|
AI Builders Digest
Tuesday, September 15, 2026
|
|
The AI safety debate keeps pulling in more voices, and today Box CEO Aaron Levie added a useful one: someone who thinks "pacing" is a reasonable word, actually. That puts him in an interesting spot relative to last week's Rauch-versus-Masad split, and it's worth watching which framing wins the broader industry conversation.
|
|
---
|
|
01
|
84% of the world has never typed a prompt. Here's what that chart looks like.
|
|
|
Product creator Peter Yang surfaced a slide from Brex CEO Pedro Franceschi that reframes the entire AI adoption debate with one image. Every dot is 3.2 million people. Gray boxes (84%) have never used AI at all. Green boxes (16%) use free chatbots. Orange boxes (roughly 0.3%) pay $20 a month for a subscription. And the red box, the one everyone in this industry obsesses over, represents the 0.04% of people using agents effectively.
|
Why it matters: If your product roadmap assumes the market is saturated with sophisticated AI users, that chart is a reality check. The competition for who builds the best agent isn't the main event yet. The main event is getting the gray boxes to turn green. Companies that figure out onboarding for genuinely AI-naive users will have more leverage than whoever ships the cleverest agent workflow.
|
|
Source →
|
|
---
|
|
02
|
Box CEO Aaron Levie: "pacing" isn't a dirty word
|
|
|
Aaron Levie pushed back on the idea that "pacing" in AI development is just regulatory capture dressed up in safety language. His argument: Anthropic CEO Dario Amodei's specific improvement goals aren't arbitrary slowdowns, they're the kind of standards we already demand in aerospace, healthcare, and financial systems. AI is going to run those same systems, so why would we hold it to a lower bar?
|
Why it matters: This is a more politically difficult position to hold than either Rauch's (move fast) or Masad's (we've already been hacked). Levie is saying: the people asking for standards aren't your enemies, and dismissing them as regulators-in-disguise loses the argument before it starts. If you're building in fintech or health AI specifically, how this debate lands will determine which compliance frameworks you have to design around in the next 18 months.
|
|
Source →
|
|
---
|
|
03
|
Claude cracked a 370-year-old cipher
|
|
|
Boris Cherny flagged that a company called Fable used Claude to solve the Cyphral Distich, a cipher that had been unsolved for 370 years. Nearly 2,000 likes suggests this landed well beyond the usual AI-twitter crowd.
|
Why it matters: Historical cryptography is a niche use case, but the underlying capability isn't. If Claude can pattern-match across a 370-year-old puzzle with no known solution key, the same reasoning applies to contract language, regulatory filings, and legacy codebases that nobody on your team fully understands anymore.
|
|
Source →
|
|
---
|
|
04
|
Google DevFest 2026 has 800+ global events
|
|
|
Google DevFest is back, with more than 800 events worldwide organized around building, securing, and scaling in what Google is calling "the agentic AI era." Sign up to find one near you.
|
|
Source →
|
|
---
|
|
05
|
Replit CEO Amjad Masad posted "Nominal determinism strikes again" with a link
|
|
|
The content of the linked resource isn't available here, so there's nothing to report on substance. The joke lands if you know what he's pointing at. Worth clicking directly if you follow Masad closely.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|