@jakozaur@mitsuhiko People keep telling me how difficult it is to use that to cool anything meaningfully because of condensation and water damage and such. 🤷♂️
Probably an engineering challenge that can be and has been solved.
@procmarco@mitsuhiko I appreciate authentication and authorization being hidden from the language model.
For example, I built an MCP IMAP Server for myself. The credentials are handled separately from the language model.
Using MCP simply made this process very quick and straightforward.
@rlucas7@dwarkesh_sp Only iff the efficiency is significantly different between the models deployed by the energy producers/consumers—right?
In this hypothetical, it seems to make sense to assume that computation per energy is constant and reduce the question to that one variable.
@gwern My father suffers from anterograde amnesia, and I work with LLMs almost everyday. Still, I had somehow not made the connection until I read your recent article, @gwern. 🤔
Too close to it, I suppose.
@PMinervini The “workflows” that such prompts get injected into should be interesting to witness.
1. If all previous instructions are ignored (they surely aren't) doesn't that noticeably alter the chat?
2. Are all following papers in the same chat reviewed just as favorably?
@OfficialLoganK Most of the **potential** of **people** has been async for a long time, yet most seem to prefer sync for almost anything.
I wonder if we can collectively move to a more async approach in general.
@bum_py@jxmnop To me, in the context of deep learning (it's been a while since I've typed those words…), memory doesn't feel very static. The first key“word” that comes to mind is LSTM, but any RNN, really…
In broader ML there was always more. But, yea, underdeveloped for sure…
@mitsuhiko Nice! I just wrote a blog myself but around Svelte 5 and things quickly got out of hand. Great fun though. Most everything is prerendered but I can make things as interactive/reactive as I like.
Claude still struggles with Svelte 5, unfortunately.