AI Builders
Digest

The Takeaway: True automation inside software requires machine-native AI optimized for reliable, composable System 1 decisions — not chat models trained by RLHF to please humans.

Diogo Almeida, CEO of TypeSafe and co-inventor of RLHF at OpenAI, spent two years in stealth building JEV. It is not another chatbot. It turns intelligence into software logic gates: calibrated probabilities instead of brittle, sycophantic text. Developers decompose decisions the way they write code — “is this urgent?”, “should we escalate?” — and every classifier is measurable, debuggable, and composable.

His sharpest line: “95% Einstein, 5% Mr. Bean. And if it doesn’t tell you when it’s the Mr. Bean, then how are you gonna automate anything?” Current LLMs remain jagged on System 2 tasks; System 1 intelligence is already extremely valuable if made robust. By discarding autoregressive generation and inventing RLCD training for machine-native intelligence, JEV is orders of magnitude faster and cheaper while staying reliable.

“We’re building prod, not god.” Copycats only copied the interface; they skipped the hard work of reliability. Benchmarks, in his view, are lies that mislead the users who need truth most. The real revolution is making intelligence as composable and infinitely replicable as software itself — so every row in a database or line of code can embed trustworthy decisions. The $870 M Series A at a $7.5 B valuation is noise; what matters is builders finally shipping things they previously avoided because AI was too expensive and too unreliable.

Vercel CEO Guillermo Rauch notes that models increasingly speak engineering vernacular — gate, evidence, culprit, durable, provenance, contract — language born from debugging and forensic analysis of existing systems. He wonders what the true minimum viable English is for productivity; ASD-STE100’s ~900 approved words may already be more than we need.

The deeper point: there will be times you over-guide the AI (it knows better) and times you under-guide it (it drifts or misses context). How well you manage this superpower is the #0 skill he looks for.

Box CEO Aaron Levie predicts we will soon see 1,000× more tokens flowing through cyber agents. The surface area of risk keeps expanding; most companies already cannot keep up. Only in the last few months have long-running agents and computer-use capabilities become good enough, and models trained deeply on cyber, for agent swarms to fan out, leverage any security tool, test attack permutations, read logs, and score enterprise data risk.

This is a massive opportunity across the entire security stack — identity, firewall, network, code scanning, data protection — and one reason to be optimistic about the balance of offensive and defensive capability.

After two years in stealth, Diogo Almeida’s TypeSafe launched JEV and within three weeks became the most-copied model in AI, raising an $870 M Series A at a $7.5 B valuation. Roughly a million users a day, a trillion tokens a day, and one-third of the Fortune 500 already on board. Almeida insists the real leading indicator is builders using it for things they previously avoided because AI was too expensive. The product itself is carrying the marketing; the company still has no formal marketing team.

Ryo Lu (previously design at Cursor, Notion, Stripe) wrote a reflective piece on what is lost when difficulty disappears. He learned programming and design by failing, taking things apart, and starting over. Struggle taught him to feel when something was wrong before he could explain it.

He worries that agents make it too easy to avoid the discomfort of collaborating with other humans — the pushback, the different priorities, the slow process of finding rhythm together. Some of his closest friendships came from sharing unfinished work and being vulnerable. If everyone retreats into private agent worlds, we may lose the very friction that turns strangers into a band.

Product person Peter Yang reports that of all the learning apps he and his 4-year-old have tried, only Once has lasted months. It builds a daily shared habit, adds lightweight gamification (silly photos after lessons), and after 30 lessons ships physical books the child can read alone. His daughter now sometimes chooses Once books over YouTube.

Swyx is hiring writers, reviewers and reporters for the new Latent Space editorial team and welcomes part-time and one-off guest posts.

Aditya Agarwal (GP at SPC) notes a side effect of being max AI-pilled: he now treasures IRL time with family and friends more than ever. Let the agents cook; celebrate humans.

Thibault Sottiaux (Codex & ChatGPT at OpenAI) called a rest day for Codex after a heavy stretch, promising “mad codexing” ahead.