Search

Jev and the System One Pivot: Why an RLHF Co-Inventor Ditched Language Models

Diogo Almeida co-invented RLHF at OpenAI and then co-invented ChatGPT. On September 15, he launched Jev, a model from TypeSafe AI that emits no text at all. Structured state in, typed probability distributions out. No chat. No autoregressive decoding. No hallucination by construction, because the model cannot invent a string outside its schema.

Almeida is public about his pivot. He wrote on LinkedIn that RLHF "bakes in people-pleasing, overconfidence, and hallucination" and that "they've taken AI down the wrong path." His new company raised $40M from DCVC and exited stealth on September 15.

The strongest version of the thesis: RLHF optimized models for humans. Reasoning models made them better at thinking. Jev makes them good at software. This is a pivot in the interface between AI and its consumers, disguised as a training-method change, exactly like RLHF was in 2022.

What Jev Actually Does

Jev offers three primitives: Choice (a distribution over options), Score (a value on an ordered scale), and Noul (a yes/no probability). All questions in a request are evaluated in parallel, not autoregressively. The model is transformer-based but not an LLM per TechCrunch. Training used RLCD (Reinforcement Learning for Calibrated Decisions) on purely synthetic data.

TypeSafe claims 70–500 ms end-to-end latency versus 3–329 seconds for frontier LLMs. Input costs $0.042/MTok. Output is free. The homepage headline claims 193.6× faster and 444.6× cheaper. Those figures are self-tested, and TypeSafe notes these figures sit at the upper end of real-world gains.

The name comes from Kahneman's Thinking, Fast and Slow and from William Stanley Jevons, the economist who argued that cheaper coal would increase total consumption. Almeida wants cheaper intelligence to drive more demand. The Jevons paradox is real; whether machine-readable decisions have unsaturated demand is the open question.

Benchmarks: Honest Enough to Trust

TypeSafe's own four-workflow benchmark: Jev scores 67.8%, tied with GPT-5.6 Terra, a few points behind GPT-5.6 Sol (74.1%) and Claude Opus 5 (73.1%). At roughly 1/200th the cost and 1/50th the latency. TypeSafe disclosed its asterisks: the workflows were authored by its own capabilities team, and the reference answers are the average of GPT-6 Astra and Claude Fable 5.1 outputs, not ground truth.

Third-party evidence is thin but positive. Every published a test: 777 judgments across 37 documents in under 0.7 seconds for roughly a quarter of a cent. In a planted-defect writing check, Jev caught 6 of 7 defects vs. Claude Fable 5.1's 7 of 7, about 25× faster at roughly 1/580th the cost. Verdict: good but not perfect.

Vercel replaced a ChatGPT Luna 5.6 safety classifier with Jev and reported 5–18× faster with higher accuracy. Bryo AI's email classification vs. Gemini showed slightly lower accuracy but 10–20× cheaper, with confidence scores as the differentiator.

The Counter-Arguments Are Real

  • "Can't hallucinate" is not an empirical claim. Zero hallucinations means schema compliance. Jev cannot output an off-schema string, but it can still confidently pick the wrong allowed option. TypeSafe concedes the 0% figure is plotted by construction, not measured.
  • The technical moat may be thin. Sean Goedecke argues most of Jev's speed is reproducible on any open LLM: prefill the JSON schema, generate one constrained token with logits readout, batched. Roughly 2–3× speedups are already demonstrated with Qwen2.5-1.5B. He doubts a substantial technical moat exists and notes that calibration, not schema constraints, is the hard part.
  • Capability ceiling. No test-time compute likely caps System One models around non-reasoning LLM strength. This is a new interface, not a new scaling axis.
  • Calibration is not the same as probability. Returning a number for confidence does three different jobs: generated confidence text, raw logits, and empirically calibrated rates. Only the last justifies thresholded automation, and it must be validated per workflow. Zero-shot entailment classification is prior art from 2019.
  • Enterprise concerns. No explanations for decisions hurts auditability in regulated industries. Single-region hosted service. Early-stage vendor risk.

As of September 19: No Big-Lab Response

OpenAI, Anthropic, and Google had not responded as of September 19. No API changes, no competitive System One products, no public commentary. This absence of evidence was meaningful: if Jev were obviously replicable or obviously doomed, one of them would have said so.

Goedecke predicts choice-only versions of small fast models ("System One Terra / System One Haiku") if Jev gets traction. The bet is that calibration, not architecture, is the durable advantage. If he is right, Jev's lead is measured in months, not years.

Update, October 7: The Response Arrived in Two Weeks

The falsifiable test below resolved faster than anyone predicted. On September 29, OpenAI announced a Decisions API at DevDay: a bounded, classification-oriented endpoint built on GPT-6 Luna that returns probability-scored choices in about 150ms. It is in limited preview with no published pricing (Eagentix). On October 1, Cloudflare released open-weight Clef and Clef-flash under Apache 2.0, and AWS shipped Strands Decider 2B, a small decision model for hybrid agents where a decider picks and an LLM reasons (AI Eating the World).

So the category question is answered: decision-native endpoints are now a product category, not one vendor's framing. What is not answered is the calibration question. OpenAI's API has no publicly documented calibration track record, and independent tests put Jev mid-pack on accuracy: strong on binary screening, several points behind the frontier on harder calls. The remaining test is evidence, not endpoints. Whose confidence scores survive contact with production thresholds is what decides this category.

The Bottom Line

Jev is real, cheap, and fast. Its founder's biography gives the pivot story unusual weight. The model achieves comparable-but-not-frontier accuracy at drastically lower cost. The architecture is unreleased; the moat is unproven; the capability ceiling is real.

The "biggest pivot since RLHF" thesis is an argument, not a measurement. RLHF changed how models interface with humans. System One changes how they interface with software. Both are interface pivots disguised as training-method pivots. The falsifiable test was whether any major lab would ship a first-class, decision-native endpoint within two product cycles. It took two weeks.

Jev created a genuine new category: cheap, fast, non-language intelligence. The open question is whether it remains a product or becomes a paradigm. For developers building automation workflows, the answer is the same either way: agentic software just got a cheaper, faster option for the decisions that used to require chat.