There are maybe 5,000 people who can train frontier models. The rest need AI engineers

The episode opens with a party trick that fits the subject. Nathan Labenz is at a wedding, so an ElevenLabs clone of his voice reads the introduction. The guest is Swyx — Shawn Wang — who runs the Latent Space newsletter, podcast, and Discord, and who has just organized the first AI Engineer Summit for October 8–10. Labenz wants a survey of the year: what the new job is, whether agents work, whether OpenAI will eat the app layer, and whether a cable-style bundle of AI apps could exist.

Swyx came to software after years as a trader and hedge-fund person, then spent five years in developer relations. He treats the talent shortage as a supply-and-demand problem, not a branding exercise. There are maybe 5,000 good LLM researchers in the world. The average company, he says, will never hire one. Models are now APIs or open weights, so demand shifts to people who can put them to use without doing foundation-model research.

That is the job he has bet Latent Space on. He is less interested in coining a name than in observing which name will win. Cognitive engineer and LLM engineer have been tried. AI engineer, he thinks, is the one that sticks: a software engineer who specializes in AI and the emerging stack. Everything after a model is in production, plus a little fine-tuning — extra training on top of pre-training already done — is their territory.

The picture he wants in your head is a left-to-right spectrum. Far left: research scientists. Next: ML engineers and data scientists. Then the new slot, AI engineers. Far right: full-stack people doing the last mile of UI. Traffic, he says, mostly runs right to left. Software engineers start learning what a tokenizer is, what an embedding is, why chain-of-thought beats zero-shot. The reverse commute exists but is rarer.

Raza Habib of Humanloop is the counterexample he names. Habib has a PhD in probabilistic programming and still chose an LLM-ops product, because more people can be served that way. Labenz, a Humanloop customer since early 2023, adds a concrete trick the platform made easy: use GPT-4's reasoning traces as fine-tuning data for GPT-3.5. The jump in quality, he says, was dramatic.

Labenz counts himself in the new category, from a nonstandard path. What employers actually need, Swyx says, is someone who can take a product idea and say whether it is buildable this month. That is a different skill from keeping up with the Discord jargon. Most people, he thinks, should cover the fundamentals, then build the projects that already make money. Being cutting-edge is a full-time job of its own, which is part of why the title deserves to exist.

Latent Space University is a free seven-day email course, about an hour a day, aimed at software engineers who have never touched these APIs. Day one is the GPT API, including the context limit that ChatGPT and Copilot hide. Day two is prompting and memory: vector databases and LangChain, with a nudge to write the abstractions yourself first. Days three through five are code, images, and speech-to-text. Day six is fine-tuning and running open-source models. Day seven is an agent.

The graduation test is practical. When a PM or CEO walks over with a product idea, you should be able to say yes or no and why. Labenz's version of that skill is an XKCD reflex: people post that GPT-4 cannot do X when they used the wrong model or the wrong prompt. Staying on the actual frontier is a different job. Those people hide in Discords and on Twitter, using jargon that will not be public for months, and the valuable rooms are private channels inside the private channels.

On tooling, five years of developer-relations work make Swyx vendor-friendly, and he is also sympathetic to "choose boring technology." The rule he repeats is buy first. Plug into a community that has already hit the bugs. The moment you understand your domain, be ready to build, because the field is immature and you may have to rewrite. The 2023 LangChain summer was the live lesson. People started with it, hit walls, ripped it out, and declared it crap. He thinks that is unfair to a company less than a year old. "If you chose to be in this field, you chose to be on the cutting edge, and sometimes the cutting edge cuts you."

The table-stakes names on their show notes are Humanloop, Guardrails, and LangChain. Humanloop owns the prompt lifecycle: workshop in a playground, hit save, get an API, change the prompt without a code deploy, log everything, export the good traces for fine-tuning. Guardrails sits on output validation, including checking that generated SQL is legal. LangChain sits on orchestration. The awkward fact is that every LLM-ops company has the same roadmap — prompt versions, evals, cost, latency — and OpenAI's JSON mode made a whole class of validators obsolete overnight.

When companies ask how to hire for a job they cannot yet define, Swyx points at the same curriculum and then past it. Twitter, the Latent Space Discord hiring channel, Hacker News "Who is hiring," and the conference floor are the matching surfaces he actually uses. The skill list is the floor. The thing you cannot test in a take-home is whether the person is entrepreneurial enough for a stack that changes every month. They have to walk in with ideas, not wait for someone to throw an LLM layer on an existing app.

The Vercel team is the role model in his AI-engineer post: not a foundation-model lab, still shipping things that attract millions of users on free projects. People who only rebrand as AI engineers after a data-science title, he says, will not put a company on the edge of what generative models can do. The job is poorly specified on purpose. Prompting is poorly specified. The model list turns over every month. Comfort with that fog is the filter.

