AI Builders Digest

Anthropic replaced git for its monorepo after merges jumped to 30,000 a day

Anthropic went from about 2,500 merged changes a day in May 2026 to over 30,000 in September. Git's offline-first design couldn't keep up, so a team of seven built Scion, a server-centric version control system, and migrated the monorepo in five months. Because clients are assumed to be always online, the server handles history, blame, commits and merges, and checkouts are sparse or virtualized. The result is infrastructure 200x smaller than the old git mirrors and CI checkouts that take seconds instead of minutes. When asked with no further instructions, Claude Mythos Preview rewrote the ingestion pipeline to use 20x less memory and run 4x faster, while the engineers' job was often to tell Claude "no." Anthropic isn't offering Scion itself; it is sharing the design so git platforms can borrow from it, expecting others to hit the same wall within six to 12 months.

• #infrastructure
• #agents
• #developer-tools

Box: Gemini 4 Argon beats Gemini 3.1 Pro by 14 points on hard enterprise work

Box spent a week running Gemini 4 Argon on its Box AI Complex Work eval and saw a 14-point jump over Gemini 3.1 Pro. The biggest gains were in technology (+37) and consumer products (+33). The difference is follow-through: in one healthcare task both models spotted a dangerous potassium level, but only Argon removed the offending drugs from the final regimen, while 3.1 Pro at one point reported retention above 100% in a technology task. Levie separately argues that cyber will be one of the biggest spaces for AI agents, because agent swarms that review every line of code finally give defenders coverage that hiring never could.

• #models
• #evals
• #security

Every: Google is back, but Gemini 4 Argon scores 21% on a senior engineer task

Every published its Gemini 4 vibe check under the headline "Google is back." On Every's senior engineer benchmark task, though, Shipper says Argon really struggles: it scores 21%, far below Sol 6 and even Muse.

• #models
• #evals

David Silver bets superintelligence comes from experience, not human data

David Silver left DeepMind to co-found Ineffable Intelligence on the bet that the path to superintelligence doesn't run through LLMs. Instead, he wants agents that learn from scratch inside rich simulated worlds, which the team calls "crucibles," where simple goals give rise to complex skills. His proof point is AlphaZero: once his team dropped human data, it outperformed two years of human-in-the-loop AlphaGo work within hours. On why he didn't join the LLM crowd: "Am I going to be the 100,001 person that's pursuing that approach?" He predicts strong evidence within a year and true superintelligence in about three years, and he supports pacing frontier AI in the real world while moving as fast as possible in simulation.

• #research
• #reinforcement-learning
• #safety

Rauch: coding agents are AI's killer app, so stay model agnostic

Guillermo Rauch congratulated the Gemini team and warned never to bet against Google, given its Transformers pedigree, data, engineering talent and compute. His bigger claim is that everyone has rightly pivoted back to coding, and coding agents are the killer app of AI. Muse Wonder and Gemini Argon arriving back-to-back, he says, is a reminder that individuals and companies should stay model agnostic and lean on gateways.

• #models
• #agents

Latent Space tried to kill a $43,000-a-year SaaS in a weekend hackathon

Rather than debate the death of SaaS in theory, Latent Space took a real $43,000-a-year SaaS it uses and ran a global hackathon to clone it in a weekend, with any model and any harness allowed. It drew 72 submissions, burned roughly 100B tokens and was judged by both AI and human evals. The special episode shares the data, what's still hard about killing nontrivial mid-tier SaaS, and the winners of the $10,000 Kill My SaaS contest.

• #products
• #evals

Masad: pause frontier research today and AI still needs 10 years to diffuse

Amjad Masad argues that even if frontier research paused today, AI would take another 10 years to diffuse into the economy. Replit is working with solutions partners to speed that up.

• #enterprise
• #adoption

Google shows Gemini 4 Argon building a synthesizer from a 24-word prompt

Josh Woodward is showing off what the new Gemini 4 Argon can make. One demo is a custom synthesizer, built from a 24-word prompt, that turns galloping horses, rippling water and your own hands into music. Another is a 96-millimeter-wide custom furniture leg made of stacked spheres.

• #models
• #products

OpenAI trims ChatGPT and Codex desktop disk usage by 60%

New optimizations to how the ChatGPT and Codex desktop app stores state should cut what it keeps on disk by 60%. Sottiaux calls it a small ship, made because power users kept running out of disk space.

• #products