Jev: The AI That Makes Decisions, Not Text - and Why It Changes Everything (Diogo Almeida)

Summary · 4 min read

Almeida is CEO of TypeSafe, the company behind Jev, and Jev’s creator. He co-invented RLHF at OpenAI.

Almeida argues Jev is a machine-native decision model that returns probabilities rather than generated prose, giving developers more control over software workflows. He says language models can sound polished while making unreliable decisions, and that RLHF and RLVR optimize for goals that do not solve dependable automation. TypeSafe’s RLCD and in-house data aim to improve decision reliability; Almeida says defining the task comes before scaling data and compute. He expects cheap decision models to create a market for intelligence embedded in software, doubts recursive self-improvement, and thinks practical automation may avert the AI winter he fears.

Minute by minute

1. 1:13

   Intelligence per dollar drives adoption. Matt Turck cites TypeSafe’s $870 million Series A at a $7.5 billion valuation and Jev’s rapid growth. Almeida says the metrics matter less than builders using Jev for applications that were previously too expensive. He argues that intelligence per dollar, rather than intelligence per second, will drive adoption, and says its cost has fallen by two orders of magnitude.

2. 3:21

   Product proof mattered more than hype. Almeida says TypeSafe stayed in stealth because he saw AI as overpromising and underdelivering; Jev’s launch gained traction when users tried it and shared results. He says a technological gap, not hype, kept him committed, and he still believes AGI could have arrived last year under OpenAI’s definition. He argues the product carries the marketing and that investors only understood his pitch after Jev showed product-market fit.

3. 9:27

   System One fits software workflows. Almeida prefers to describe Jev as machine-native AI, with System One as an implementation detail. He says AI is strong at fast, intuitive System One tasks, while models’ System Two reasoning remains fragile. He argues that reliable small units can be composed into more capable software workflows.

4. 15:17

   Email decisions need legible rules. Almeida uses email scheduling to contrast LLMs’ large natural-language prompts with Jev-based classifiers for individual decisions, such as whether a message is inbound or urgent. He says prompts can become difficult to control as complexity grows, while classifiers make each rule visible and measurable. Developers can then test errors and add them to evaluations.

5. 19:57

   Probabilities give developers control. Almeida says Jev returns probabilities so developers can set thresholds for actions, rather than forcing a yes-or-no decision. He points to inconsistent refusal behavior in language models and says different applications, such as Disney and Character AI, need different thresholds. Developers can tune for the errors they most want to avoid.

6. 23:38

   Classification is a broad intelligence interface. Almeida attributes Jev’s speed and lower cost partly to avoiding slow, autoregressive text generation, but says the model’s training and optimization matter too. He rejects the idea that Jev is narrow, arguing that a general-purpose classifier can support a wide range of tasks and calling classification AI-complete.

7. 27:12

   RLHF optimized models for conversation. Almeida describes RLHF as making models conversational and pleasing to humans. He says optimizing for human preferences encouraged stylistic, agreeable answers, contributing to sycophancy, repetitive outputs and refusals. He argues that this can make models seem more capable than their decision-making reliability warrants.

8. 34:18

   RLVR gains do not ensure automation. Almeida says RLVR rewards models for verifiable results, such as passing unit tests or solving math problems, and can produce strong benchmark results. But he argues that benchmark gains have not made automation reliable: models can be highly capable on some tasks and still fail unpredictably. He describes that unevenness as a barrier to trusting them in software.

9. 37:41

   RLCD targets calibrated decisions. Almeida says TypeSafe developed RLCD to optimize for machine-native intelligence, while stressing that the training method is only one part of the work. He says the team had to define the task, its inputs and outputs, and the data needed to train for it. TypeSafe built that data in-house because it did not already exist in the right form.

10. 41:40

    Define the task before scaling. Almeida argues that the Bitter Lesson oversimplifies the role of compute: real-world scaling also requires data, and data depends on first defining a valuable task. He says TypeSafe uses a combination of open-weight pretrained models and did not do pretraining. He calls pretraining a waste of resources and says he hopes the company can avoid doing it.

11. 45:47

    Jev puts intelligence inside software. Almeida says calling Jev a classifier is a compliment: classifiers are an interface for putting intelligence into software and guiding code’s branching behavior. He argues that Jev can make this kind of intelligence programmable across many tasks. Rather than act like a software engineer that writes code, he wants Jev to function like a logic gate inside software.

12. 53:02

    Copying the interface misses reliability. Almeida says reproducing Jev’s interface is easy, but building reliable behavior took years; he wants to measure how well competing models perform. He argues that benchmarks can mislead users who lack time to test models themselves. TypeSafe has not committed to open weights; he says an API enables faster iteration, though smaller and on-device models could change the decision.

13. 59:59

    Cheap intelligence could expand demand. Almeida sees decision models and language models as distinct markets, even though some users are switching from LLMs for workflow automation. He says cheaper, reliable intelligence embedded in software could create new uses and drive more demand, as applications call decision models repeatedly within larger workflows.

14. 1:02:43

    Almeida doubts recursive self-improvement. Almeida says past small efficiency gains have often led people to overstate the prospect of recursive self-improvement, while diminishing returns limit what those gains can deliver. He thinks AI may still face a market correction, but says the automation failure he feared is less likely now that reliable decision models are emerging. He argues that practical automation could avert an AI winter.

Published Oct 15, 2026.