If the door is that open, why hasn't every software engineer walked through it? Swyx lists three hangups. Some people got burned by crypto and treat every new wave as a fad. Some assume you need a PhD. Some are still stuck on the stochastic-parrot argument, sure that multiplying matrices cannot approach intelligence. His reply to the credentials panic is a calendar. Transformers are six years old. GPT-3 is three. Most of the companies in the stack are under a year old. Everybody is uncredentialed.

Andrej Karpathy, formerly Tesla's director of AI, has backed the title in public: there should be more AI engineers than ML engineers, and they can be very successful without training anything. Swyx writes a monthly recap and jokes that being factual makes him an "OpenAI maxi." The OpenAI news block takes two screens. The rest of the frontier-model forum is five bullet points. Labenz's own answer to the fad question is code: Copilot and GPT-4 are already a multiple speedup in the editor, and Nat Friedman has about six orders of magnitude of scale left in the decade.

The summit is sold out on the floor, still open online, and built as a ten-year bet. Machine-learning researchers have ICML. Engineers, Swyx says, have nothing equivalent, so he is trying to grow one. Twenty-eight speakers, AutoGPT as presenting sponsor, OpenAI, Microsoft, Notion, Amazon, and GitHub on stage, plus first public demos from Adept and New Computer and Lindy's first public talk. He mixed in people who already run this in production: Eugene on Amazon Books with language models, and the author who had just finished the O'Reilly book on LLM ops.

He splits agent projects into two cultures. Open-source lines like AutoGPT and BabyAGI optimize for the most impressive thing you can fit in a tweet or a short video. Closed companies — Adept, Lindy, New Computer — chase mundane daily tasks that might support a $20 or $100 subscription. Both are valid. The cold water is autonomy. AutoGPT asks for confirmation before every action. That is not autonomy. That is assisted prompting. The destination is fire-and-forget, the way he already gets into a Cruise or Waymo in San Francisco, says nothing, and arrives. Agents, he thinks, may be five to ten years from that, about where self-driving was a decade ago.

The most useful "agent" he uses today has no LLM in it at all. Calendly and SavvyCal send a link, the other person picks a time, reschedules happen on their side, and he never touches it. That is the feeling he wants from every agent. The loop that has already proven itself is Code Interpreter: generate code, run it, read the output, decide whether to fix or stop. Planning, tools, memory, and a way to touch the outside world sit on top of that loop. For a survey he points to OpenAI researcher Lilian Weng's agent post. A narrow case he likes is CodiumAI, an Israeli company that raised an $11 million seed to do one job: wander a repo generating tests for a human to accept or reject.

The fight is issue-to-PR. Sweep.dev and others promise that a ticket becomes a pull request. Swyx has only seen copy-change demos work; complicated tasks never do. Software architecture and turning requirements into code, he is happy to call five years out. Labenz takes the other side. Google's Med-PaLM M produced chest-x-ray reports that were preferred to a real radiologist's 40 percent of the time. If a problem that hard is already near expert, he expects complicated prompt-to-PR inside two or three years. They agree to refine the wager offline. Swyx adds the acceptance test: a non-technical person managing a project by filing issues. If you still need a technical person writing YAML and nursing prompts, you have invented a new programming language. AGI-in-two-years talk he will not take. It is not falsifiable. It is dinner conversation.

Labenz describes himself as an adoption accelerationist and a hyperscale pauser: take the gains from current systems, stay nervous about putting a million times GPT-4's compute into one model. That is close to Eric Drexler's comprehensive AI services picture — many narrow, even superhuman systems as a buffer against a general one. Swyx's version is a slogan. Programmers put the shoggoth in the box by shrinking models to a domain and composing them into software that keeps humans first. Nobody is against that path. Asked for p(doom), he says somewhere between 5 and 95 percent and refuses to narrow it. Five percent is already enough to worry; 95 percent is not a reason to lie down.

The product fight he is sure about is multimodality. The summit will host the first public DALL·E 3 and GPT-4 Vision talk. Labenz had just added a feature to a React app he had never written in, pasting a file tree and screenshots into GPT-4, which found the defect in the picture and patched the code. Vision, Swyx says, is the more impressive sense in a demo. Voice is the one you will use every day: hands-free, always on, patient. Greg Brockman had just tweeted a ChatGPT voice session teaching Russian, with the line that Duolingo should not exist anymore.

