ChatGPT Co-Inventor Launches Model That Makes LLMs Look Like Overkill for Decisions

What actually dropped today

Diogo Almeida , the OpenAI researcher who co-invented ChatGPT and RLHF , just launched TypeSafe AI and its first model, Jev. The headline calls it a new class of decision-making model. What that means in practice: Jev doesn't generate text. It outputs type-safe probabilities trained via Reinforcement Learning for Calibrated Decisions (RLCD). No tokens, no sampling, no temperature parameter. You feed it structured input, it returns a probability distribution over your defined schema. That's it. The blog post went live this morning alongside TechCrunch and InfoWorld coverage. Vercel is already using it in production for safety classification.

Why this breaks the LLM paradigm

Every LLM you've used is a generative model pretending to be a classifier. You prompt it, hope the output parses, then regex the JSON out. Jev skips the middleman. RLCD trains directly on the decision boundary , not next-token prediction. The model learns to output calibrated probabilities that map to your actual labels. This isn't fine-tuning a chat model. It's a different objective function entirely. The architecture is smaller, deterministic, and doesn't need a tokenizer. When Almeida says it can't hallucinate, he's not making a safety claim. He's describing a structural property: there's no generation step to go off the rails.

The hallucination claim is structurally real

We've heard 'reduced hallucinations' from every lab since 2022. Usually it means better RLHF or a classifier head on top. Jev is different. A model that outputs a probability distribution over a fixed schema cannot invent facts , it can only assign weight to the options you gave it. If your schema has three labels, the output space is three numbers summing to one. There's no vocabulary to drift into. No chain-of-thought to derail. The failure mode shifts from 'making things up' to 'miscalibrated confidence on known categories.' That's a debugging problem we actually know how to solve. This is the first time the architecture itself eliminates the failure class.

Numbers that change what you can build

TypeSafe claims 40-400x cheaper and 40-200x faster than LLMs for decision tasks. Let that sink in. A classification that costs $0.01 with GPT-4o drops to $0.00025. Latency goes from 800ms to 4-20ms. At that price-performance, you can run Jev on every log line, every API request, every database write , not just the expensive human-facing ones. The Vercel deployment validates this: they're using it for safety classification at 5-18x speedup over OpenAI's own Luna model. Luna was built for exactly this workload. A two-person team's first model beats OpenAI's dedicated safety classifier on their own turf. That's not incremental.

Almeida just obsoleted his own previous work

This is the part nobody's emphasizing. Diogo co-invented RLHF , the technique that made ChatGPT usable and spawned the entire alignment industry. RLHF teaches models to generate text humans prefer. RLCD teaches models to make decisions machines can trust. They're solving different problems. For the last two years, we've been jamming LLMs into classification pipelines because they were the only tool that understood nuance. Almeida just built the tool that actually belongs there. The irony is delicious: the person who gave us the generative paradigm just shipped the discriminative alternative that makes generative models look like the wrong primitive for half the use cases we've been forcing them into.

Vercel in production means this isn't a demo

TechCrunch says Vercel uses Jev for safety classification. InfoWorld confirms it's handling real traffic. This matters because safety classification is the canary in the coal mine , high volume, low latency budget, zero tolerance for false negatives. If Jev survives there, it survives anywhere. Luna was OpenAI's answer to this exact problem. Jev beating it 5-18x on speed while being 40x cheaper suggests the architecture gap is real, not benchmark-chasing. We've seen too many 'faster than GPT' claims that evaporate under load. A production deployment at Vercel's scale with those numbers is the strongest signal we've seen since the original ChatGPT launch.

This kills the LLM-as-classifier pattern

Right now, every team building AI features does the same dance: call GPT-4o with a structured prompt, parse JSON, retry on failure, cache aggressively, pray the schema holds. It works but it's absurd , burning a 1.8 trillion parameter model to pick between 'spam' and 'not spam.' Jev makes that pattern obsolete for any decision with a fixed label set. Content moderation, routing, feature flags, anomaly detection, fraud scoring , all of it moves to a model that costs pennies per million decisions and returns typed probabilities in milliseconds. The second-order effect: your LLM budget shifts entirely to actual generation tasks. The classifier tax disappears.

What this signals for the next six months

TypeSafe calls these 'System One Models' , fast, intuitive, machine-readable. Kahneman's System 1 vs System 2 mapped to AI. The implication: we're about to see a split in the model market. Generative models (System 2) for reasoning, writing, coding. Decision models (System 1) for classification, routing, gating, scoring. They'll compose: Jev routes the request, the LLM handles the complex part. Almeida's team is small and just launched. Expect copies within weeks , but the RLCD training data and calibration curves are the moat. If you're building AI infrastructure, stop optimizing LLM classifiers. Start designing for a decision layer that's deterministic, typed, and cheap enough to run everywhere.