Learning to Reason with LLMs • Hacker News Board

Re: Learning to Reason with LLMs

#261

Post by narrator » Thu, Sep 12, 2024, 6:16 PM UTC

notamy wrote: post #46

https://openai.com/index/introducing-openai-o1-preview/ > ChatGPT Plus and Team users will be able to access o1 models in ChatGPT starting today. Both o1-preview and o1-mini can be selected manually in the model picker, and at launch, weekly rate limits will be 30 messages for o1-preview and 50 for o1-mini. We are working to increase those rates and enable ChatGPT to automatically choose the right model for a given p…

The human brain uses 20 watts, so yeah we figured out a way to run better than human brain computation by using many orders of magnitude more power. At some point we'll need to reject exponential power usage for more computation. This is one of those interesting civilizational level problems. There's still a lack of recognition that we aren't going to be able to compute all we want to, like we did in the pre-LLM days.

Re: Learning to Reason with LLMs

#262

Post by apsec112 » Thu, Sep 12, 2024, 6:16 PM UTC

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]

AFAIK, they are the least open of the major AI labs. Meta is open-weights and partly open-source. Google DeepMind is mostly closed-weights, but has released a few open models like Gemma. Anthropic's models are fully closed, but they've released their system prompts, safety evals, and have published a fair bit of research (https://www.anthropic.com/research). Anthropic also haven't "released" anything (Sora, GPT-4o realtime) without making it available to customers. All of these groups also have free-usage tiers.

Re: Learning to Reason with LLMs

#263

Post by mewpmewp2 » Thu, Sep 12, 2024, 6:16 PM UTC

vessenes wrote: post #201

djoldman wrote: post #79

> THERE ARE THREE R'S IN STRAWBERRY Ha! This is a nice easteregg.

I appreciated that, too! FWIW, I could get Claude 3.5 to tell me how many rs a python program would tell you there are in strawberry. It didn't like it, though.

I was able to get GPT-4o to calculate characters properly using following prompt:

""" how many R's are in strawberry?

use the following method to calculate - for example Os in Brocolli.

B - 0

R - 0

O - 1

C - 1

O - 2

L - 2

I - 2

Where you keep track after each time you find one character by character

"""

And also later I asked it to only provide a number if the count increased.

This also worked well with longer sentences.

Re: Learning to Reason with LLMs

#264

Post by vessenes » Thu, Sep 12, 2024, 6:16 PM UTC

Note that they aren't safety aligning the chain of thought, instead we have "rules for thee and not for me" -- the public models are going to continue have tighter and tighter rules on appropriate prompting, while internal access will have unfettered access. All research (and this paper mentions it as well) indicates human pref training itself lowers quality of results; maybe the most important thing we could be doing is ensuring truly open access to open models over time.

Also, can't wait to try this out.

Re: Learning to Reason with LLMs

#265

Post by 93po » Thu, Sep 12, 2024, 6:16 PM UTC

dinobones wrote: post #60

thelastparadise wrote: post #41

Earlier quoted context omitted.

Why not? Isn't that basically what humans do? Sit there and think for a while before answering, going down different branches/chains of thought?

This new approach is showing: 1) The "bitter lesson" may not be true, and there is a fundamental limit to transformer intelligence. 2) The "bitter lesson" is true, and there just isn't enough data/compute/energy to train AGI. All the cognition should be happening inside the transformer. Attention is all you need. The possible cognition and reasoning occurring "inside" in high dimensions is much more advanced than any…

Karpathy himself believes that neural networks are perfectly plausible as a key component to AGI. He has said that it doesn't need to be superseded by something better, it's just that everything else around it (especially infrastructure) needs to improve. As one of the most valuable opinions in the entire world on the subject, I tend to trust what he said.

source: https://youtu.be/hM_h0UA7upI?t=973

Re: Learning to Reason with LLMs

#266

Post by csomar » Thu, Sep 12, 2024, 6:16 PM UTC

I gave the Crossword puzzle to Claude and got a correct response[1]. The fact that they are comparing this to gpt4o and not to gpt4 suggests that it is less impressive than they are trying to pretend.

[1]:

Based on the given clues, here's the solved crossword puzzle: +---+---+---+---+---+---+ | E | S | C | A | P | E | +---+---+---+---+---+---+ | S | E | A | L | E | R | +---+---+---+---+---+---+ | T | E | R | E | S | A | +---+---+---+---+---+---+ | A | D | E | P | T | S | +---+---+---+---+---+---+ | T | E | P | E | E | E | +---+---+---+---+---+---+ | E | R | R | O | R | S | +---+---+---+---+---+---+ Across:

ESCAPE (Evade) SEALER (One to close envelopes) TERESA (Mother Teresa) ADEPTS (Initiated people) TEPEE (Native American tent) ERRORS (Mistakes)

Down:

ESTATE (Estate car - Station wagon) SEEDER (Automatic planting machine) CAREER (Profession) ALEPPO (Syrian and Turkish pepper variety) PESTER (Annoy) ERASES (Deletes)

Re: Learning to Reason with LLMs

#267

Post by CooCooCaCha » Thu, Sep 12, 2024, 6:17 PM UTC

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/

My guess is they're going to incorporate all of these advances into gpt-5 so it looks like a "best of all worlds" model.

Re: Learning to Reason with LLMs

#268

Post by tedsanders » Thu, Sep 12, 2024, 6:17 PM UTC

p1esk wrote: post #58

Do people see the new models in the web interface? Mine still shows the old models (I'm a paid subscriber).

They're rolling out gradually over the next few hours. Also be aware there's a weekly rate limit of 30 messages to start.

Re: Learning to Reason with LLMs

#269

Post by from-nibly » Thu, Sep 12, 2024, 6:17 PM UTC

seydor wrote: post #241

wmf wrote: post #184

Earlier quoted context omitted.

People have been celebrating the fact that tokens got 100x cheaper and now here's a new system that will use 100x more tokens.

If it 's reasoning correctly, it shouldnt need a lot of tokens because you don't need to correct it. You only need to ask it to solve nuclear fusion once.

As someone experienced with operations / technical debt / weird company specific nonsense (Platform Engineer). No, you have to solve nuclear fusion at . You gotta do it over and over again. If it were that simple we wouldn't have even needed AI we would have hand written a few things, and then everything would have been legos, and legos of legos, but it takes a LONG time to find new true legos.

Re: Learning to Reason with LLMs

#270

Post by evrydayhustling » Thu, Sep 12, 2024, 6:17 PM UTC

mewpmewp2 wrote: post #244

changoplatanero wrote: post #227

Earlier quoted context omitted.

Hmm, are you sure it was using the o1 model and not gpt4o? I've been using the o1 model and it does consistently well at solving rotation ciphers.

Does it do better than Claude, because Claude (3.5 sonnet) handled ROTs perfectly and was able to also respond in ROT.

Just tried, no joy from Claude either:

Can you decrypt the following? I don't know the cypher, but the plaintext is Spanish.

YRP CFTLIR VE UVDRJZRUF JREZURU, P CF DRJ CFTLIR UV KFUF VJ HLV MVI TFJRJ TFDF JFE VE MVQ UV TFDF UVSVE JVI