redox99

7 ms·

121.
by redox99 2mo ago
I have no problem running two or three sequences of qwen 27B with a 3090. It's basically the recommended way, LLM inference without batching is super inefficient.
122.
by redox99 2mo ago
Most people probably don't care about nginx performance. It shouldn't be your bottleneck unless you serve massive amounts of static data.
123.
by redox99 2mo ago
You don't really need to train a 10T model to test cerebras against a 10T model. You can feed it an untrained (randomly initialized) model and benchmark it. Result will be gibberish but performance the same.
124.
by redox99 2mo ago
Correct me if I'm wrong, but it didn't actually generate CP, it changed people into bikinis, not actual nudity. You can go to the beach or (I assume) search on Google and find people of any age wearing bikinis. And that's fin
125.
by redox99 2mo ago
It doesn't.
126.
by redox99 2mo ago
Obviously the missing part (which we already have for software and math) is that we need agents to be able to run automated loops in the real world. That basically requires robots. I think we'll be there in less than 5 years.
127.
by redox99 2mo ago
Usually at least here it's tens of thousands of laws like "replace article 3 paragraph 2 with:" and it's not like you can ctrl f "lemonade stand" and get something useful. A strong llm without that knowledge ca
128.
by redox99 2mo ago
The human genome is 800MB, (and 500MB with compression according to GPT) so that part is easy. I think the problem is that the conversation drifted from "all useful human knowledge" to "enough knowledge to do my job". Bo
129.
by redox99 2mo ago
I think a specialized model could squeeze all you need to know about a certain profession in 30GB. But not all professions at once, which is what these models try to do. Or maybe not, maybe there's a world model needed for human level
130.
by redox99 2mo ago
It's definitely trained on generating SVGs. Not that they trained on the pelican prompt itself, but probably a more general RL step that has it generate SVG and a stronger vision model judges it. It's up to you if that counts as b
131.
by redox99 2mo ago
(mumbles) Shannon entropy... Kolmogorov complexity... something, something... On a more serious note, it depends on your cutoff for "entire human knowledge". It's easy to prove for a generous interpretations of "entire h
132.
by redox99 2mo ago
Yeah but results are worse, see https://news.ycombinator.com/item?id=49301574
133.
by redox99 2mo ago
Asking it factual information[1]. You just can't compress the entire human knowledge into a 30GB file. [1] Without searching the internet. And even if you allow it, you'll get much worse results because search means browsing and p
134.
by redox99 2mo ago
Yep. These small models are actually worse than GPT 3.5 at some tasks (like recalling facts). You can definitely make models smarter at specific tasks (like tool calling, coding) but you can't compress the entire human knowledge into a
135.
by redox99 2mo ago
Benchmarks mean very little. The difference between Luna and Sol in the real world is massive.
136.
by redox99 2mo ago
Almost surely the free model on the web ui. So a dumb model, no harness, and probably a bad prompt.
137.
by redox99 2mo ago
The cache reads are really insanely expensive.
138.
by redox99 2mo ago
Yeah [1] is really a thing and SlopCodeBench ( https://www.scbench.ai/ ) kind of measures that. You need to manually push models to clean up the slop every now and then otherwise it becomes chaotic. And every change with LLMs
139.
by redox99 2mo ago
I'm sure Grok 4.6 is not Fable level. Benchmarks are almost useless. Having said that, Grok 4.6 (1.5T params) is without a doubt way smaller than Fable, maybe a Fable sized Grok would be Fable level?
140.
by redox99 2mo ago
Yeah the speed was very nice indeed.
141.
by redox99 2mo ago
That's pretty surprising. Idk about Grok 4.6, but Grok 4.5 was clearly below Fable, Opus 5 and GPT 5.6 Sol.
142.
by redox99 2mo ago
It's a bit worse. I haven't tried so it's pure speculation based on benchmarks, but I'd assume Grok 4.6 is around Opus 4.8 in real world use, but clearly below Opus 5.
143.
by redox99 2mo ago
They weren't sold out, they had bad seats or didn't have 4 contiguous ones.
144.
by redox99 2mo ago
For a Linux user, you can already build such a system yourself quite trivially by...
145.
by redox99 2mo ago
Computer use will suck until we get 500+ tk/s
146.
by redox99 2mo ago
The internet sucks. Yesterday, I had my agent search for openings for The Odyssey that fit my requirements and then book them. It was way better than manually looking at seat maps for 20 different showtimes and going through 10 steps just t
147.
by redox99 2mo ago
They're improving but you definitely need both a strong AGENTS.md and also usually many LLM passes (one for implementing a feature, another one for code quality, large passes every now and then for major refactors, etc). That's no
148.
by redox99 2mo ago
Obviously sleep is necessary. I'm saying that there isn't always a large pressure to minimize sleep. Long sleep can be beneficial just to preserve energy.
149.
by redox99 2mo ago
That one for some reason is [flagged]
150.
by redox99 2mo ago
You definitely see a bit of this in other primates. They sleep on trees (much safer) and for a lot more hours than humans.

More ›