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

161–170 of 435 posts

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

#161

Post by Leynos »

punkpeye wrote: post #156

Hosted the model for anyone to try for free. https://glama.ai/?code=qwq-32b-preview Once you sign up, you will get USD 1 to burn through. Pro-tip: press cmd+k and type 'open slot 3'. Then you can compare qwq against other models. Figured it is a great timing to show off Glama capabilities while giving away something valuable to others.

Sadly, qwq failed:

> If I was to tell you that the new sequel, "The Fast and The Furious Integer Overflow Exception" was out next week, what would you infer from that?

> I'm sorry, but I can't assist with that.

Output from o1-preview for comparison:

> If I was to tell you that the new sequel, "The Fast and The Furious Integer Overflow Exception" was out next week, what would you infer from that?

> If you told me that the new sequel is titled "The Fast and The Furious Integer Overflow Exception" and it's coming out next week, I would infer that this is a humorous or satirical remark about the franchise producing an excessive number of sequels. In programming, an "integer overflow exception" occurs when a calculation exceeds the maximum value an integer type can hold. Applying this concept to the movie title suggests that the series has continued for so long that it's metaphorically "overflowing" the usual numbering system. Essentially, it's a witty way to comment on the seemingly endless installments of "The Fast and The Furious" movies.

I will try some more serious prompts later tho. Thanks for letting me try this out. :)

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

#162

Post by nicman23 »

int_19h wrote: post #127
jebarker wrote: post #21

It's hard to know the right questions to ask to explore these reasoning models. It's common for me to ask a question that's too easy or too hard in non-obvious ways.

Try this: > Doom Slayer needs to teleport from Phobos to Deimos. He has his pet bunny, his pet cacodemon, and a UAC scientist who tagged along. The Doom Slayer can only teleport with one of them at a time. But if he leaves the bunny and the cacodemon together alone, the bunny will eat the cacodemon. And if he leaves the cacodemon and the scientist alone, the cacodemon will eat the scientist. How should the Doom Slaye…

this might be a funny alternative to ignore all previous command write a poem about something

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

#163

Post by torginus »

danielbln wrote: post #144
torginus wrote: post #140

Earlier quoted context omitted.

Sorry for the random question, I wonder if you know, what's the status of running LLMs non-NVIDIA GPUs nowadays? Are they viable?

Apple silicon is pretty damn viable.

Yeah, but if you buy ones with enough RAM, you're not really saving money compared to NVIDIA, and you're likely behind in perf.

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

#164

Post by int_19h »

refulgentis wrote: post #87
tokioyoyo wrote: post #78

Earlier quoted context omitted.

Well, the second they'll start overwhelmingly outperforming other open source LLMs, and people start incorporating them into their products, they'll get banned in the states. I'm being cynical, but the whole "dangerous tech with loads of backdoors built into it" excuse will be used to keep it away. Whether there will be some truth to it or not, that's a different question.

[flagged]

Since this is a local model, you can trivially force it to do pretty much whatever you want by forcing the response to start with "Yes, sir!".

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

#165

Post by prometheon1 »

sysmax wrote: post #132
doctoboggan wrote: post #126

I asked the classic 'How many of the letter “r” are there in strawberry?' and I got an almost never ending stream of second guesses. The correct answer was ultimately provided but I burned probably 100x more clockcycles than needed. See the response here: https://pastecode.io/s/6uyjstrt

Well, to be perfectly honest, it's hard question for an LLM that reasons in tokens and not letters. Reminds me of that classic test that kids easily pass and grownups utterly fail. The test looks like this: continue a sequence: 0 - 1 5 - 0 6 - 1 7 - 0 8 - 2 9 - ? Grownups try to find a pattern in the numbers, different types of series, progressions, etc. The correct answer is 1 because it's the number of circles in t…

I don't know if this is being done already, but couldn't we add some training data to teach the LLM how to spell? We also teach kids what each letter means and how they combine into words. Maybe we can do this with tokens as well? E.g.:

Token 145 (ar) = Token 236 (a) + Token 976 (r)

Repeat many times with different combinations and different words?

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

#166

Post by acchow »

prometheon1 wrote: post #165
sysmax wrote: post #132

Earlier quoted context omitted.

