AI Builders Digest

Altman: pacing means slower than possible, not stopping

Two ways AI progress goes very badly: losing control of the future to AI, and power concentrating so much that one person, company, or country imposes its worldview on everyone. OpenAI now writes explicit safety cases before frontier reinforcement learning runs it expects to jump capability, a shift from the old Preparedness Framework posture that focused on finished models at deployment time. He welcomes a federal framework with consistent safety requirements and says no antitrust exemption or legislation is needed before labs start building that confidence themselves. The blunt line: no amount of American competitive pressure justifies recklessness, and capabilities must not get ahead of alignment and monitoring.

• #policy
• #safety

Levie: "pacing" sounds like a handbrake, but the goals are table stakes

The word itself is a trigger because it reads as an arbitrary slowdown or a way to hobble competitors through regulation. The actual improvement goals in Dario's essay are absolute necessities, and equivalents already exist in aerospace, life sciences, and health care. AI is heading into financial trading systems, medical devices, biotech, defense, and government workflows, so wanting those systems safe and aligned is not a controversial ask. The hard part, and he calls it one of the most complex questions of the century, is getting there without meaningfully slowing innovation or reducing competition.

• #policy
• #safety

Arm's Rene Haas: compute demand isn't close to oversupplied

Chip design runs 24 to 36 months, and the bulk of that is not architecture but verification, validation, debug, and documentation, which is exactly what AI is good at. Between 80 and 90% of Arm engineers use it daily, and taking it away would be like restricting internet access to two hours a day: "the genie's out of the bottle." On the bubble question, valuations aside, supply versus demand is "not even close," and the next bottleneck is data center construction rather than wafers or memory, partly because of organized local pushback rooted in job-loss fear he says is not well grounded. He also pushes back on accelerator tunnel vision: something has to orchestrate where all those tokens go, and that is what CPUs do.

• #hardware
• #compute

A higher entry valuation and higher comp is usually a bug, not a feature

Founders text him about this daily: a candidate treats a big last round as proof a company is safer and better to join. Raising high does not make you more secure, and a company worth 100x ARR has to grow into that number. Do the work yourself on market, traction, positioning, your own exit estimate, and your 409a price and tax exposure, because right now the numbers are not rooted in reality. He has lived it: a company he worked at raised $60M from one of the best investors alive and folded three years in with all equity wiped out.

• #startups
• #funding

0.04% of the world uses agents effectively, per Brex CEO's chart

Pedro at Brex uses a chart where each dot is 3.2 million people. Gray boxes are the 84% who have never used AI at all, green is the 16% on free chatbots, orange is the roughly 0.3% paying $20 a month, and one tiny red box is the roughly 0.04% using agents effectively. The data is from February 2026 so the numbers have likely moved, but the shape of it holds: while everyone argues about AI outcomes for humanity, most of humanity has not started.

• #adoption
• #agents

Worktrees get ~80% faster using filesystem folder clones

The next release, or the dev channel, makes worktree creation roughly 80% faster by using copy-on-write folder clones on APFS, btrfs, XFS, and ReFS, which also saves a lot of disk space. It is written in Rust, because you cannot escape Rust. He plans to test it for a few weeks and, if it helps most users, try to get it into Codex.

• #products
• #developer-tools

Astra agrees with your correction and then does nothing

A specific failure pattern: the model reports it did X, you say that is wrong and it should do Y, it replies "you're right, I should do Y," and then stops there. Other models just go do Y. Agreement without follow-through is its own kind of broken, and it is the sort of thing that only shows up in daily use rather than benchmarks.

• #agents
• #evals

Masad: users were priced out of AI coding, now it's free to build again

He calls the stretch where costs pushed people out of AI coding painful to watch. That window has closed on his end, and building is free again.

• #products
• #pricing

Fable cracked the Cyphral Distich, a 370 year old cipher

A 370 year old cipher fell to Fable, which Cherny flags as a genuinely novel use of Claude rather than another coding demo.

• #research
• #products