kekebo

I don't think it's necessarily a bug per se, but a central tradeoff in system prompt length between well-documenting the environment (harness specifics, exposed tools, tool use instructions etc) to the llm, vs the initial prompt stage ("prefill") growing so large that it results in an unpleasant lag to first response, and reduced available context, which is most noticeable with open models on resource-constrained consumer hardware.

You can use llm to optimize some of this, I condensed the tool descriptions of some larger LM Studio plugins to shrink prefill by almost 10k tokens. But there's a soft limit to this, if you don't want to under-document available tools and let the model guess (/behave unsafely).

One optimization around this is called "smart tool selection", which only sends tool descriptions when the model indicates need for a certain tool (suite), not all of them upfront.

I have ocd so I concur with only #2 for equilibrium.

Do you think they're unreliable based on the position in the conversation or other factors?

Supposedly there was a time when drinking beer was often safer than drinking water from available sources, including for children. I guess there are levels

You could reverse the last sentence to "freed disk space reduces the need to clean the hd". And I don't even mean physical limitation as much as the compulsion to manually clean ballooning cache dirs once in a while.

Maybe this is a useful way to address maladaptive llm use for some, but it seems to focus the discussion on either-or (using llms) as opposed to how (to use them).

For things you want to learn: Are you asking to have something solved, or are you asking for information so you can solve it?

There may be detrimental learning effects even with asking llms for information, but I think there's a fundamental difference in addressing it as an intern/slave vs teacher/tutor.

Thanks for sharing.

Clouds of intracellular water more efficiently retained in muscle due to creatine supplementation[?]. Hydration is controlled for in some studies but not all.

I keep having had the best time with Andrej Karpathy's Youtube intros into LLM math. But I haven't compared scope or quality to this submission

Thank you for the comment. https://github.com/umlx5h/zsh-manpage-completion-generator appears to adapt this to ZSH. Have yet to try through

:)) Apparently. I don't have a better guess. Well spotted

I apologize for shortness but <3

Thank you

I just switched away from it again after it got delisted from nixpkgs due to lack of development[0][1]

[0] https://search.nixos.org/packages?channel=25.05&from=0&size=...

[1] Should be this one https://github.com/NixOS/nixpkgs/pull/351205

w/ signed counts

Overton window calculation usually starts at a minimum of two inputs without proximity requirements (except technical requirements like linguistics/ functional semantics)*

[ * Only because an opinion may appear too far removed from a given perceived spectrum-threshold for 'reasonable reasoning'.. should not necessitate collapsing the contrasting input to some purely sarcastic/humorous telos, especially when this stochastically undermines one's own chances for being afforded the inversely congruent gesture]

Is this intentional devils advocacy for the sake of balancing an expected narrative? Outside of the rarely normative definition threshold as to what constitutes spyware or not, on what data / references (if any) do you base your impression on?

And how does a perception of company trustworthiness correlate with telemetry ethics that don't infringe in some way on 'basic digital human rights' (as defined by GDPR et al, say)?

Such confidence ratings over unavailable/lost/missing data appear like a luxury to where I'm from (science wise). Does historical analysis come with heuristic privileges over other fields in that regard?

While partly debated, the first part sounds like a prime candidate for Dunning Kruger effect in (historic) action

Nice, thanks for sharing!

In this case both options seem to apply at the same time

(Corresponding wiki entry for the remote version: https://en.wikipedia.org/wiki/Telephone_card#Remote_memory_s... )

Oh boy, this takes me back. This is from memory, so take it with a grain of salt: The calling cards I mainly used as a kid in the 90s were ones with a sim-like chip on it that you would insert into a phone booth until it ran out at which point you'd throw them away. Earlier/alternative ones worked, iirc, by calling a free phone line which prompted you to enter a valid calling card number which had credit assigned to it, and then you could make your call. If there was a way to reverse engineer the algorithm for valid calling card numbers (similar to a keygen) I guess you could hack it for free calls. If someone has a more plastic / accurate memory of this please correct me

Pretty much

A common trajectory seems to be your sentiment, followed by a couple of reluctant retries, followed by eventual, disproportional reverence. I'm no different

The people who are likely have a limited capacity to.

You may be seeing a feature as a bug

I can't speak to the differences of Openrouter to Together but the Openrouter endpoint should work as a drop-in replacement for OpenAI api calls after replacing the endpoint url and the value of $OPENAI_API_KEY. The model names may differ to other apis but everything else should work the same.

Just watch your phases a bit:

"Several studies of PSG-measured sleep report increased SWS [25, 75], decreased REM sleep [74, 75, 160], and decreased REM density (e.g., number of eye movements during REM sleep) [74, 75]. However, this pattern is not always replicated [150]."[0]

(SWS =slow wave sleep)

[0] https://link.springer.com/article/10.1186/s13722-016-0056-7