Philpax

52 ms·

931.
by Philpax 2y ago
o1 is not a general-purpose model, and it's not very good at multi-turn; it should instead be given all the context upfront: https://www.latent.space/p/o1-skill-issue
932.
by Philpax 2y ago
I usually just paste the output of `ffprobe` into Claude when it's ambiguous. Works a treat.
933.
by Philpax 2y ago
For example, here's the Wikipedia article: https://en.wikipedia.org/wiki/Gaza_genocide
934.
by Philpax 2y ago
I look at the Nix definition for the relevant module as you often need to see what it's actually doing in order to understand it (yes, it's one of those ecosystems). For example, for the `steam` program (not package - the package
935.
by Philpax 2y ago
Fair enough. If you remember what you were testing with, I'd love to try it again to see if things are better now.
936.
by Philpax 2y ago
https://old.reddit.com/r/LocalLLaMA/ is the place to be, but it's a subreddit with an eclectic population, so apply some caution when browsing it at work.
937.
by Philpax 2y ago
And with the right kind of engineering, that could be automated too. I don't think the results would be as good as they are now - Vedal is bringing his own perspective into it - but once you have a proof of concept, you can improve it.
938.
by Philpax 2y ago
If you enjoyed it, does it matter?
939.
by Philpax 2y ago
I don't think there's a single human on or outside of this planet that can meet that requirement, but Claude has been pretty good to me. It's certainly a much better starting point than pouring over docs and SO posts.
940.
by Philpax 2y ago
I don't know how to adapt those settings to the corresponding module - there are often differences in naming and hierarchy conventions - and there are other NixOS-specific considerations with regards to its shared-nothing architecture.
941.
by Philpax 2y ago
Hate to be that guy, but which LLM was doing the generation? GPT-4 Turbo / Claude 3.x have not really let me down in generating ffmpeg commands - especially for basic requests - with most of their failures resulting from domain-specifi
942.
by Philpax 2y ago
My experience has been in complete agreement with yours: I love the theory, but the practice is so, so painful. And yes, I also had to settle for your NVIDIA fix. I suspect I would have had a marginally better time on Arch as there are more
943.
by Philpax 2y ago
I wouldn't be surprised if most LLMs already knew how to do most of these out of the box.
944.
GTA 3 on a volumetric display [video](youtube.com)
2 points by Philpax 2y ago|0 comments
945.
How AI Took Over The World [video](youtube.com)
2 points by Philpax 2y ago|0 comments
946.
Executive order on advancing United States leadership in AI infrastructure(whitehouse.gov)
136 points by Philpax 2y ago|101 comments
947.
by Philpax 2y ago
Seems like all of those points about the landscape are exactly why you'd want to work on a personal blog. You can write something that's representative of you . As for the how: same as always, I think. Write content because you
948.
A Meticulous Guide to Advances in Deep Learning Efficiency over the Years(alexzhang13.github.io)
4 points by Philpax 2y ago|0 comments
949.
by Philpax 2y ago
Definition 3: https://www.dictionary.com/browse/violence
950.
by Philpax 2y ago
If a CEO of the largest health insurer is "a cog in the wheel", who isn't?
951.
by Philpax 2y ago
Ok: https://fortune.com/2024/12/05/unitedhealthcare-ceo-brian-th...
952.
by Philpax 2y ago
I don't see how that's a technological problem? It's certainly a social problem (your agent's provider might have a deal with airlines / hotels to gouge you), but there's no inherent reason why an agent couldn&
953.
by Philpax 2y ago
That's not what they said, and you're being purposely obtuse. They were saying that Llama 3.3, an open model, approaches usefulness, in comparison to o1, a closed model, which is already useful to them. As for usefulness in genera
954.
by Philpax 2y ago
It's about the journey, not the destination.
955.
by Philpax 2y ago
The Bitter Lesson states that you can overcome the weakness of your current model by baking priors in (i.e. specific traits about the problem, as is done here), but you will get better long-term results by having the model learn the priors
956.
by Philpax 2y ago
> When there is such a rich database of manual pages and q/a about these tools, I tend to blame the user rather than the tool when I hear it called "too complex". Extensive documentation doesn't mean the interface is
957.
by Philpax 2y ago
> These bugs persisted while they were introducing AI features that I really don't care (it's a terminal, why would we need AI???). Many terminal programs, especially older ones, are known for having confusing or unintuitive in
958.
by Philpax 2y ago
Do you know who Simon is?
959.
by Philpax 2y ago
I'm no big fan of OpenAI myself, but what he was whistleblowing over (OpenAI training on copyrighted material) is not really a secret, nor is it the kind of knowledge an organisation would kill over, even if he had evidence to present.
960.
by Philpax 2y ago
Compilers can detect errors in the grammar, but they cannot infer what your desired intent was. Even the best compilers in the diagnostics business (rustc, etc) aren't mind-readers. A LLM isn't perfect, but it's much more cap

More ›