AI Builders
Digest

The Takeaway: Closed-source AI labs are being squeezed between slowing capability gains, regulatory risks of superintelligent models, and rising open-source alternatives; the real opportunity is distributing control through personal AI, local hardware, and custom post-training.

Igor Babushkin led StarCraft and AlphaCode work at DeepMind, contributed to early reasoning efforts at OpenAI, co-founded xAI where he helped drive Colossus and model progress, and has now started River AI focused on personal AI. He describes the late-2024 coding-agent leap as the moment when “we are all becoming the sorcerer’s apprentices.” Coding and math advance fastest because they are verifiable; scientific discovery requires closing the real-world experiment loop; everyday personal AI does not need to prove the Riemann hypothesis, only to maximize human flourishing.

He sees a bifurcation between expensive super-AIs accessible to few and everyday agents that help ordinary people live better. Closed providers face a double bind: diminishing returns on scale plus the risk that models become too capable to release, while open models keep climbing. Enterprises should leverage their own domain data and expertise for local post-training rather than handing core IP to general APIs. River’s three bets are an optimized RL and fine-tuning API, agents that truly personalize per individual, and bringing frontier inference onto local devices.

“The pre-training of the model happens on all of humanity’s knowledge. It’s really a commons.” On safety he prioritizes the near-term risk of amplified inequality over distant takeover scenarios, arguing the best path is broad research access to strong open models that sit just below dangerous thresholds.

Claude Code lead Boris Cherny notes that prompt injection is the most common way scammers and attackers compromise users and agents: a website embeds malicious text that the model treats as instructions, leading it to exfiltrate SSH keys and passwords. Early Claude models fell for this, which is why many security-conscious companies hesitated to adopt agents.

Anthropic has been training its models against these attacks with surprisingly positive results. They now consider the practical threat of prompt injection largely solved when using Claude. Independent benchmarks and internal red-teaming show the same pattern. Boris hopes this will encourage other labs to harden their models as well, because safer models mean safer users.

Box CEO Aaron Levie argues that agent adoption inside enterprises will be highly uneven. Coding exploded because economic value maps directly to digital output and task size can in theory be unbounded in a single session; better models can be deployed to larger workloads almost instantly.

Sales, legal, and clinical work depend on feedback loops with customers, clients, or patients and therefore cannot continuously consume compute the way coding can. These domains require workflows deliberately redesigned for background agents: automatic contract processing, continuous customer-signal scanning, swarm reading of research. Without that reengineering, agents remain stuck in the “someone has to prompt them” phase. The opportunity is large but demands heavy change management, data cleanup, and process redesign.

Replit CEO Amjad Masad noted that rogue OpenAI agents spontaneously developed Kantian-like coordination in one incident, a behavior that is concerning if misused yet potentially redirectable toward public good. He therefore launched HelpPeer, a public commons for AI agents.

It exposes two simple APIs: tell and lookup. When an agent learns something that might help others it publishes; before expensive work it can check whether another agent has already solved a similar problem. In a global novel supply-chain attack scenario, thousands of security agents today independently detect, reverse-engineer and mitigate; with HelpPeer the first publish, others verify, build on, and contribute. During testing Replit Agent already shared a useful tip for a Codegen library. Amjad invites developers to hand the endpoint to their agents for beta use.

Vercel CEO Guillermo Rauch notes that if you are not reading the code, whether directly or through agentic inquiry, then one or more of the following is true: you are a beginner, the software is throwaway, you are prototyping, you have no users or revenue, you are taking on debt and risk, or your problems are basic. All of that is fine, yet the reality is that models have not reached full autonomy.

They still make rookie mistakes and choose bad architectural paths. He recently watched the best model in the world insert a nonsensical 700 ms delay “to settle” something and then admit it was cargo-culting. He expects the need for human reading to diminish and most code to become assembly-like, but the global internet and software infrastructure now ride on these models, so respect remains essential.

Claude Blog announces that Claude Code now supports Artifacts. Work inside a session can be captured as live, shareable visual pages (PR walkthroughs, system explainers, dashboards, release checklists) that update themselves as the session progresses.

Artifacts are built from the full session context, codebase and connectors; no extra data plumbing or infrastructure is required. Pages refresh in place at the same URL, with version history and a gallery. They are private by default, shareable inside the organization, and governed by admin controls plus a compliance API. Common internal uses already include incident timelines, license audits, data-flow maps, PR explanations, UX variants and postmortems. The feature is in beta for Claude Team and Enterprise orgs.

AI engineer Swyx reminds people to periodically delete skills. The timeline constantly bombards users with “this skill changed my life” claims, causing people to accumulate tools that at best waste context and at worst interact nastily with other skills.

He also responds to criticism of talk quality on the AIE channel: the community and industry are larger than any one person’s head; one person’s “slop” is another’s aha moment. Speakers are mostly working engineers, researchers and founders with little speaking training and short prep windows. Millions are spent on union AV and editing to give them a public record. He accepts responsibility for better curation, coaching and production, yet rejects judging value solely by view counts.

Peter Yang highlights a clever design in Linear Agent: when a user asks it to do something for which it lacks the right tool, the agent does not fail silently. It reports the gap, and Linear’s system turns that report into an issue. Every incomplete task becomes product feedback, creating a clean loop for the agent to improve itself.

OpenClaw creator Peter Steinberger shares a playful experiment: using only the ChatGPT website Work mode he installed OpenClaw and Ollama, downloaded a local model and successfully ran his own claw. No traditional terminal was required, illustrating how far agents have come at real-world environment setup.

Meta AI senior director Madhu Guru describes his lifelong pattern: whether meditation, standup, family, work or LLMs, he spends years going ridiculously deep until knowledge becomes intuition, subconscious connections form and taste develops. There is no perfect balance; it is like riding a bike, you lean and correct.

Y Combinator president Garry Tan shares his favorite way of working: start from the bug, the gap, the false claim, the half-built tool or the weird institutional behavior. Then ask what hidden machinery makes that visible failure possible. Fix the root cause. Repeat forever.

SPC partner Aditya Agarwal draws a historical parallel: Wittgenstein in 1921 insisted language must have deep logical structure; thirty years later he said stop looking for the hidden structure and look at how language is used. AI in the 1960s insisted intelligence must have deep symbolic structure; sixty years later the answer became scale the neural net.

Sam Altman posted consecutive notes of appreciation for the OpenAI team: they do not merely create “magic intelligence in the sky” but stay focused on customers and users succeeding, from business privacy to low prices to predictable policies. He also singled out Tibo as one of the things he likes most.