Yup, LLMs much more relevant than DeepRL
MACHINE LEARNING
-
Large financial firms adoption of deep learning since 2016
By
–
Some large hedge funds, investment banks, and financial firms started getting into deep learning around 2016 (though probably not into Deep RL). That probably paid off well!
-
Deep RL for Real Robots: Limited Success in 2016
By
–
Deep RL for real robots didn’t really work so well back in 2016.
-
Deep RL at Peak Hype in 2016, AGI Hopes
By
–
Deep RL was at peak hype during NIPS2016. Everyone thought training a ConvNet to play a few Atari games using Q-Learning would lead to AGI.
-
AI knows URLs from training data, cannot browse web
By
–
I assure you it can't browse the web, it just knows the URLs for many, many websites. It's seen many URLs in its training data, and can both recite real ones and hallucinate plausible fake ones.
-
Sliding window of 8192 tokens confirmed in latest update
By
–
See my update in the replies — I now believe it's simply a sliding window of length 8192 tokens.
-
ChatGPT uses a simple sliding window of 8192 tokens
By
–
I now believe ChatGPT uses a simple sliding window with a context window of 8192 tokens:
-
ChatGPT context length 8192 tokens inferred from repetition test.
By
–
Update: Recollection of the start of the conversation reliably fails after 7 repetitions of the ~1024-token input. This suggests that text-davinci-002-render, the model underlying ChatGPT, has a likely context length of 8192 tokens, with some tokens consumed by formatting.
-
ChatGPT recalls content beyond 4096-token context window
By
–
A demonstration that ChatGPT is able to recall the content of conversations beyond the length of a 4096-token context window. Uses four messages of 1024+ tokens, each ending with instructions to demonstrate that input is not truncated before the model sees it.
-
Falcon: Model-Based RL for Extreme Turbulent Environments
By
–
— Prof. Anima Anandkumar (@AnimaAnandkumar) 3 décembre 2022
Tune in for our talk on Falcon, first model-based RL method in real-world extreme turbulent environments (Reynolds Num. 230000!), AI4Science workshop @ #Neurips2022, today @ 1:20pm. Falcon achieves 37% improvement over SOTA in stabilization of an airfoil using only 8 min of data!