AI Builders
Digest

The Takeaway: Real automation does not come from models that chat better, but from reliable, composable, system-one intelligence that can live inside software.

Diogo Almeida is CEO of TypeSafe and creator of Jev. A former OpenAI researcher who worked on RLHF, he launched Jev a month ago to explosive traction, followed by an $870M Series A at a $7.5B valuation. Almeida argues that current LLMs were trained by RLHF to please humans in chat, producing sycophancy, over-refusal, stylistic repetition, and the jagged “95% Einstein, 5% Mr. Bean” behavior—models that are brilliant yet never signal when they are being foolish, making true automation impossible.

He advocates machine-native AI: intelligence shaped as classifiers and decision layers that plug into software like logic gates—composable, measurable, and threshold-tunable. Instead of giant system prompts that hope the model behaves, developers decompose behavior into verifiable smart if-statements and receive calibrated probabilities to set their own thresholds. Classification is not an outdated technique; it is the correct interface for putting intelligence into software branching. “Classification is AI-complete.”

Almeida’s slogan is “build prod, not god.” He calls pre-training a dumpster fire of wasted resources. TypeSafe starts from open-weight models, creates its own data and RLCD training in-house, and optimizes purely for machine-native reliability. Most copycats merely force LLM outputs into probability format without the hard robustness work. He expects a Jevons-paradox effect: once intelligence is cheap and trustworthy enough, it will be stacked like semiconductors across every layer of software.

“We're building prod, not god.”

TypeSafe emerged from two years of stealth with Jev, gained massive attention within a month, and closed an $870M Series A at a $7.5B valuation. CEO Diogo Almeida said the team deliberately chose under-promise, over-deliver because the AI field has long suffered from hype. The product itself carries the marketing—they have no marketing team and even rewrote their fundraising announcement to be direct and free of fluff. Almeida notes that much current usage is workflows migrating from LLMs that should never have been LLM tasks. He expects a distinct system-one market that will expand rapidly via Jevons paradox.

Box CEO Aaron Levie argues that most knowledge work is still rate-limited by human bandwidth for processing information. Chatbots were only the first step; the next is spinning up background agent swarms that “come back when the goal is done.” This shift will drive 100-1000x more token consumption and create large amounts of net-new work—codebase protection, exhaustive document risk review, infrastructure QA, and more. He also praised Meta’s new Muse open-weights model, saying enterprises will mix frontier models for orchestration with cheaper, tailored models for repeatable workflows.

Anthropic’s Boris Cherny (Claude team) noted significant progress on prompt-injection defenses for Claude models over the past few years. As agents take on more responsibility, solving the problem industry-wide has become urgent. He announced a partnership with Gray Swan to host a competition and invited the community to find issues.

SPC General Partner Aditya Agarwal called Goodfire potentially one of the most important companies to emerge from SPC. The team is demonstrating that interpretability and alignment can be approached as concrete engineering problems rather than pure research questions.

Every CEO Dan Shipper offered a sharp, satirical take: discovering a neuron that lights up for the Eiffel Tower does not tell you what the model is actually planning. Local interpretability wins are easily mistaken for solved alignment.

OpenAI Codex product lead Nan Yu argued that equating “taste” with aesthetics is the shallowest possible definition—and uninteresting precisely because it is so common. Real taste is a deeper form of judgment.

Vercel CEO Guillermo Rauch quipped that the company got a Vercel zeppelin and flying cars before GTA6, while also highlighting other impressive technical demos.

OpenAI’s Thibault Sottiaux (Codex & ChatGPT) posted: “Either we are out of ideas or we are cooking. Time will tell,” sparking widespread speculation.

Y Combinator President & CEO Garry Tan highlighted recent open-weight model progress with the simple observation “Open weight models rising.”

Builder Peter Steinberger (OpenClaw / OpenAI) called crabbox underrated and shared the relevant resource.