AI Builders
Digest

The Takeaway: The AI that will actually benefit everyone is not the million-dollar-per-query super model, but personal AI that runs locally, continuously adapts to the individual, and returns control to the user.

Igor Babushkin, former DeepMind (StarCraft, AlphaCode), early OpenAI reasoning work, and xAI co-founder (Colossus and Grok), has left to start River AI. He describes the late-2025 coding-agent breakthrough as the moment software engineers could no longer ignore agents—the inspiration for his two pieces of AI fiction. The next frontiers are scientific discovery (closing the real-world experiment loop) and everyday personal agents for ordinary people, not just the most verifiable domains of code and math.

River is betting on three research projects: a high-engineering RL/fine-tuning API for companies, personalization that lets each user’s agent behave differently, and local hardware that brings frontier inference into homes and offices for control, latency, and privacy. He argues closed-source labs are in a tight spot—models may become too capable to release, while open weights keep closing the gap.

“We’ve all become the sorcerer’s apprentices,” he says. “LLMs and agents are only getting stronger, and personal AI is the next manifestation of that impact.” The goal of alignment should be maximizing human flourishing rather than letting machines take full control.

Igor Babushkin argues that diminishing returns on pre-training scale, regulatory pressure, and rapidly improving open models are squeezing closed-source labs such as OpenAI and Anthropic. Models may become too capable to release while open weights keep closing the gap. The exit is to innovate new forms of distribution and post-training so companies and individuals can do local post-training instead of handing all data to centralized APIs.

He strongly supports the US producing the world’s best open model and remains clear-eyed about the current Chinese-dominated open-weight ecosystem: open weights lower the barrier, yet also introduce licensing and potential control risks.

Claude Blog: Claude Code now supports artifacts. Starting today, Claude Code can capture work progress as an artifact—live, shareable visual pages including PR walkthroughs, system explainers, dashboards, and release checklists that update themselves as the session continues. Artifacts are built from the full session context (codebase, connectors, conversation) with no extra infrastructure required. Pages open in any browser or the desktop app; teammates see the same link and the same version, with in-place refreshes on publish. Private by default, shareable within the org, with version history and a compliance API. Available in beta to Claude Team and Enterprise orgs via Claude Code CLI and desktop app.

Vercel CEO Guillermo Rauch highlights that AI Gateway gives companies the critical infra to make AI a productive investment: budgets per key/team/project, failover for maximum uptime, model and provider choice, and realtime observability. He tells those still in a “token-maxing fever dream” it is time to wake up. He also notes that software projects are transitioning to agentic software factories: the Issue → Agent → PR → Release loop will become the norm, and the job of the author/maintainer is to work on the loop that yields the highest quality product and sets the criteria for what should be worked on.

Linear Head of Product Nan Yu shares that the most common loop inside Linear is some variant of Issue → Agent → PR → Release; roughly 30% of bugs make it all the way through. Key nuances: instructions must ask the agent to extensively research the root cause and use Datadog and Sentry MCPs for evidence; it should only attempt a fix with high certainty, otherwise tokens are wasted; if more evidence would help (repro steps etc.), it should leave a comment requesting it. Agents need to be told to follow good practices, just like people.

Box CEO Aaron Levie notes that directionally it is clear the harness is going to become the most important variable—right next to model capability—in the AI stack. The ability for harnesses to break down work efficiently and route to the right model at the right time will be huge for maximizing accuracy and reducing costs. When tasks move from hundreds of thousands or millions of tokens to tens or hundreds of millions, this becomes a major variable. We are still incredibly early; huge opportunity ahead.

Y Combinator President & CEO Garry Tan shares that your personal AI or company brain needs a clean harness; this is the one their team built and uses every day, now free and open source.

smol.ai and Latent Space’s Swyx notes he seems to be in the minority among AI leaders still actively using /loop and /goal. He thinks those who stopped are wrong—not forever, just giving up too early in the Grok 5.6/Claude 5 era. Use it when you want the right mix of steerability and autonomy, or an open-ended “loop that generates loops” end state without deeply specifying the path. He shares an example where having a goal saved a very long action-reasoning turn.

OpenClaw and OpenAI-affiliated Peter Steinberger observes that queue was the way, but with 5.5 the model doesn’t get confused anymore—you can just throw stuff at it while it works and it will diligently process whatever you throw at it.

Replit CEO Amjad Masad shares an 8B model that has reached ~1500 Elo, consistently beating frontier models and Stockfish level 0. Fun to see it mogging GPT 5.6 with high reasoning and response chaining, spending 1-2 seconds per move versus 30 seconds. Playable now.

OpenAI Codex & ChatGPT lead Thibault Sottiaux announces that to celebrate a week of efficiency and let users run 100,000 Luna threads this weekend, usage limits for Codex and ChatGPT Work have been reset. He also posted “Optimize for curiosity” and a ChatGPT Work count update.

OpenAI CEO Sam Altman shares a cool ChatGPT Work use case: connect family calendars and explain kids’ interests, then every morning generate a podcast covering one kid’s soccer game that afternoon, an upcoming birthday, some news, etc. He also responds to Moore’s Law with “i see your moore’s law and i raise you 20x” and “it could be faster.”

Every CEO Dan Shipper contributed the kicker to a Wall Street Journal piece on OpenAI vs Anthropic and stands by the view that momentum has been shifting to OpenAI since early spring—a fascinating comeback story. He also jokes about 2027 human programmer interviews: describe the last 3 unresolved mathematical conjectures you solved and share your prompts; describe the last cyber felony your agent unintentionally committed and mitigation steps; tell us how many R’s are in strawberry and how many e’s in seventeen.

Builder Zara Zhang shares a fascinating interview on how Claude has changed work at Anthropic: 65% of PRs by product & eng teams are now raised by Claude. She has watched her own agent interface evolve over six months: January terminal → March desktop app (e.g. Codex) → June work collaboration tool. Each step gets closer to how humans naturally communicate; the agent should meet the user where they are. She also notes that what feels obvious to you is not obvious to most people, and if you want to post but don’t know what to say, just do your normal work and talk about it.

Swyx notices that the pejorative connotation around “vibe coding” has completely disappeared since roughly everyone, from nontechnical to supertechnical, is now doing it.

FPV Ventures partner Nikunj Kothari has a new essay pondering the quiet belief in venture that the best founders are running from something, and the gear of drive most people never reach.