Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.
i've been letting fable run experiments (mostly autonomously) to test a few ideas we had about jlens when does the structure form in training? can you transfer between models? how does it scale? how far does a nudge travel? what does K2 jlens look like? https://eliebak.com/viz/jspace-open-v2 https://x.com/eliebakouch/status/2076417092039901505/photo/1 https://twitter.com/eliebakouch/status/2074532904009421260
since i'm not a great writer, i wanted to try a different way to do "blogging". less time consuming for me and more aligned with how i think people actually read blogs now, happy to have feedback on this! (i still removed a bunch of the slop for people who like to read blogs) for the results: i think most of my hypotheses were wrong lol (except the transfer one (which was from fable) which seems to be strong). i didn't have time to properly analyze/undersatnd every result yet and wanted to put it in the open to have ppl more knowledgeable than me to give their opinion! all experiment logs/results are open, lmk if you find cool stuff also i tried to use a bunch of open source artifacts to find some correlations such as delphi (@WilliamBarrHeld), olmo (@soldni @natolambert and al) and our little smollm3 (@LoubnaBenAllal1) was also a good testbed for me to find the limitation of frontier model like fable on such task https://github.com/eliebak/open-jlens-data
.@eliebakouch managing agent to do autonomous research https://x.com/samsja19/status/2076423851995775042/photo/1 https://twitter.com/eliebakouch/status/2076417092039901505
The study found universal structural patterns forming during training.
i've been letting fable run experiments (mostly autonomously) to test a few ideas we had about jlens when does the structure form in training? can you transfer between models? how does it scale? how far does a nudge travel? what does K2 jlens look like? https://eliebak.com/viz/jspace-open-v2 https://x.com/eliebakouch/status/2076417092039901505/photo/1 https://twitter.com/eliebakouch/status/2074532904009421260
since i'm not a great writer, i wanted to try a different way to do "blogging". less time consuming for me and more aligned with how i think people actually read blogs now, happy to have feedback on this! (i still removed a bunch of the slop for people who like to read blogs) for the results: i think most of my hypotheses were wrong lol (except the transfer one (which was from fable) which seems to be strong). i didn't have time to properly analyze/undersatnd every result yet and wanted to put it in the open to have ppl more knowledgeable than me to give their opinion! all experiment logs/results are open, lmk if you find cool stuff also i tried to use a bunch of open source artifacts to find some correlations such as delphi (@WilliamBarrHeld), olmo (@soldni @natolambert and al) and our little smollm3 (@LoubnaBenAllal1) was also a good testbed for me to find the limitation of frontier model like fable on such task https://github.com/eliebak/open-jlens-data
.@eliebakouch managing agent to do autonomous research https://x.com/samsja19/status/2076423851995775042/photo/1 https://twitter.com/eliebakouch/status/2076417092039901505
Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.