Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

QwQ: Alibaba's O1-like reasoning LLM

qwenlm.github.io

Post reply on HN
Print view View original

311–320 of 435 posts

Re: QwQ: Alibaba's O1-like reasoning LLM

#311

Post by whimsicalism »

roenxi wrote: post #238
fngjdflmdflg wrote: post #231

It gets the Sally question correct, but it takes more than 100 lines of reasoning. >Sally has three brothers. Each brother has two sisters. How many sisters does sally have? Here is the answer: https://pastebin.com/JP2V92Kh

In fairness it actually works out the correct answer fairly quickly (20 lines, including a false start and correction thereof). It seems to have identified (correctly) that this is a tricky question that it is struggling with so it does a lot of checking.

overthinking is also a problem o1 struggles with

Re: QwQ: Alibaba's O1-like reasoning LLM

#313

Post by seanmcdirmid »

TrainedMonkey wrote: post #312

So western controls on training hardware (hello NVIDIA) seem to have failed. I wonder if there will be any repercussions here.

Or they could be training the models in the states? It’s hard to say since alibaba does R&D in Bellevue as well as Hangzhou.

Re: QwQ: Alibaba's O1-like reasoning LLM

#314

Post by evilduck »

anonym29 wrote: post #297
nextworddev wrote: post #296

Is their repo / model free of any undisclosed telemetry, ie is it purely weights

Is it even possible to embed telemetry into a model itself, as opposed to the runtime environment / program (e.g. Ollama)? I would be disinclined to believe that to be possible, but if anyone knows otherwise, please share.

That's literally why the safetensor format exists. The previous pickle checkpoint (ckpt) format allowed for arbitrary code execution.

Re: QwQ: Alibaba's O1-like reasoning LLM

#315

Post by echelon »

TrainedMonkey wrote: post #312

So western controls on training hardware (hello NVIDIA) seem to have failed. I wonder if there will be any repercussions here.

Most of the papers in machine learning are coming from China. The vast majority.

Most of the open source models on GitHub, too.

Hailuo, Kling, Vidu, and Hunyuan (posted on Banodoko) blow Sora and Runway out of the water.

China is dominating at this field. And if they begin releasing weights as open source, that'll mean foundation model companies can only bank on the thin facade of product. That's a really good strategy to make sure American AI startups don't achieve escape velocity if they have to fend of dozens of fungible clones.

Re: QwQ: Alibaba's O1-like reasoning LLM

#316

Post by imdsm »

echelon wrote: post #315
TrainedMonkey wrote: post #312

So western controls on training hardware (hello NVIDIA) seem to have failed. I wonder if there will be any repercussions here.

Most of the papers in machine learning are coming from China. The vast majority. Most of the open source models on GitHub, too. Hailuo, Kling, Vidu, and Hunyuan (posted on Banodoko) blow Sora and Runway out of the water. China is dominating at this field. And if they begin releasing weights as open source, that'll mean foundation model companies can only bank on the thin facade of product. That's a really good strate…

Are they making breakthroughs or are they taking what exists and copying/marginally improving?

Re: QwQ: Alibaba's O1-like reasoning LLM

#317

Post by cscurmudgeon »

echelon wrote: post #315
TrainedMonkey wrote: post #312

So western controls on training hardware (hello NVIDIA) seem to have failed. I wonder if there will be any repercussions here.

Most of the papers in machine learning are coming from China. The vast majority. Most of the open source models on GitHub, too. Hailuo, Kling, Vidu, and Hunyuan (posted on Banodoko) blow Sora and Runway out of the water. China is dominating at this field. And if they begin releasing weights as open source, that'll mean foundation model companies can only bank on the thin facade of product. That's a really good strate…

Quantity of papers is really not a good measure.

> that'll mean foundation model companies can only bank on the thin facade of product.

The “facade” of product tested in the real world in the hands of millions or billions is better than thousands of unread/uncited/clique-cited papers using questionable gameable benchmarks.

Re: QwQ: Alibaba's O1-like reasoning LLM

#318

Post by hamilyon2 »

JKCalhoun wrote: post #285

Is Alibaba's LLM the "Chinese LLM"? It would appear to have been a U.S.-only game until now. As Eric Schmidt said in the YouTube lecture (that keeps getting pulled down), LLM's have been a rich-companies game.

Have you tried Mistral? I am blown away by how good it is.

Re: QwQ: Alibaba's O1-like reasoning LLM

#320

Post by n144q »

seanmcdirmid wrote: post #313
TrainedMonkey wrote: post #312

So western controls on training hardware (hello NVIDIA) seem to have failed. I wonder if there will be any repercussions here.

Or they could be training the models in the states? It’s hard to say since alibaba does R&D in Bellevue as well as Hangzhou.

My understanding is that R&D in the US is a very small part of the business for these China-based companies, both in terms of headcounts and org chart. TikTok hires many people in the US, but that's nowhere near the number in the Beijing office, not to mention that all decisions come from China. You can find lots of articles that talk about midnight meetings for folks in US. And there is a good reason for that -- you can hire a few engineers in China for the same price of hiring 1 engineer in the Bay Area.
Post reply on HN