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

901–910 of 1001 posts

Re: Learning to Reason with LLMs

#901

Post by strgcmc »

ActorNightly wrote: post #794
crazygringo wrote: post #310

Earlier quoted context omitted.

Seriously. I actually feel as impressed by the chain of thought, as I was when ChatGPT first came out. This isn't "just" autocompletion anymore, this is actual step-by-step reasoning full of ideas and dead ends and refinement, just like humans do when solving problems. Even if it is still ultimately being powered by "autocompletion". But then it makes me wonder about human reasoning, and what if it's similar? Just fo…

Again its not reasoning. Reasoning would imply that it can figure out stuff without being trained in it. The chain of thought is basically just a more accurate way to map input to output. But its still a map, i.e forward only. If an LLM coud reason, you should be able to ask it a question about how to make a bicycle frame from scratch with a small home cnc with limited work area and it should be able to iterate on an…

I think you need to re-calibrate your expectations... I'm not saying this is a solved problem by any means, but I just tried this out with Claude Sonnet 3.5, and these instructions seem quite reasonable and detailed to me (about what I might expect if I spoke to a human expert and they tried to explain the steps to me over the telephone, for example). Does this mean this LLM is able to "reason"? I don't know that I would make THAT bold of a claim, but I think your example is not sufficient to demonstrate something that LLMs are fundamentally incapable of... in other words, the distance between "normal LLM statistical tricks" vs "reasoning" keeps getting smaller and smaller.

---

My base prompt:

> Here is a hypothetical scenario, that I would like your help with: imagine you are trying to help a person create a bicycle frame, using their home workshop which includes a CNC machine, commonly available tools, a reasonable supply of raw metal and hardware, etc. Please provide a written set of instructions, that you would give to this person so that they can complete this task.

First answer: https://claude.site/artifacts/f8af03ba-3f2c-497d-b564-a19baf...

My follow-up, pressing for actual measurements:

> Can you suggest some standard options for bike geometry, assuming an average sized human male?

Answer including specific dimensions: https://claude.site/artifacts/2f5ea2f3-69d8-4a1b-a563-15d334...

Re: Learning to Reason with LLMs

#902

Post by benreesman »

cs391231 wrote: post #408

Student here. Can someone give me one reason why I should continue in software engineering that isn't denial and hopium?

I was debugging an issue the other day where either of sentencepiece or gRPC linked into a C++ program worked fine, but both at once caused a segfault before even getting to main deep in the protobuf initialization stuff in some arena management code and left a fairly mangled stack even pwndbg struggled with legible frames.

It wasn’t trivial that combination was even the culprit.

I’ve been around the block with absl before, so it wasn’t a total nightmare, but it was like, oof, I’m going to do real work this afternoon.

They don’t pay software engineers for the easy stuff, they pay us because it gets a little tricky sometimes.

I’ll reserve judgement on this new one until I try it, but the previous ones, Sonnet and the like, they were no help with something like that.

When StackOverflow took off, and Google before that, there wide swaths of rote stuff that just didn’t count as coding anymore, and LLMs represent sort of another turn of that crank.

I’ve been wrong before, and maybe o1 represents The Moment It Changed, but as of now I feel like a sucker that I ever bought into the “AI is a game changer” narrative.

Re: Learning to Reason with LLMs

#903

Post by thelastparadise »

rpmisms wrote: post #773
sashank_1509 wrote: post #677

Earlier quoted context omitted.

I have access to this and there is no way I spend more than 50$ on OpenAI api. I have ChatGPT + since day q though (240$ probably in total)

You missed your raise key on "day q"

Raise it up just one

Re: Learning to Reason with LLMs

#905

Post by Axsuul »

geenkeuse wrote: post #563

Average Joe's like myself will build our apps end to end with the help of AI. The only shops left standing will be Code Auditors. The solopreneur will wing it, without them, but enterprises will take the (very expensive) hit to stay safe and compliant. Everyone else needs to start making contingency plans. Magnus Carlsen is the best chess player in the world, but he is not arrogant enough to think he can go head to h…

I think this is a common fallacy and an incorrect extrapolation, especially made by those who are unfamiliar with what it takes to build software. Software development is hard because the problems it solves are not well defined, and the systems themselves become increasingly complex with each line of code. I have not seen or experienced LLMs making any progress towards these.

