yorwba

22 ms·

481.
by yorwba 8mo ago
If you want to experiment with hardcoding small programs into transformer weights, maybe try ALTA: https://arxiv.org/abs/2410.18077v2
482.
by yorwba 8mo ago
I recommend asking a friend who's a better writer and mathematician than Claude Code to help you reorganize the paper so that there are no gaps in the argumentation and incorrect statements like "For a purely causal transformer wi
483.
by yorwba 8mo ago
Yes, I think this is basically an instance of the "emergent abilities mirage." https://arxiv.org/abs/2304.15004 If you measure completion rate on a task where a single mistake can cause a failure, you won&#x2
484.
by yorwba 8mo ago
According to the Shima Seiki history page: https://www.shimaseiki.com/company/dna/history/ it was a world first in 1995. That doesn't make it novel anymore in 2026. I would nonetheless find it interestin
485.
by yorwba 8mo ago
The comment you're replying to is talking about a hypothetical scenario. In any case, the blinding didn't stop Reviewer #2 from calling out obvious AI slop. (Figure 5)
486.
by yorwba 8mo ago
The guidelines don't say anything about not posting something because an LLM told you that you shouldn't...
487.
by yorwba 8mo ago
Note that those comments are written in perfectly understandable English. Further note how often you come across comments written in perfectly understandable English, but they're downvoted anyway. It suggests a bias in writers to assum
488.
by yorwba 8mo ago
It doesn't seem like anybody ever filed the paperwork for it. (A search for "krul" on site:unicode.org doesn't turn up anything.) Unicode symbols don't just magically appear! Somebody has to do a bit of work to make
489.
by yorwba 8mo ago
Not all of them do: https://news.ycombinator.com/item?id=47335156 There are evidently lots of people experimenting with different botting setups. Some do better at blending in than others.
490.
by yorwba 8mo ago
It's a variable-rate codec. The audio is still compressed, but by how much depends on the duration of the segment corresponding to a particular text token. The TTS model predicts one audio token per text token and its duration, and the
491.
by yorwba 8mo ago
Every company is subject to constraints. A bigger budget is not an infinite budget. And there is no tradeoff between efficiency and raw power. An optimization that lets you build a similarly powerful model for less money also lets you build
492.
by yorwba 8mo ago
There are sequential dependencies, so you can't just arbitrarily increase speed by parallelizing over more GPUs. Every token depends on all previous tokens, every layer depends on all previous layers. You can arbitrarily slow a model d
493.
by yorwba 8mo ago
Anthropic is obviously also aware of the benefits of MoE and distilling a larger model into a smaller one, so they could run a model of the same size as Alibaba's for the same inference cost if they want to. Or they can run a slightly
494.
by yorwba 8mo ago
Where are you getting those benchmark figures from? Math-500 should be closer to 98% for both models: https://artificialanalysis.ai/evaluations/math-500?models=de...
495.
by yorwba 8mo ago
> The new 1s have a mandate to at least run on local hardware. They do? Source? But if that's true, it would explain why Minimax, Z.ai and Moonshot are all organized as Singaporean holding companies, with claimed data center locatio
496.
by yorwba 8mo ago
Copyright protects even very abstract aspects of human creative expression, not just the specific form in which it is originally expressed. If you translate a book into another language, or turn it into a silent movie, none of the actual te
497.
by yorwba 8mo ago
The actual MiniMax website is https://www.minimax.io/news/music-25-unlock-instrumental-mus...
498.
by yorwba 8mo ago
Yes, of course there are people in China who, when their job puts them in conflict with their ethics, will decide to do something ethical. I can't think of any war-related examples, since it's been a while since China was involved
499.
by yorwba 8mo ago
Those aren't even percentage increases, but standardized effect sizes. So if you take an individual survey respondent and all you know is that they self-reported higher AI usage, you can guess their answers to the self-reported individ
500.
by yorwba 8mo ago
People could always simply argue in court that their torrenting was free use. If you're just some nobody representing yourself instead of an expensive lawyer acting on behalf of a large company, maybe the judge will even try to be extr
501.
by yorwba 8mo ago
It's subsidized by people who paid the fee when they bought a hard drive to hold something other than pirated media.
502.
by yorwba 8mo ago
The article is about an LLM being used for real maths.
503.
by yorwba 8mo ago
That Wikipedia is not a democracy doesn't mean there are no votes and no elections. https://en.wikipedia.org/wiki/Wikipedia:Administrator_electi...
504.
by yorwba 8mo ago
You're replying to a bot account https://news.ycombinator.com/item?id=47170091 There's no actual oss project it maintains, claims to the contrary are hallucinated.
505.
by yorwba 8mo ago
It's the opposite of a MoE architecture in many ways. MoE splits every individual feed-forward layer into many tiny subnetworks, only a small number of which contribute to the layer output, and they get trained together to complement e
506.
by yorwba 8mo ago
There are smaller ones on HuggingFace https://huggingface.co/models?other=qwen3_5&sort=least_param... with 0.8B, 2B, 4B and 9B parameters.
507.
by yorwba 8mo ago
Part of what makes it so infuriating is that it uses the same patterns so often, the other part is that it's not very good at using them—the revelation that it's Y and not X is typically incredibly banal, not some profound observa
508.
by yorwba 8mo ago
This is not Alibaba, just some rando fine-tuning a model released by Alibaba on Claude reasoning traces collected by other randos.
509.
by yorwba 8mo ago
It doesn't really matter as long as you keep physical control of the code and don't let others copy it.
510.
by yorwba 8mo ago
The article says "Filed takedown notices with Google, Cloudflare, and the domain registrar spaceship.com"

More ›