AI Builders
Digest

The Takeaway: The real bottleneck is not more compute or larger models, but architectures that cannot learn continuously at test time; the path forward is new architectures plus a highly automated research lab.

Core Automation co-founders Jerry Tworek (former OpenAI VP who led the Strawberry and reasoning teams) and Rohan Anil (one of the four Gemini pre-training leads, previously at Google Brain and Anthropic) argue that today’s models still fail on real-world tasks because of the gap between lab training and deployment. Reinforcement learning has been scaled, yet it is not the final form of learning from experience; richer algorithms closer to how humans actually learn are needed.

Jerry’s core point: “The first step to replacing transformers is appreciating deeply how far they were able to carry us.” Transformers are computationally shallow and rely on one-token-at-a-time chain-of-thought, making inference-time scaling inefficient. What is required are systems that continuously adapt on user data and real distributions. Big labs, locked in quarterly model races, have little appetite for high-risk architectural bets, which is why the pair founded an independent lab whose mission is to become “the most automated lab in the world.”

Rohan stresses that optimization and architecture must be co-designed end-to-end; current pre-training + RL pipelines still leave orders-of-magnitude efficiency on the table. Automating high-performance kernel generation is a key inner-loop bottleneck: a human-plus-search competition already delivered a 60× speedup on QR factorization that no existing model can match.

Their practical success metric is simple: go on vacation as a team and see whether the lab produces better work while they are gone.

Box CEO Aaron Levie highlights a counter-intuitive dynamic: the “hardest” domains—math, cyber, and code—are actually the first to be automated because their outputs are objectively verifiable. Clear reward signals accelerate training and allow scalable testing at inference time. In contrast, legal drafting, marketing campaigns, sales messaging, and financial planning lack instant verifiability, depend on subjective risk tolerance, and often cannot be judged correct for months. Even as model capability grows exponentially, substantial applied-AI work and process redesign will still be required, and we may need entirely new methods to “test” knowledge work.

Every CEO Dan Shipper maps the psychological cycle that follows when a model suddenly performs a task that once required your continuous involvement. He calls it an “agency rupture”—you are no longer required and had no say in the matter, a kind of identity death. The pattern is predictable: (1) initial rupture, where only the AI is visible and language becomes “AI solved X”; (2) seeing the scaffolding, the human prompting, babysitting and correction work; (3) agency reconstruction, where the model’s contribution fades into the background and you simply say “I did this.” The ability to metabolize each rupture into playfulness and curiosity is a leading indicator of who thrives in the new AI economy.

Vercel CEO Guillermo Rauch reports that the company’s internal agent @v now sits at the center of daily operations across finance, communications, docs, marketing, engineering and analytics. It maintains per-user memory and workflows and continuously improves. Rauch stresses the difference from third-party “Big AI Slack bots”: if agents become the foundation of modern companies, full control from source to runtime to data is non-negotiable. @v also acts as a router that delegates to specialized sub-agents, solving the UX problem of every team spinning up its own isolated agent.

Replit CEO Amjad Masad has deployed his LLM chess engine on Lichess, where it currently sits at 1253 Elo and plays multiple concurrent games against humans and bots. The system runs fully autonomously and games can be watched live on the site.

Practical AI educator Peter Yang asked Nous Research co-founder how Hermes avoids accumulating low-quality skills. The answer: a background process called Hermes Curator continuously audits skills and memory, removing inefficiency. Because the system is open-source, users can supply their own definition of “slop” and the curator rewrites its cleanup loop accordingly.

Andrej Karpathy posted a follow-up to Simon Willison’s “pelican on a bicycle” test, uploading the source so the prompt is playable directly in the browser and easily forkable for further experimentation.

Anthropic Claude Code engineer Thariq notes that AI has made mathematics easier to understand, yet the volume of mathematical activity has increased and experts now spend more time discussing ideas at higher levels of abstraction. Demand for people who can truly think mathematically is therefore rising—an outcome he sees as parallel to what happened in chess.

FPV Ventures partner Nikunj Kothari observes that early-to-mid-stage venture has become pure “vibes capital”: rounds with little substance close easily if the sector is fashionable, while fundamentally stronger companies struggle. He expects the dynamic to persist at least 12–18 months. Public markets show the same mood swings—even trillion-dollar names move >5% on model releases. Long-term fundamentals still win, but short-term survival requires understanding the new game.

Y Combinator President & CEO Garry Tan argues that growth is good and that AI will unlock unimaginable economic expansion—the strongest white-pill narrative available. He also notes that true meritocracy judges the territory (did you make something people want?) rather than the map of titles and stories.

Cursor designer Ryo Lu reflects on how early products such as Rdio, Mailbox and Apple taught him patterns that made software feel simple and tactile, then asks what parts of software will remain visible—and how they will feel—once the discrete-app paradigm recedes.

Vercel CEO Guillermo Rauch notes that AI alone is cool, but mastery combined with creativity and AI reaches an entirely different level. He urges people not to be discouraged from pursuing excellence and craft.