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

Learning to Reason with LLMs

openai.com

Post reply on HN
Print view View original

321–330 of 1001 posts

Re: Learning to Reason with LLMs

#321

Post by randomdata »

oblio wrote: post #308
randomdata wrote: post #211

Earlier quoted context omitted.

They partnered with Microsoft, remember? 1985 – Windows 1.0 1987 – Windows 2.0 1990 – Windows 3.0 1992 – Windows 3.1 1995 – Windows 95 1998 – Windows 98 2000 – Windows ME (Millennium Edition) 2001 – Windows XP 2006 – Windows Vista 2009 – Windows 7 2012 – Windows 8 2013 – Windows 8.1 2015 – Windows 10 2021 – Windows 11

Why did you have to pick on Windows? :-( If you want real atrocities, look at Xbox.

Honestly, it is the only Microsoft product I know. Xbox may be a better example, but I know nothing about the Xbox. But I am interested to learn! What is notable about its naming?

Re: Learning to Reason with LLMs

#322

Post by qqqult »

93po wrote: post #230
rglullis wrote: post #192

Earlier quoted context omitted.

When are they going to change the name to reflect their complete change of direction? Also, what is going to be their excuse to defend themselves against copyright lawsuits if they are going to "understandably" keep their models closed?

[flagged]

> literally anyone can use it for free, you don't even need an account

how can you access it without an account?

Re: Learning to Reason with LLMs

#323

Post by Hizonner »

minimaxir wrote: post #48

> Therefore, after weighing multiple factors including user experience, competitive advantage, and the option to pursue the chain of thought monitoring, we have decided not to show the raw chains of thought to users. What? I agree people who typically use the free ChatGPT webapp won't care about raw chain-of-thoughts, but OpenAI is opening an API endpoint for the O1 model and downstream developers very very much care…

They they've taken at least some of the hobbles off for the chain of thought, so the chain of thought will also include stuff like "I shouldn't say ".

Re: Learning to Reason with LLMs

#324

Post by Cu3PO42 »

cja wrote: post #288
cal85 wrote: post #54

Sounds great, but so does their "new flagship model that can reason across audio, vision, and text in real time" announced in May. [0] [0] https://openai.com/index/hello-gpt-4o/

Recently I was starting to think I imagined that. Back then they gave me the impression it would be released within week or so of the announcement. Have they explained the delay?

It is definitely available today and I believe it was available shortly after the announcement.

Re: Learning to Reason with LLMs

#325

Post by jsheard »

swatcoder wrote: post #251
ARandumGuy wrote: post #137

One thing that makes me skeptical is the lack of specific labels on the first two accuracy graphs. They just say it's a "log scale", without giving even a ballpark on the amount of time it took. Did the 80% accuracy test results take 10 seconds of compute? 10 minutes? 10 hours? 10 days? It's impossible to say with the data they've given us. The coding section indicates "ten hours to solve six challenging algorithmic…

> Did the 80% accuracy test results take 10 seconds of compute? 10 minutes? 10 hours? 10 days? It's impossible to say with the data they've given us. The gist of the answer is hiding in plain sight: it took so long, on an exponential cost function, that they couldn't afford to explore any further. The better their max demonstrated accuracy, the more impressive this report is. So why stop where they did? Why omit actu…

See also: them still sitting on Sora seven months after announcing it. They've never given any indication whatsoever of how much compute it uses, so it may be impossible to release in its current state without charging an exorbitant amount of money per generation. We do know from people who have used it that it takes between 10 and 20 minutes to render a shot, but how much hardware is being tied up during that time is a mystery.

Re: Learning to Reason with LLMs

#326

Post by utdiscant »

Feels like a lot of commenters here miss the difference between just doing chain-of-thought prompting, and what is happening here, which is learning a good chain of thought strategy using reinforcement learning.

"Through reinforcement learning, o1 learns to hone its chain of thought and refine the strategies it uses."

When looking at the chain of thought (COT) in the examples, you can see that the model employs different COT strategies depending on which problem it is trying to solve.

Re: Learning to Reason with LLMs

#327

Post by skywhopper »

ARandumGuy wrote: post #137

One thing that makes me skeptical is the lack of specific labels on the first two accuracy graphs. They just say it's a "log scale", without giving even a ballpark on the amount of time it took. Did the 80% accuracy test results take 10 seconds of compute? 10 minutes? 10 hours? 10 days? It's impossible to say with the data they've given us. The coding section indicates "ten hours to solve six challenging algorithmic…

Yeah, this hiding of the details is a huge red flag to me. Even if it takes 10 days, it’s still impressive! But if they’re afraid to say that, it tells me they are more concerned about selling the hype than building a quality product.

Re: Learning to Reason with LLMs

#328

Post by Hizonner »

trash_cat wrote: post #169
gliiics wrote: post #20

Congrats to OpenAI for yet another product that has nothing to do with the word "open"

It is open in the sense that everyone can use it.

Only people who exactly share OpenAI's concepts of what "alignment" and "safety" should mean can use it to its full potential.

Re: Learning to Reason with LLMs

#330

Post by fraboniface »

logicchains wrote: post #161
k2xl wrote: post #88

Pricing page updated for O1 API costs. https://openai.com/api/pricing/ $15.00 / 1M input tokens $60.00 / 1M output tokens For o1 preview Approx 3x the price of gpt4o. o1-mini $3.00 / 1M input tokens $12.00 / 1M output tokens About 60% of the cost of gpt4o. Much more expensive than gpt4o-mini. Curious on the performance/tokens per second for these new massive models.

I guess they'd also charge for the chain of thought tokens, of which there may be many, even if users can't see them.

That would be very bad product design. My understanding is that the model itself is similar to GPT4o in architecture but trained and used differently. So the 5x relative increase in output token cost likely already accounts for hidden tokens and additional compute.
Post reply on HN