www.stat.cmu.edu/~aramdas/icm...

you've fancied some e-values and processes, @sharky6000.bsky.social , have you not

eh... if only it worked...

sure, but isn't overall training objective imitational? why can't the data collected from the open loop + closed loop systems can't be used for the offline RL?

world model part with online off-policy algorithms looks clearer

scrolled though it, very nice survey, very topical

hope it brings back importance of performing offline RL and not behavioural cloning on anything apart from mujoco benchmarks

offline RL should be better but it's not

nice!

congrats on 2 mil

yes, watched it recently. Thank you for asking my question, appreciate the answer Shimon gave

lowkey true: public perception on twitter/bluesky or through arxiv is not much worse than an average academic review

LLM pyramid scheme

But, does it really raise any questions?Papers can be different: from a single but concise experiment to methodological studies or big benchmarks. Some people haven't got resources for their experiments or environment facilitating the dev.It's not only about AI, the latter highlights the difference.

additionally, if you can post any sort of investigation report, would be nice to look what people tried, what worked and what not

thank you

ChatGPT definitely reshaped the landscape

maybe these models are so mind-boggling, no idea

hope people at Sakana will go further with their CTMs, and we will be able to talk about something more than “foundational models”