AI Builders
Digest

Anthropic Claude Code lead Boris Cherny shared an experiment where Claude handles day-to-day maintenance of their apps. In a Slack channel called proj-claude-maintains-apps, Claude runs daily routines including crash fuzzing in simulators, unifying near-duplicate abstractions, removing dead code, and fixing leaky abstractions across iOS, Android, Desktop, web, CLI, and Agent SDK. Over recent weeks it opened 388 PRs, 180 of which were merged after Claude Code Review plus human review. Claude usually gets them right on the first try; otherwise the team asks it to tune its own routines. They are now looking at ways to streamline merging these mechanical changes.

Replit CEO Amjad Masad predicted that by next year using a computer will be optional and work will radically change. He noted that ARC-AGI-3 is nearly solved simply by adding a coding harness, confirming that coding generalizes LLMs. He added that this is already true inside Replit.

Box CEO Aaron Levie called the hypothesis that AI would eliminate engineers absurdly wrong. AI gave engineers a power tool that accelerates development of whatever we want. Automating drug discovery, manufacturing, or larger software projects still requires engineers. AI causes companies to take on far more work, creating demand for more experts to oversee it. Even as models improve, experts utilize them better than novices. Great time to be an expert.

Meta AI Senior Director Madhu Guru argued that prompt debt is the new tech debt. With every model update you should cut at least 50% of your prompts. The typical pattern is model fails then add 10 rules, tool call fails then add 10 examples, weird output then add formatting constraints. Three months later the system prompt is a novel while the model has gotten smarter and the micromanaging prompts turn it into a dumb rules machine.

The Takeaway: Humans still want to compete and create with other humans; even after three decades of superhuman chess engines, more people play chess than ever.

Erik Allebest is CEO of Chess.com. In 2005 he and a friend bought the chess.com domain at a bankruptcy auction for roughly $56,000 after every VC told him it was an uninvestable niche. They raised zero capital, bootstrapped with proceeds from earlier exits and a small loan from a friend’s mother, launched in 2007, turned profitable within 18 months by charging for lessons, and never looked at fundraising again. Today the company has more than 250 million registered members, roughly 10 million daily actives, over $200 million in annual revenue, and a 650-person fully remote team whose mission is to get a billion people playing chess.

After the COVID and Queen’s Gambit spike they feared a flash in the pan; the 2023 wave driven by short-form video, the Mittens bot and cheating scandals then lifted the baseline permanently. Only in 2024 did they finally accept that chess is a large and durable business. Allebest notes: “We always thought we’d never have more than 50 or 100 people, but chess kept surprising us.”

On AI he observes that early engines made chess boring by pushing perfect but dry play, while neural nets such as Leela Chess Zero restored creativity and aggression. AI coaches, personalized puzzles and game review now help humans improve faster. There are still no shortcuts: “Take any given day and multiply it by a thousand—that’s what your life looks like.” The company is applying the same rating philosophy to poker via Gambit so skill, not stack size, becomes the signal.

His advice to founders is the opposite of the standard playbook: ignore most founder advice, decide what you want to exist, prove it with minimal resources, work with great people, and just execute.

“Stop listening to people just giving you advice on what to do. Just go do it.”

AI engineer Swyx (smol.ai, Latent Space) modified /align-me after listening to Matt Pocock and Thariq so it accepts batches of questions instead of round-by-round interaction. The intuition mirrors speculative decoding: look ahead 2-10 steps to speed things up. He reports it works incredibly well for design explorations. He also highlighted the latest frontier-skills posts from Matt, Thariq and Philipp Schmid.

OpenAI Codex & ChatGPT team member Thibault Sottiaux showed that you can now work with Google Docs, Sheets and Slides directly inside ChatGPT. He says it has changed how he writes, brainstorms and proofreads—open the file and chat or talk through changes without leaving the flow. He also recommends installing the Computer History plugin and asking ChatGPT to roast your computer usage after a day; the results are insightful and often painfully accurate.

Vercel CEO Guillermo Rauch predicted that a single command managing uptime, model choice, lower costs, observability and ZDR will become the default way to use coding AI at scale. The command already configures existing harnesses such as Claude Code and Codex. He also highlighted free GLM 5.2 (up to 500 TPS) from Blackbox AI.

Google Labs / Gemini VP Josh Woodward announced 3.7 Flash: faster, 50% cheaper, and shipped in roughly three weeks.

Meta AI Senior Director Madhu Guru noted the irony that the AI industry claims unlimited creativity while naming almost every product “Studio.” He counted more than 20, including Google AI Studio, Vertex AI Studio, Copilot Studio, Agentforce Studio, SageMaker Studio and LangGraph Studio. Only a few actually make studio-like tools.

Builder Zara Zhang observed that many expected AI coding to devalue engineers, yet the most in-demand roles all contain the word “engineer”: forward-deployed engineer, design engineer, product engineer, growth engineer.

FirstMark Capital’s Matt Turck described the current startup climate: either you are an AI-native rocketship (permanent fundraising at ever-scarier valuations, death-match for talent and customers, sacrificing gross margins) or you are left for dead regardless of how good the underlying company is. The binary has existed for a while but recently feels escalated.

Several builders (Ryo Lu, Nikunj Kothari, Peter Yang) praised the Matic home robot as the first they own that genuinely feels smarter with every update. It has become a “fifth family member” and keeps floors spotless—sometimes too diligently. The company is described as a true ten-year overnight success.

Sam Altman and Thibault Sottiaux both highlighted /ultrafast. Cat Wu invited non-engineering roles to book Cowork office hours. Peter Yang found AI most useful during a family health situation for navigating healthcare bureaucracy rather than researching the illness itself. Nikunj Kothari debated one master agent versus specialized sub-agents, guessing the long-term winner will be a master orchestrator. Garry Tan declared YC is becoming the YC for hard tech.