I built non-autoregressive decision models with RL a year ago

Both, loosely. The submitted post is from the open-model publisher arguing it built non-autoregressive RL decision models a year earlier, and the entry notes the story URL points at that publisher rather than TypeSafe. The real subject is who arrived first.

The argument is more conceptual than technical: commenters say constrained choice over predefined options is an established pattern and that naming and packaging are what make it look novel. The entry's own caveat applies — these are commenters' readings of priority, not adjudicated findings.

One commenter walks through the earlier project's training script arguing target leakage, tracing an outcome field from the metrics object into the conversation state vector and back in as model input. That is an unverified accusation made in comments, not a reviewed result.

Several commenters say no, describing it as relying on retrieval and an orchestrator so the title oversells the method. One also claims the project was later reshaped into the open-weight model after Jev shipped, which is likewise an unverified comment.

The captured page shows 1,363 points and 319 comments. That is the reason the thread is worth reading: the volume of interpretation is where the substance is, since the entry's own caveat is that point totals move and that priority claims here are commenters' readings rather than adjudicated findings.

Several do. One argues marketing and branding are part of what makes a product a product and that Jev is unusually well branded and advertised as something more general. Another agrees it is easier to market with a name brand behind you, comparing it to the same observation on Twitter from someone known versus no one. A third frames it as a familiar ML pattern of parallel rediscovery and sour grapes.

It cuts both ways in the thread. One commenter describes the pattern as routine in ML, where the same intuition gets reached by many people under different names. A reply says that comparison is obnoxious specifically because Schmidhuber rarely executed the idea correctly, which makes priority claims hard to take seriously when the execution is thin.