going to pitch this? Which segment of the market are we going after? Then you have more go-to-market thinking built into the product from the start.
Mm-hmm.
A lot of engineer- or researcher-defined audiences don’t have that go-to-market mindset early enough. They usually leave it until the end. The most junior part of the company becomes the go-to-market organization: some interns, or a 22-year-old running social media.
No. Try to be serious about who’s supposed to use this thing, where those customers spend their time, what problems they face, what words they use to describe those problems, and how their lives will be better after they see your product.
Right.
Those are very simple things that most researcher and engineer founders haven’t thought about. Having a go-to-market partner can be very useful. When it comes to developer tools, I find a lot of Chinese founders are in robotics.
Right.
Robotics is a very interesting space because it has a much longer time horizon, and it’s still very early. Why do all these Chinese robotics founders default to laundry?
Mm-hmm.
Unitree is the best.
Yeah.
At least, Unitree is still doing Unitree things. I’d like to see them do things you’d never expect Unitree to do. But the thing they just launched this week is doing kung fu moves.
Kung fu, yeah.
I’ve seen that before. Show me something I haven’t seen, and then I’m really going to pay attention.
Mm.
That’s hard to show in robotics. But with models, you can still do it, because it’s so early. You can also do it with innovative product ideas for consumer experiences. So I think showing the demo is very important. TypeSafe is very recent, but also very good. Diogo didn’t come out with the Doom demo, but the Doom demo is what everyone talks about.
Mm.
The Doom demo was done by DevRel, by go-to-market. That’s a core principle in marketing: show, don’t tell.
Yeah.
Show me the benefits of this thing, but don’t tell me in bullet points what it is.
Don’t tell.
Don’t give me a PDF or a presentation that tells me all these things. Just show me.
Right. And maybe a follow-up is about taste. Everyone’s talking about taste, and I think go-to-market and building the product also require taste.
Yeah, yeah, yeah. That’s probably very true.
It’s actually becoming a little hard for me to define what taste is, because investors talk about it. I wonder, in your view, how do you define taste when you see it in engineers or founders?
I can define taste in a way that’s not helpful. Taste is when, out of a large set of possibilities, you’re able to narrow it down quickly and cheaply to the set of possibilities that, looking back, are obviously good.
At the start, you look out and there are 1,000 different things you could do. Which one do you choose? You can brute-force it and try 1,000 things until you find a good one. But a tasteful person can say, “Yeah, obviously these are bad. I can’t explain it, but obviously they’re bad. This tool looks interesting.”
Mm.
At the end of the journey, they pick one, and you look back and think, “What do you mean we were considering the other 999? Obviously this was the right one.”
There are many applications of this, but the one that comes to mind is the iPhone. For a long time, smartphones had a keyboard. It took Steve Jobs to say, “No, we’ll do whatever it takes to have a touchscreen and virtualized screen, because it needs to look like a screen, like a flat slab.”
Mm-hmm.
It took 10 years of R&D, because a company called General Magic tried to do this and failed. But obviously, now every phone in the world looks like this. That’s taste. In 1990, when people were starting to have smartphones, you could do a huge number of things in terms of form factors.
Form factors, yeah.
But it takes someone with taste to say, “This is what we want. I’m going to say no to everything else, and the first phone I ship is going to be this.”
Right. Looking back at all those AI products, could you share an example that you think has taste?
I actually think everything Cursor does has a lot of taste. They were first to the VS Code thing before us, and now they’re doing Groqbot, which is a good-taste thing. Now OpenAI has to copy them, and Claude has to copy them.
Yeah.
You know you have taste when other people copy what you just shipped.
Yeah.
It’s interesting to look below the labs and think about the more boring problems people solve. Legora, in the legal field, has this massively parallel document-review process.
The legal—
Once you see it, you’re like, “I don’t want to do my work anywhere else. I have to do it in Legora.” Again, anyone could have come up with it. Ideas are cheap.
Yeah.
But you didn’t do it.
No.