Re: Learning to Reason with LLMs

#906

Post by rchaves »

thomasahle wrote: post #170
islewis wrote: post #124

My first interpretation of this is that it's jazzed-up Chain-Of-Thought. The results look pretty promising, but i'm most interested in this: > 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. Mentioning competitive advantage here signals to me that OpenA…

> my gut reaction is this negatively impacts model usability, but i'm having a hard time putting my finger on why. This will make it harder for things like DSPy to work, which rely using "good" CoT examples as few-shot examples.

yeah I guess base models without built-it CoT are not going away, exactly because you might want to tune it yourself. If DSPy (or similar) evolves to allow the same or similar than OpenAI did with o1, that will be quite powerful, but we still need the big foundational models powering it all

on the other hand, if cementing techniques in the models becomes a trend, we might see various models around with each technique for us to pick and choose beyond CoT without need for us to guide the model ourselves, then what's left for us to optimize is the prompts on what we want, and the routing the combination of those in a nice pipeline

still the principle of DSPy stays the same, have a dataset to evaluate, let the machine trial an error prompts, hyperparameters and so on, just switch around different techniques (possibly automating that too), and get measurable, optimizable results

Re: Learning to Reason with LLMs

#907

Post by lukev »

ramesh31 wrote: post #621
lukev wrote: post #512

This is a pretty big technical achievement, and I am excited to see this type of advancement in the field. However, I am very worried about the utility of this tool given that it (like all LLMs) is still prone to hallucination. Exactly who is it for? If you're enough of an expert to critically judge the output, you're probably just as well off doing the reasoning yourself. If you're not capable of evaluating the outp…

>If you're enough of an expert to critically judge the output, you're probably just as well off doing the reasoning yourself. Thought requires energy. A lot of it. Humans are for more efficient in this regard than LLMs, but then a bicycle is also much more efficient than a race car. I've found that even when they are hilariously wrong about something, simply the directionality of the line of reasoning can be enough t…

Look, I've been experimenting with this for the past year, and this is definitely the happy path.

The unhappy path, which I've also experienced, is that the model outputs something plausible but false but that aligns with an area where my thinking was already confused and sends me down the wrong path.

I've had to calibrate my level of suspicion, and so far using these things more effectively has always been in the direction that more suspicion is better.

There's been a couple times in the last week where I'm working on something complex and I deliberately don't use an LLM since I'm now actively afraid they'll increase my level of confusion.

Re: Learning to Reason with LLMs

#908

Post by natch »

In practice, this implementation (through the Chat UI) is scary bad.

It actively lies about what it is doing.

This is what I am seeing. Proactive, open, deceit.

I can't even begin to think of all the ways this could go wrong, but it gives me a really bad feeling.

Re: Learning to Reason with LLMs

#909

Post by natch »

romeros wrote: post #428
bartman wrote: post #367

This is incredible. In April I used the standard GPT-4 model via ChatGPT to help me reverse engineer the binary bluetooth protocol used by my kitchen fan to integrate it into Home Assistant. It was helpful in a rubber duck way, but could not determine the pattern used to transmit the remaining runtime of the fan in a certain mode. Initial prompt here [0] I pasted the same prompt into o1-preview and o1-mini and both c…

is it better than Claude?

Way worse than Claude for solving a cipher. Not even 1/10th as good. Just one data point, ymmv.

Re: Learning to Reason with LLMs

#910

Post by attentive »

OkGoDoIt wrote: post #522

Some practical notes from digging around in their documentation: In order to get access to this, you need to be on their tier 5 level, which requires $1,000 total paid and 30+ days since first successful payment. Pricing is $15.00 / 1M input tokens and $60.00 / 1M output tokens. Context window is 128k token, max output is 32,768 tokens. There is also a mini version with double the maximum output tokens (65,536 tokens…

So, basically, it's chain of thought as a service?

Not a model, per se, but a service that chains multiple model requests behind the scene?

Post reply on HN