AI Builders Digest

Claude now opens and co-edits Google Docs, Sheets and Slides beside the chat

Paste a Google file link or ask for a new doc, sheet or deck, and it opens next to the conversation so you and Claude can edit it together. Access follows your existing Google sharing permissions, and it's in beta on all paid plans. Anthropic also spotlighted Every, which built a company agent on Claude Managed Agents that the whole team uses in Slack to share skills when a new model comes out, then released it to subscribers once it caught on internally.

• #products
• #agents

Applied Compute CEO: post-training wins inference, and your evals are the moat

Applied Compute's CEO, who mostly worked on Codex while at OpenAI, argues that the hard part of RL is defining what to optimize, so companies should guard their evals as closely as their employees instead of handing them to labs or public benchmarks. The core thesis is that post-training wins inference: the biggest workloads get custom-trained first, and you can cut a token bill by making a model more token-efficient at the same eval score, not just by tuning serving. Rubric-based RL works surprisingly well in domains without one right answer, but RL transfers far less than pretraining, and continual learning is still blocked on learning efficiently from sparse rewards. "What we have with RL is we essentially have a hill climbing machine. The hardest part is actually defining the hill to climb."

• #post-training
• #evals
• #inference

Boris Cherny: there's no secret to prompting, talk to Claude like a coworker

Boris Cherny, who works on Claude Code at Anthropic, shared real prompts and says most tasks don't need heavy scaffolding or prescriptive instructions: give Claude a goal and it will figure it out. Back in the Sonnet 3.5 days, the wording of a prompt mattered a lot. Now the three things worth communicating are what you want done, how much effort to spend, and how the model should verify it did the right thing.

• #prompting
• #coding

Thariq: Claude's brains are moving to the cloud, with local hands on your computer

Thariq, who works on Claude Code at Anthropic, says things are increasingly moving toward Claude's 'brains' living in the cloud while 'local hands' operate on your computer. The tricky part is availability: if Claude can only reach your files while your computer is online, work can stall until it comes back, so some kind of sync may be needed, and sync has its own edge cases. Thariq also pushed back on the idea that coding agents let you skip the lower layers, since working at a higher level of abstraction has always required understanding the ones below.

• #agents
• #coding

Aaron Levie: cyber will be one of AI's most defining domains

Levie expects AI to pile new work onto security teams through vibe-coded bugs, agentic attacks and even accidental agent swarms hunting for data, and calls OpenAI + Hugging Face just a preview. Security teams are often the most resource-strapped in the enterprise, so agents become the fix as well, with new agentic products for protecting code, enterprise systems, critical infrastructure and data. The bottom line: a booming market for security professionals who can actually deploy agents.

• #security
• #agents

Amjad Masad: AI decompilation will make all software de facto open source

Masad calls what's happening in AI-powered reverse engineering and decompilation 'absolutely insane' and predicts that pretty soon all software will be de facto open source. The blunt summary: 'AI is coming for everything and everyone.'

• #security
• #open-source

Peter Steinberger hooked the team's claw agent up to X to dispatch work

Anyone can grab an unassigned session, and the agent looks up who last worked on the related code and pings them on the team server. The whole setup took one prompt, and the team server extended itself because plugins are now hot-reloadable.

• #agents
• #coding

OpenAI's Thibault Sottiaux processes a community-voted reset and keeps the upgrades

Sottiaux, who works on Codex and ChatGPT at OpenAI, says Day 2 shipped four things judged good to great plus some math proofs, but the community vote still came out for a reset, which has now been processed. Sottiaux admits the game seems rigged in reset's favor under the current rules, and says the improvements won't be unshipped. Day 3 is tomorrow.

• #products

Nikunj Kothari: VC rage-bait on X is costing other people real deals

Kothari says too many VCs chase dopamine or deliberately rage-bait for views, and since X has no room for nuance, nothing said there can really be taken back. The damage lands hardest on the people 'under' them: Kothari knows two people who lost deals that were all set because of X drama. With capital a commodity and so many venture funds around, few seem to be playing long-term games.

• #funding

Guillermo Rauch bets on confidence-threshold thinking for AI decisions at scale

Rauch highlighted a feature he sums up as 'thinking fast and a bit less fast under a confidence threshold', calling it very simple but likely to be extremely impactful for AI decision-making at scale.

• #products