this is the most epic meme arising from interprtability research lol, good for committing to the bit
@swyx
-
Thoughts on 8x22b Release Without 8x7b Model
By
–
its honestly a bit odd they uploaded 8x22b first without 8x7b. but surely it cant be far behind now esp if its really sparse upcycled (do we know for sure?)
-
Grouping Neurons Doesn’t Create New Knowledge
By
–
not at all bc it only groups existing neurons together but doenst add new knowledge
-
Anthropic Keynote Coming Soon at AI Engineer Conference
By
–
do also stay tuned for the Anthropic keynote at @aidotengineer 🙂 https://
ai.engineer -
Claude Sonnet/Opus Outperforms GPT-4o in Summarization Quality
By
–
claude sonnet/opus still has the best summarization quality vs gpt4o. i've run both for ainews for a while and maybe its a prompting skill issue but claude just "gets" my prompts a lot better.
-
Request for Demo API to Test Feature Clamping
By
–
i'd LOVE a demo api for the rest of us to try out these features.. just send dict of feature id and clamp value? you guys would be first to market on an API for this!!!
-
Prompt Engineering Evolution: From Art to Science for AI
By
–
it doesnt kill prompt engineering, but by god it simplifies tweaking model behavior from a vague art for wordcels toward a science for tensor rotators. more in ainews
-
Anthropic’s Progress Toward Productizable AI Interpretability
By
–
IMO @AnthropicAI is very close to making a breakthrough in productizable interpretability. For ~4 years all we've had to really control LLMs is temperature/top_p and logit bias. We recently got `seed` and constrained structured output, with `interactive=false` on the way. But
-
Hyperwrite Assistant Limitations: Recording Requirements Instead of Direct Instructions
By
–
perhaps @mattshumer_ can correct me but i dont think hyperwrite assistant does what i ask. just like adept workflows, it requires me to record what to do first (exactly the thing i'm trying to skip) instead of just giving it instructions and then checking in on it every now and
-
Training Time vs Committee Discussions on Voice Selection
By
–
this will take like 2 hrs of training, 2000 hours of committees on what voice to train