tim ganiev

Nanni thought that this would show how bad of a merchant I am but little does he know collecting his complaints it’s the reason my heart still beats

there's no answer on context / memory layer IMO. currently it's a big part of the harness itself. additionally, there's too many stuff must be solved: correct data provenance, importance for the current task, the shape of memory itself – ideally, you don't want to shit the context window with not related pile of stuff. i still don't like that there's too few teams working on it and that their approaches basically set of fine-grained heuristics.

i want my Agent (coding, chat, or otherwise) to have a fuzzy sense of everything we've ever talked about, and be able to recall deeper at will feels like a primitive version of this could be something like a JSON topic cloud, where models can "click in" to keyphrases, which each link to most relevant threads with single-sentence summaries, which then can be expanded into longer summaries, or just text search over the trace directly is there any interesting work along these lines? we have all the pieces, it just seems like we don't have proper language for comparing and contrasting these kinds of methods, and hence you don't see much academic discussion about it, just "infinite memory agent" hypeslop

we're getting close to AGI-as-tool which basically is a semantic SAT solver given an input task description, either find a solution, or return that no solution has been found after N tokens of search RSI here just looks like 9s + efficiency doesn't auto imply AGI-as-employee

@saprmarks oh and with Aryan Bhatt ofc!

but at least i had an opportunity to talk with @saprmarks, which is pretty cool

just fucking kill me

me deciding how to show up to work this morning

@saprmarks either way, not the best start of the weekends

another day another cyberbullying attempt from LLM

imo the meta-optimization + emergent features are the future of agent engineering. you don't know which concepts the model actually operates on (reminds me of polysemantic neurons), so it ends up improving its harness in ways you'd never hand-engineer pretty cool huh

New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.

@thdxr Idk dude. Opentui is a gem, had no render issues with it (especially after painful experience with ink)

oh look everyone yelling at us for doing a fullscreen tui where are you now where the fuck are you huh WHERE?

@Fried_rice @apples_jimmy everyone right now:

@badlogicgames context management >> long context

just get that these are different limits 10 iq post, sorry

bruh

@boxmining it's unethical to use openclaw

is it unethical to ask my openclaw agent to move myself to hermes?

@Teknium what do you want from a real estate agent lmao

They really just hate local models lmao Local models passed GPT-3.5 a year and a half ago and they don't even know this

just discovered @graphite and their stacked PRs handling idk how to live without it anymore, diffs are much cleaner now

@Teknium probably a pretty solid invoice from ant

But what it cooks is the question! 👁️👀👁️

like how will brown just does pro bono rl consulting in your replies