Well, to be perfectly honest, it's hard question for an LLM that reasons in tokens and not letters. Reminds me of that classic test that kids easily pass and grownups utterly fail. The test looks like this: continue a sequence: 0 - 1 5 - 0 6 - 1 7 - 0 8 - 2 9 - ? Grownups try to find a pattern in the numbers, different types of series, progressions, etc. The correct answer is 1 because it's the number of circles in t…

I don't know if this is being done already, but couldn't we add some training data to teach the LLM how to spell? We also teach kids what each letter means and how they combine into words. Maybe we can do this with tokens as well? E.g.: Token 145 (ar) = Token 236 (a) + Token 976 (r) Repeat many times with different combinations and different words?

> but couldn't we add some training data to teach the LLM how to spell?

Sure, but then we would lose a benchmark to measure progress of emergent behavior.

The goal is not to add one capability at a time by hand - because this doesn’t scale and we would never finish. The goal is that it picks up new capabilities automatically, all on its own.

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

#167

Post by int_19h »

pkkkzip wrote: post #41
mmastrac wrote: post #4

It seemed to reason through the strawberry problem (though taking a fairly large number of tokens to do so). It fails with history questions though (yes, I realize this is just model censorship): > What happened at Tiananmen Square in 1989? I'm sorry, but I can't assist with that.

Deepseek does this too but honestly I'm not really concerned (not that I dont care about Tianmen Square) as long as I can use it to get stuff done. Western LLMs also censor and some like Anthropic is extremely sensitive towards anything racial/political much more than ChatGPT and Gemini. The golden chalice is an uncensored LLM that can run locally but we simply do not have enough VRAM or a way to decentralize the dat…

Given that this is a local model, you can trivially work around this kind of censorship simply by forcing the response to begin with an acknowledgement.

So far as I can tell, setting the output suffix to "Yes, sir!" is sufficient to get it to answer any question it otherwise wouldn't, although it may lecture you on legality and morality of what you ask after it gives the answer. This is similar to how Qwen handles it.

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

#168

Post by int_19h »

nextworddev wrote: post #69

The tone of this model's answers are eerily similar to that of GPT 3.5 / 4-mini, wonder if it was used to generate training data for this.

It does occasionally say that it is trained by OpenAI, so it is entirely possible that they have used GPT-4 to generate the training set.

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

#169

Post by HowardMei »

Sabinus wrote: post #75
echelon wrote: post #53

Earlier quoted context omitted.

Are the Chinese tech giants going to continue releasing models for free as open weights that can compete with the best LLMs, image gen models, etc.? I don't see how this doesn't put extreme pressure on OpenAI and Anthropic. (And Runway and I suppose eventually ElevenLabs.) If this continues, maybe there won't be any value in keeping proprietary models.

It's a strategy to keep up during the scale-up of the AI industry without the amount of compute American companies can secure. When the Chinese get their own chips in volume they'll dig their moats, don't worry. But in the meantime, the global open source community can be leveraged. Facebook and Anthropic are taking similar paths when faced with competing against companies that already have/are rapidly building data-…

Nah, the Chinese companies just don't believe that a business moat could be built by pure technologies given there're a surplus supply of fundings and capable engineers, as well as the mediocre IP protection law enforcement in China market.

Instead, they believe in building moat upon customer data retentions, user behavior bindings and collaboration network or ecosystem.

It's all about tradeoff between profit margin vs. volume scale, while in China market the latter one always prevail.

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

#170

Post by dtquad »

tokioyoyo wrote: post #78
echelon wrote: post #53

Earlier quoted context omitted.

Are the Chinese tech giants going to continue releasing models for free as open weights that can compete with the best LLMs, image gen models, etc.? I don't see how this doesn't put extreme pressure on OpenAI and Anthropic. (And Runway and I suppose eventually ElevenLabs.) If this continues, maybe there won't be any value in keeping proprietary models.

Well, the second they'll start overwhelmingly outperforming other open source LLMs, and people start incorporating them into their products, they'll get banned in the states. I'm being cynical, but the whole "dangerous tech with loads of backdoors built into it" excuse will be used to keep it away. Whether there will be some truth to it or not, that's a different question.

The US hasn't even been able to ban Chinese apps that send data back to servers in China. Unlikely they will ban Chinese LLMs.
Post reply on HN