There is a new approach called RLCD which is reportedly used by typesafe AI team. But it appears that it is plagiarized and build on someone elses work without being open source/open weight claiming as if it there own. reddit.com/r/LocalLLaMA/…/i_literally_built_the_jev_architecture_one_year?solution=…&jsc_token=…&… typesafe.ai I haven’t read the papers associated with the thread, just found it and am wondering if it is so. The painful part is i know a few people in that team. I think everyone is following hype cycle and fund_raising and chasing validation and claiming to be the good guys/philanthropic.

Flyte 2 now has support for “System one types” union.ai/docs/…/system-one-types, just includes Jev, but I see Laya and a bunch of other OSS alternatives have popped up

I tried both of it. To put it bluntly i am not very impressed. I still have to read the paper to understand what it does, though from a superficial read it appears that the “head” is removed to make it fast at the cost of accuracy. I saw the use cases and i can say for sure one thing all of these cases can be ahndled by classic ML/ regression/ classification at almost no cost with sub mili second latency for a particular domain.

> for a particular domain I think the point is that it generalizes over many domain though? The main qualitative improvement of large-scale LLMs was their generalizability… if you reduce them to a low enough level, LLMs are “just classification in token space”. A classical ML classifier over the same token space won’t be able to do what LLMs do today, just as a classical ML classifier over the a similar domain won’t be able to do what Jev (and now similar OSS “Decision models”) can do

Ya may be. But here is a question : Can we define a state in detail with all the output types and input types. I was just thinking where and why will anyone use it. Apart from playing video games taking json as a state ot may be a llm router like Openrouter. I was thinking about one of our product which on a high level detect fraud call/transactions etc, will i use it? For the time being i am not sure. I have also read about its RAG abilities for rerank etc, but i have not tried it

• Customer solutions triaging and resolution: one typically pre-defines triage levels, resolution types, etc. • Code review: folks typically also define category scheme of errors/bugs/code smells to catch • etc. etc. Basically any application where architects define an ontology of entities/relationships that guardrails the application

Ya may be you are right. I just don’t see it right now. Laya -> Faster , less accurate Jev -> Slower, Bit more accurate Based on my experiments

it’s just the AI hype cycle

I just want to see people work on something other than transformers, so was exited to see it

what do you work on?

Mostly post training if LLMs Other wise classical ML -> Regression/classification, good old days of rand forest etc But in my spare time i am working on spatial representation of action/trajectories -> which is a facy way of saying i want to figure out out of all the representation possile to find a particular probabilty distribution to represnt a data whcih/how to find the best one and adapt as thing change