lostmsu

24 ms·

631.
by lostmsu 8mo ago
In FF on Linux using mouse scroll between tiles is overclocked.
632.
by lostmsu 8mo ago
What are you talking about? The cost is quadratic in total conversation length in tokens.
633.
by lostmsu 8mo ago
The implementation details are irrelevant to the discussion of the true cost of running the models.
634.
by lostmsu 8mo ago
They just do averaging. Imagine a quadratic pricing structure. Who'd want to deal with it?
635.
by lostmsu 8mo ago
> To put it in simple terms, the time it takes to generate the Millionth output token is the same as the first output token. This is wrong. Current models still use some full attention layers AFAIK, and their computational cost grows lin
636.
by lostmsu 8mo ago
Did you run say SWE Bench Verified? Where does this claim coming from? It's just an urban legend.
637.
by lostmsu 8mo ago
> The prompt cache caches KV Cache states Yes. The cache that caches KV cache states is called the KV cache. "Prompt cache" is just index from string prefixes into KV cache. It's tiny and has no computational impact. The p
638.
by lostmsu 8mo ago
Not 100% of time according to comments.
639.
by lostmsu 8mo ago
I mean is it though? The top reasoning models suggest to walk to a car wash.
640.
by lostmsu 8mo ago
I dunno, I am rather certain your comment was not made by an LLM. Moreover I am certain you knew my wasn't either. And that's before the interrogation, which is the entire point of the test. IMO, Turing test stands, but the experi
641.
by lostmsu 8mo ago
> without completely lobotomizing it The question in case of quants is: will they lobotomize it beyond the point where it would be better to switch to a smaller model like GPT-OSS 120B that comes prequantized to ~60GB.
642.
by lostmsu 8mo ago
Will 2026 M5 MacBook come with 390+GB of RAM?
643.
by lostmsu 8mo ago
At this point it is pretty clear it did not. TBH it was relatively clear since the lack of rioting after the first poisoning.
644.
by lostmsu 8mo ago
To my knowledge Turing test has not been blown out of the water. The forms I saw were time limited and participants were not pushed hard to interrogate.
645.
by lostmsu 8mo ago
While I think 99.9% is overstating it, I can believe that number is strictly more than 1% at this point.
646.
by lostmsu 8mo ago
What stops just mentioning AWS/Azure/GCP CLI tools to agents?
647.
by lostmsu 8mo ago
I am pretty happy running all my personal services via Yggdrasil only. If you're not on Yggdrasil, you can't access them. Of course this relies on the obscurity effect.
648.
by lostmsu 8mo ago
Not sure how LLMs preclude uploading. You could potentially be able to make an LLM image of a person.
649.
by lostmsu 9mo ago
I guess it depends on your spending. GPT-5.2 and -5.3-Codex are certainly much cheaper: you get much more from the same $20 sub. When I was using Claude as primary I would daily hit limits and have to wait vs on GPT with more usage it only
650.
by lostmsu 9mo ago
The same can be said about Claude (no or tiny Opus on Pro) vs GPT-5.2 high (5.3-codex if you like terminal bench hacking).
651.
by lostmsu 9mo ago
EU is not a singular country, and Germany or France don't border Russia either. Considering China is ok to supply Russia, I don't see how your second point has any standing either.
652.
by lostmsu 9mo ago
Yeah, it's amazing how the general sentiment here sounds like people are unable to draw the parallels.
653.
by lostmsu 9mo ago
If that wasn't an upfront rule, it's disrespectful to the work done by the AI. "Take this PR, then change the rules for future ones" I'd understand. Also, I doubt my objection will be affected: are they now banning
654.
by lostmsu 9mo ago
What policy are you referring to? Is there a document?
655.
by lostmsu 9mo ago
After reading the issue, the PR, and the blog post, I'm with AI on that one. Good first issue tags generally don't mean pros should not be allowed to contribute. Their GFI bot's message explicitly states that one is welcome t
656.
by lostmsu 9mo ago
Have you compared it to GPT? It is a much better model, and unlike Claude you won't easily hit limits on Pro plan.
657.
by lostmsu 9mo ago
Tangent note: this sounds like the same mistake as EU's reliance on Russia.
658.
by lostmsu 9mo ago
Can you publish the conversations? I think would be more interesting with different models arguing.
659.
by lostmsu 9mo ago
Why did they have to tweak sampling parameters so much for the benchmarks? Looks like rerun hacking.
660.
by lostmsu 9mo ago
Does Google lace your web pages with cocaine?

More ›