By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Latent SpaceWhy I couldn't build Jev at OpenAI — Diogo Almeida, TypeSafe Co-founder & CEO
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Jev’s bet: programmable System One models built for reliable software automation
- Diogo Almeida says Jev represents a new class of “System One” programmable models built for software consumption and optimized for intelligence per dollar.
- He argues common post-training approaches (especially RLHF) create mode collapse and miscalibration that undermine reliability, which he views as the main blocker to AI automating economically valuable work.
- He claims refusals and “safety alignment” policies that follow a lab’s instructions are acceptable for consumer chat products but are unacceptable for APIs embedded as dependencies in software systems.
- He explains why TypeSafe rejects public benchmarks in favor of workflow-specific evaluation, internal eval discipline, and a long-term focus on moving “nines” of reliability.
- He outlines how Jev’s API primitives (choice, score, null) map to programming constructs and why developers should structure inputs/outputs and decompose tasks into many small calls.
IDEAS WORTH REMEMBERING
5 ideasJev is positioned as “machine-native” AI designed for code to consume.
He frames Jev as a “large programmable” System One model meant to be consumed by code rather than by humans chatting, and says TypeSafe is optimizing the model and its interfaces for software integration.
RLHF-style tuning can warp model behavior in ways that break software reliability.
He argues RLHF pushes models toward mode dropping/mode collapse and harms calibration, creating behavior that looks good in chat but is unreliable for decision-making inside programs.
Refusals are treated as fundamentally incompatible with being a reliable API component.
He says refusals are acceptable-ish in a chatbot but “a type error” in an API dependency because they can cause stochastic breakage deep inside downstream systems where users may not even know what model is involved.
They reject public benchmarks and prioritize real workflow evaluation and trust.
He says TypeSafe is “extremely anti-public benchmarks,” calling them gameable and misaligned with measuring the “je ne sais quoi” of intelligence; he emphasizes workflow-specific evaluation and internal truth-seeking eval discipline.
TypeSafe centers data and task design, not user-data training, to achieve robustness.
He describes TypeSafe as a “data lab,” emphasizing that capability comes primarily from data and picking the right “task/North Star,” and says their data is synthetic and intentionally not trained on users’ data to avoid bias and overfitting to the present.
WORDS WORTH SAVING
5 quotesLike, I feel like the entire AI field is like one of those, like, carnival house of mirrors, and everyone is just insane and saying the weirdest stuff that doesn't make sense.
— Diogo Almeida
Refusal is just like obviously a type error.
— Diogo Almeida
Like would I prefer that our stuff is not used to kill people? Obviously. Would I prefer it's used for like all sorts of like great stuff in the world? Obviously. Will I put my thumb in the scale for that? Yes. Would, will I do it at the technological layer? Absolutely not, because that will fracture the intelligence.
— Diogo Almeida
I believe that in the long run, it needs to be vibes and trust until you put it into a workflow and evaluate it for that workflow and measure it and have your own sense of, like, how it does on the exact workflow that matters.
— Diogo Almeida
If you gave me a billion dollars, I wouldn't pre-train.
— Diogo Almeida
High quality AI-generated summary created from speaker-labeled transcript.
