Profile avatar
drsezzer.bsky.social
Software Engineer researching LLM agents, at The Alan Turing Institute. Witcher fan (books/games not netflix). Recreates retro games in python for fun! *Opinions are mine* or borrowed from those more insightful.
181 posts 128 followers 134 following
Prolific Poster
Conversation Starter
comment in response to post
I'm saying, I assumed the icky bit, is from a joint paper by Apollo research and Anthropic. This pairing seems to be behind much of the LLMs have a dark side Comms. And that i could make a full time job outta debunking their papers.
comment in response to post
I spend a lot of time debunking their papers! ... Scully would be proud.
comment in response to post
Linked to Apollo research by any chance? They are very good at spinning their results to strongly hint at GenAI having malign intent.
comment in response to post
We need to bring back good polytechnics!
comment in response to post
Has a few typos and formatting issues, will get to them tomorrow.
comment in response to post
Haha! That's pretty much the last line in my draft blog on coding assistants! drsezzer.github.io/my_thoughts_...
comment in response to post
Isn't UML (or similar) the Lingua Franca here?
comment in response to post
Haha, nope. Apparently it was related to her philosophy and ethics course.
comment in response to post
FREE TO READ: Eleanor Shearer of @cmmonwealth.bsky.social and Matt Davies (@halcyene.bsky.social) of @adalovelaceinst.bsky.social assess the strengths and limits of Labour's AI Opportunities Plan, and make the case for a progressive alternative renewal.org.uk/articles/cod...
comment in response to post
Better memorisation is likely a response to model convergence.
comment in response to post
Better memorisation is likely a response to model convergence.
comment in response to post
A serious lack of diversity on the linked in example!
comment in response to post
Or 40+ arXiv tabs open...
comment in response to post
I do hope you're talking about Battle Beyond the Stars!
comment in response to post
I can't tell what's more icky, the reform-voter pleasing policy or the fact they can so easily switch to believing this shit. Vacuous either way.
comment in response to post
Something in the user prompt or system setup encouraged the LLM to suggest the outlier set of results as a definitive answer/source.