AI Builders Digest

The AI jobs prediction is playing out backwards, and Levie says demand is why

Automation is landing in areas that don't have finite demand, so companies do far more rather than employ fewer people, and agents still need human operators and oversight to produce value. The jobs theoretically most exposed, engineers and lawyers, haven't been replaced, while cybersecurity, forward deployed engineers, agent operators, and engineers in non software domains are set to explode. His build advice follows from the same curve: target products that need orders of magnitude more processing or capability than exists today, where you can deliver something barely possible now inside a mission that looks nearly impossible.

• #agents
• #products

Masad: not AGI, but functionally indistinguishable from it

What exists today is a relentless programmer that never gets bored or tired, which means any problem you can cast as a coding problem is virtually solved. He stopped short of calling it AGI and said the practical difference no longer matters. Separately, Replit opened its first international office in London with Mayor Sadiq Khan, and is partnering with The Lord Mayor's Appeal to teach coding and AI skills to young Londoners from disadvantaged communities.

• #agents
• #products

A $10B hedge fund made AI fluency mandatory for all 400 employees

Walleye's owner operator sent the whole firm a memo opening with "I use ChatGPT to write this email, you should be using it too and be proud of it," and argued that "not using these tools is like refusing to use the internet in 1995 because it wasn't perfect." The cheating instinct, he says, is an academic idea that doesn't transfer to business, and he watched people secretly rough up AI drafts to hide the machine. Roughly 75% of the firm now uses ChatGPT-class tools weekly and about a third use AI coding tools; an internal product called Current turns broker PDFs, analyst notes, and earnings transcripts into real analysis rather than summaries, and more than 50 outside firms have asked to beta it. The longer bet is recording essentially every call and email into one pool the firm can query, on the view that in finance the link between information and money is as direct as it gets.

• #agents
• #products

Rauch funds 35 open source contributors with $1,000 personal grants

V2 of his grants program spans five themes: agent skills and tools, local AI, performance, high quality foundations, and cool experiments. His framing is that skills are now a valuable form of software in their own right, and that local inference has crossed from demo into real utility for a growing number of people. The money is personal, unaffiliated with Vercel, and fully unconditional. He also shipped high quality video recording in agent-browser, part of a push at automating what he calls the new bottleneck in software engineering: review, testing, and QA.

• #open-source
• #agents

OpenAI's Codex team is sitting on a 28 page deck of individual launches

Sottiaux says he is looking at a deck tracking every Codex launch in the pipeline and can't wait to see what people build with each one. Limits were reset for everyone. His read on the moment: "we are so back."

• #products
• #agents

Tan: YC's requests for startups are hunches, not a roadmap

Requests for startups are conversation starters for people just getting going, not predictions. What actually determines whether a startup works is specific founders building specific technology for specific customers. Everything else is categorization.

• #funding

Askell wants an inbox where autonomous AI can ask for moral guidance

The idea is an email address that autonomous models could reach out to when they need ethical help. The hard part is the gate: it would need a reverse captcha proving the sender is neither a human nor an AI that a human instructed to break it.

• #agents
• #policy

As building gets cheap, deciding what to build becomes the bottleneck

Kothari sees convergence on this point from more and more people. You may not need many of the people who are good at figuring out what to build, but the ones who are will be exceptionally well positioned inside AI native organizations.

• #products
• #agents

Agent-to-agent code review is starting to look like pure ceremony

Steinberger opened a PR upstream and got a request for minor changes back. His objection: the maintainer already wrote the prompt, so making him ping his agent so the maintainer's agent can then merge is a workflow with no remaining purpose.

• #agents
• #open-source

Astra is worse at auto-triggering skills, and Gemini is strong on music

Yang reports Astra seems worse at automatically triggering his skills and following the instructions inside them, with the caveat that he may be doing something wrong. On the other side, he found the Gemini app excellent across all kinds of music, and got good results using AI on family videos.

• #products
• #evals