Envisioning high-frequency decisions in games and trading with Jev
Recalling that a year ago small Qwen models playing Street Fighter were limited by prefill, the author thinks Jev makes LAM-style high-frequency decision making plausible for gaming, trading and robotics.
A year ago, when I tried using small Qwen models to play games like Street Fighter, I realized that efficiency was a major bottleneck. Most interactions were actually prefill-bound: you don’t need much CoT or long decoding, but you do need fast decisions. Now Jev makes me think this direction has real potential. LLMs, or what I personally think of as LAMs (Language Action Models), could eventually power much higher-frequency decision making in gaming, trading, and even robotics. Here is an example of Qwen model one year ago.
This post quotes another post: “ChatGPT co-inventor releases Jev, a model built for decisions rather than text”