That line is what makes Labenz nervous. He lived in the same college dorm as Mark Zuckerberg and watched Facebook invite every developer onto the platform, then ship the high-value cases in-house until the leftover product was Log in with Facebook. He sees the same movie starting. OpenAI has a fine-tuning jobs page that looks like Humanloop, ChatGPT Enterprise with native retrieval over Google Drive, Notion, and Dropbox, plus voice and vision in the app. All of it, he expects, is coming to the API. Sam Altman did promise not to compete with developers, in a conversation with Raza Habib whose transcript was later deleted. Swyx's verdict: the promise is not ironclad, and they do not want to stand behind it. Every large platform hits this tension. Apple even has a name for it. Sherlocking.

He does not treat that as a dead end. OpenAI is about 600 people, not a thousand, and they will spend those people on big swings — the rumored Jony Ive phone, not an A/B-testing fine-tune console. The best life for a startup next to a foundation lab is to prove it is a strong engineering team. If you are good enough they will buy you. Money is cheap for them. He picks ElevenLabs for ease of use and PlayHT when he wants a wider emotional range; OpenAI, in the short term, will not bother with horror-movie and shooter voices. On chat product craft he actually ranks Character.ai higher than OpenAI, even without the best model. His own research bet is a GPT-4.5 thesis: long inference. What happens if you give the model a year to think instead of a few milliseconds? Underneath that, text remains the universal pipe. Speech and images still pass through it, and humans understand high-quality text data better than any other modality.

Labenz had pitched OpenAI a quality-of-life idea: Log in with OpenAI. Waymark, his video company, spends about 15 cents in hard cost on every free trial. He would rather the user carry that cost. Swyx thinks the idea is reasonable and may still miss OpenAI's bar for interesting work. Building it would also announce that the lab is becoming a fourth public cloud, which is a different identity from a research organization. Character.ai, not OpenAI, is the product he cites for context that lives with the user.

His September recap had a hardware theme: AI pendants. Rewind's necklace, Avi Schiffmann's Tab, Humane's AI Pin, all pointed at all-day capture. He is already wearing an Oura ring that only logs a heartbeat. What he wants is a log of every conversation and a digital twin he can talk to, at least as a note-taker, maybe later as something exposed to the world. Those objects will be multimodal. If there is a B2C form factor after the phone, he says, the AI engineer will build it. Next quarter's summit, he already wants a whole stage for that hardware.

Then Labenz unloads the bundle. His personal stack — ChatGPT Pro, Claude Pro, Copilot, Replit Ghostwriter, Perplexity Pro, Rewind — already rivals an old cable bill and still misses hundreds of $20 apps he will not buy. On the supply side, Waymark needs one paid user on a $30 plan for every 200 free trials just to cover tokens, and users cancel the minute they have the video they wanted. The pitch: $100 a month for a thousand AI apps, flagships up front, long-tail tools like Waymark behind them, each app taking an average of 10 cents per subscriber. A million bundle subscribers would mean a million dollars a year for Waymark before any upsell.

Swyx does not kill the idea. The closest existing object is Setapp, the Mac subscription bundle. The hard questions are the split — paid whether or not anyone opens the app, or paid on usage — and the politics of pricing. SaaS companies want to own the customer; they will not enjoy asking Mr. Bundle for permission to change a price. The San Francisco infrastructure shops he knows are chasing $10,000 to $1 million annual contracts, not a $20 month. A bundle only fits Waymark-like B2C products. Nathan Baschez of Lex and Ben Thompson of Stratechery have thought about this math longer than he has. His advice is not to sit in the argument. Find the smallest version that produces information. Maybe five apps. Labenz already knows the other constraint: without ChatGPT Pro, Claude Pro, and Perplexity Pro as anchors, three hundred unknown apps for $10 will not sell.

A mentor's line is the one Swyx leaves Labenz with on the bundle, and it applies to the rest of the hour. Whenever you are stuck arguing that something could work, take the smallest step that produces information. Five apps in a pack will tell you more than another essay about cable television. The same rule sits under the seven-day course, the buy-then-build tooling advice, and the hiring filter: do the hour of API work, keep the module that still feels wrong after a few weeks, and recruit the person who already shipped.

Tickets for the in-person summit are gone. An online ticket still buys the livestream and a Slack at ai.engineer, a domain he enjoys saying out loud. The conference, the newsletter, the job board, and the Latent Space podcast are all pointed at one claim. Machine-learning researchers already have ICML. Engineers did not. He is treating this as year one of a ten-year commitment to give them one.

What the conversation keeps returning to is a division of labor. Foundation labs will train and, eventually, Sherlock. Agents will look like toys until they can be fired and forgotten. A $100 bundle will not happen until someone puts ChatGPT, Claude, and Perplexity in the same cart and measures churn. The models, Swyx says, belong to someone else. The hands that turn them into products are becoming an industry, and that is the bet he has put a decade on.