-
Reasoning/planning
- Can we conduct thinking step-by-step in latent space? Similar to process supervision, but keep the intermediate steps in latent space without needing them to be materialized (similar to diffusion steps)?
- Can we use something like diffusion for complex reasoning processes?
-
Self-teaching (i.e. thinking, curiosity)
- Can an agent reflect and meditate on some codebase, think about something from various angles / think about something to its logical conclusions, coming up with its own questions (”active learning”), and even experiment?
-
Learning, domain adaptation
- How do we take few-shot learning, and make it work at “training time”? How do we make it work “at scale” beyond a single prompt, so that a model can be adapted to complex new domains? How do we enable efficient learning of a complex new domain, without many examples?
-
Language modeling
- Can we use diffusion for language modeling - or fuse it with autoregression to handle the thinking part?
-
how is research organized? bottom-up, top-down?
-
what are the main focuses?
-
reasoning/planning?
-
have you found that the way research works with product evolved? more focused on LLMs?
-
what are your own goals for the research group?
-
MIT high scaling
-
infer
-
plasmic, built it up from scratch into a profitable business,