Legora did. Someone at Legora had the taste to look at the billion possible things they could do and define a product that becomes the way people do their work.
Yeah. Coming to an end, we’ve talked about the AI scene in Shanghai, where you’ll be arriving in about 2 weeks. What do you hope to see or learn from the builders or developers there? What topics are you most interested in?
It’s tough, because it’s more of a differential for me. I want to see how different Shanghai is from San Francisco. I think San Francisco has a B2B SaaS bubble, but in Shanghai, you guys use ComfyUI. We don’t use it here.
Okay.
We like Canva. We like Fal.ai.
Yeah.
We don’t use ComfyUI. It’s more of an open-source tool. It’s the same with OpenClaw: it had more traction here than ComfyUI.
Mm-hmm.
But OpenClaw really has traction here.
It has real traction in China. Yeah, I bet you’ve heard that.
Yeah, of course I know. But I haven’t seen that kind of energy myself.
Yeah.
I think it’s very different. For you, a medium-sized community is 1 million people. That’s huge here.
Yeah. A million is already top-tier in China as well, but—
No, look, I’ve seen those Chinese influencer farms, with all these Chinese girls doing this with their phones.
Yeah.
The scale in China is just so different, and it’s so competitive. People say they work 9-to-6 in San Francisco, but they don’t really. On Saturday and Sunday, they’re still kind of chilling. But in China, they’re really 996.
Well, it’s kind of normal in startups and big tech as well.
Yeah, but I think that’s really intense. Places like Toutiao and Meituan—
Yeah, the same.
Right?
Yeah.
Intense.
Yes.
Right? People here don’t work that hard.
I heard that engineers or startup founders work that hard in San Francisco.
I work hard because I have the most equity. But employees—
Employees—
Legally, you cannot ask people to work that hard here. It’s illegal, okay?
Yeah.
Yeah, I’ve seen some of the most competitive, top-tier U.S. startups, and I’m sure they’re not working as hard as people in China. There’s BAT, then there’s the first generation, and then the second generation.
Second generation, yeah.
Whatever the—
Yeah, yeah.
—the thing is. Yeah, I mean, those guys—oh my God.
How did you learn… Did you—
They’re such good founders.
Yes.
They’re such good founders. If they lived in the wrong country—if they lived in the U.S.—they’d kill Amazon, they’d kill Facebook, right? The only reason they’re not richer is that they live in China.
Also SaaS, software as well.
Yeah. Even when I studied abroad at Beida in 2007, BYD was still like, “Oh, we’re up and coming. We’re China-based. We’ll always be Chinese.”
Yeah, yeah.
People love BYD here.
Oh, really?
BYD is taking over the U.S.
Oh, okay.
It’s crazy. I never would have thought BYD would come here. It’s insane. I don’t know what conclusion to draw from this. I think Chinese energy, spirit, creativity, and grit are things people don’t have much exposure to. And that’s not even talking about the model labs, like DeepSeek and all those labs. There are many reasons to be excited about exposing more people to that.
Especially with Chinese events, there are regulations against livestreaming, and they’re presented differently from the way Western audiences like to consume them, so they don’t have the same reach.
Right.
Right? Aliyun did this thing where they presented the whole Alibaba roadmap, with Qwen and everything.
Yeah.
Nobody saw it here. Why? Because it was in the middle of a corporate presentation, presented by other people. So I had to go and translate everything—
Okay.
—for the U.S. audience.
Yeah.
But just do it with me, right?
Yeah.
Just do it at AI Engineer, and people will see it.
Yeah, yeah, yeah.
Yeah.
I also saw that some open-source model companies attended AI Engineer this year in San Francisco—MiniMax.
Yeah. We hope to have representation from everybody. I think a lot of people are interested in that cross-cultural exchange and aren’t interested in the politics.
Right.
We won’t do politics; we’ll only do tech.
Yeah.
Okay.
That’s great. I’m really looking forward to it. I guess that’s all my questions, but I still have 3 quick-fire questions. Just say whatever first comes to mind. First, what AI development are you most excited about right now, and what worries you most?
