Jev Launch Post

Text generation. Almeida says outright in a follow-up that the gains are not free because Jev cannot generate text, which is why the thread pairs each demo with ordinary code doing the work around the decision. The parallelism argument is the same one that took Transformers past RNNs.

Not from this post. The author says extraordinary claims need extraordinary evidence and points to the release blog for technical detail. The numbers come from workflow evals TypeSafe built itself, and the thread never names the comparison models, the workload, or the measurement method.

Input tokens, at $42 per billion, or $0.042 per million. The stated reason output is free is that the architecture makes it too cheap to meter. That is a vendor pricing policy rather than a measured rate, so do not build a long-run cost model on it.

A waitlist. The thread ends by pointing to an early-access signup and to the release blog for the technical writeup. The demos shown are a Doom run at roughly 10 calls per second, about $7 an hour, and a Wikipedia link race across hundreds to thousands of choices.

After Jevons paradox, per the author. Jevons observed that making a resource more efficient tends to increase total consumption of it rather than reduce it, which is the economic argument Almeida attaches to cheap decisions: if each decision costs close to nothing, far more of them can be made.

He describes himself as an AI capabilities researcher who co-created RLHF and ChatGPT at OpenAI and is now chief executive of TypeSafe. The launch is framed around that background and around RLCD, the training method he says he spent two years developing in stealth.

It is the training method Almeida says he spent two years building and positions alongside RLHF and RLVR. The post does not define it, and the technical detail it points to lives in the linked release blog rather than in the thread, so the launch post gives the label and the claim but not the mechanism.