The AI development that excites me most is System 1 models. That’s the clear number-one answer, so it’s boring, and I’m not going to say it. The thing that people are under the radar about is interaction models.
Interaction models?
This is something Thinking Machines Lab is working on.
Yeah.
For people who haven’t seen the announcement, look it up. It’s full-duplex, back-and-forth communication at all times. You don’t start, then take turns: you take a turn, I take a turn. No, you’re listening while you’re speaking.
Mm-hmm.
And I’m listening to you while you’re speaking. I can interrupt you, and you’ll immediately respond and adjust.
Yeah.
Right?
Yeah, I saw the demo.
That should be the way of AI, right? There’s a corollary to that, which is continual-learning AI, where you don’t have pre-training, post-training, and then you’re done. You’re learning, learning, learning as I work with you, and that’s a very interesting and important field. The last part is world models.
Mm-hmm.
They’re obviously doing very well.
Yeah.
I don’t have to explain that. As for what I’m worried about, I think we’re in an AI summer. This is probably the longest AI summer we’ve ever had in human history, but AI has been around for about 50 years. There are always summers when everything’s going to work, then winter, summer, winter. If we ever have a winter, when is it going to happen, and am I going to go bankrupt? That’s a real concern for a conference company that relies on hype.
Yeah.
More broadly, there’s a real concern about safety. My friends at all the labs are really worried that society can’t handle or adjust quickly enough. It’s not because of fake news. It’s not because of people losing their jobs. It’s something worse. Right now, the way we express it is cybersecurity.
Mm-hmm.
It’s like, “Oh, I get hacked all the time.”
Mm-hmm.
Every day it’s hack, hack, hack, hack, hack. I feel like there’s something even worse on the horizon that we don’t know about. The most extreme version is bioweapons, right? “Oops, I’ve invented COVID-2.”
Mm-hmm.
What if it’s worse than that?
Mm-hmm.
The way we phrase it on our podcasts is that the default assumption for most people should be AI doom. In all of human history, when a more advanced civilization comes along, the less advanced civilization goes extinct. This thing is evolving roughly 1,000 times faster than us, so at some point it’s going to be more advanced, and then we have to figure out what to do.
Mm-hmm. Cool. The second question is: what other newsletters, podcasts, or content do you recommend?
I have a list called “Good AI Newsletters and Podcast List.”
Okay. Is it on the—
It’s on my GitHub. For something more recent, you can check aeo.latent.space. I asked every model what the best AI podcasts for engineers are—
Mm-hmm.
—and it gives me a list. I transcribe the results and put them in a ranking.
Oh, cool.
I asked 200 of these questions.
Okay.
What’s the best payment startup? What’s the best bank? What’s the best financial advisor?
Oh.
What’s the best database? Everything—every question I can think of.
That’s why you call it AEO?
Yes, because I’m agent-optimizing for what the models answer.
Right.
It’s probably a pretty decent proxy for what the people choosing the datasets in the labs choose. And they choose Latent Space.
Okay, the top one. The first one.
It does very well. It’s either number one or somewhere between number one and number three. So, yeah, that’s the answer.
Cool, cool, cool. I’ll leave the last question for you to ask me, if you have any questions, and then we’re here—
What are your hopes for AI Shanghai?
Of course. I hope to hear more about topics like the infrastructure stuff you mentioned—for example, safety, payments, and things that unlock more applications through different kinds of infrastructure, like agent infrastructure. I’m very curious about those topics because they’re discussed less in China. But after spending a few weeks here, I’ve found more people talking about them, including people working on RSI and other areas of frontier research.
I hope it can be a great conversation between the East and the West, between the researchers here and the people in China. I hope it’s successful, and that we can keep hosting it next year.
I hope so too. I wish the team a lot of luck doing the first one of anything.
Yeah.
The first one is the hardest.
Right. Agreed.
Yeah. I’m so impressed that they’re doing it.
Thank you so much, Shawn.
Yeah. Thank you.
Thank you.