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

Many in the AI field think the bigger-is-better approach is running out of road

economist.com

Post reply on HN
Print view View original

81–90 of 354 posts

Re: Many in the AI field think the bigger-is-better approach is running out of road

#81

Post by YetAnotherNick »

michaelteter wrote: post #63

Isn't the fundamental problem that LLM's don't actually understand anything (as greater concepts), but rather operate as complex probability machines? My 2 month active experience with ChatGPT-4 gave me the following takeaways: - when it's right, it's amazing; and when you, the operator, can recognize the niche use case where it performs really well, it can be a game-changer (although you could have programmed a tool…

I have very similar observation, and while it is amazing at times at helping me with complex tasks, it just isn't like any human. If we rank the tasks by difficulty for human and AI, it just has very small correlation. Also GPT 4 is idiot in conversation, compared to its problem solving skill. I would have expected completely opposite trajectory for AI. eg it's hard for it to get it to ask good clarifying questions. There are lot of cheap tricks in prompting like asking it to act as expert, or prompt for chain of though that works.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#82

Post by RossBencina »

michaelteter wrote: post #63

Isn't the fundamental problem that LLM's don't actually understand anything (as greater concepts), but rather operate as complex probability machines? My 2 month active experience with ChatGPT-4 gave me the following takeaways: - when it's right, it's amazing; and when you, the operator, can recognize the niche use case where it performs really well, it can be a game-changer (although you could have programmed a tool…

I like your takeaways and reflection, especially the "changes the game" idea. There is an analogy with pocket calculators and mental arithmetic. Personally I'm more comfortable reaching for the pocket calculator than offloading all thinking to an LLM. On the other hand, it's not so long ago that manual calculation was a specialized occupation. I could maybe see software coding becoming automated just as calculation was -- except for the fact that calculations are much easier to specify than software.

I don't quite get where you're coming from with "LLM's don't actually understand anything (as greater concepts)". I have heard the view coming from researchers that the larger models do form representations of higher-order information structures ("concepts"). Perhaps what you're getting at is that current models don't encode enough higher-order structure to deal accurately with your domain? Whether the models can be made to do so seems like an open question to me. The boosters say it will be here by next year.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#83

Post by croes »

letitgo12345 wrote: post #67
michaelteter wrote: post #63

Isn't the fundamental problem that LLM's don't actually understand anything (as greater concepts), but rather operate as complex probability machines? My 2 month active experience with ChatGPT-4 gave me the following takeaways: - when it's right, it's amazing; and when you, the operator, can recognize the niche use case where it performs really well, it can be a game-changer (although you could have programmed a tool…

The next iteration will be trained on your own data where "when it's a little wrong, you (the expert) can fix the issue and move on without friction" so that case will become "when it's right" and some amount of "when it's any amount of wrong" cases will become "when it's a little wrong". A few more cycles of this and we could be looking at GPT-10 which is a complete replacement for most tasks.

Better result from less data?

I doubt that.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#84

Post by YetAnotherNick »

bluecoconut wrote: post #12

Another recent (but not called out in this article) is the "Textbooks Are All You Need" paper [1]; the results seem to suggest that careful curation and curriculums of training data can significantly improve model capabilities (when training domain specific, smaller models). Claiming a 10x smaller model can outperform competitors. (Eg. phi-1 vs. starcoder) [1] https://arxiv.org/abs/2306.11644

TBH, it looks like metric manipulation to me. They have used GPT-3.5 to generate their data(and not use textbooks at all like the title suggests). And their dataset is very much like their benchmark data. While there was some filtering, but still it is very possible that lot of the benchmark questions were in training data.

We likely wouldn't ever know how good the model is as it not only closed but they haven't provided access to anyone.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#85

Post by croes »

Animats wrote: post #11

We need a way to make tight little specialist models that don't hallucinate and reliably report when they don't know. Trying to cram all of the web into a LLM is a dead end.

They never know.

What we call hallucination is just when the resulting text is wrong but the underlying probabilities could be high.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#86

Post by croes »

rco8786 wrote: post #19
Animats wrote: post #11

We need a way to make tight little specialist models that don't hallucinate and reliably report when they don't know. Trying to cram all of the web into a LLM is a dead end.

Bingo. I've been beating this drum since the initial GPT-3 awe.. The future of AI is bespoke, purpose-driven models trained on a combination of public and (importantly) proprietary data. Data is still king.

This would only work if thinks like GPT were actually intelligent.

Train on dataset A to learn to think, use thinking on dataset B to become an export in B's field.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#87

Post by usrbinbash »

Animats wrote: post #11

We need a way to make tight little specialist models that don't hallucinate and reliably report when they don't know. Trying to cram all of the web into a LLM is a dead end.

> and reliably report when they don't know.

Then we need a new system, because LMs, no matter if they are large or not, cannot do that, for a very simple reason:

A LM doesn't understand "truthfulness". It has no concept of a sequence being true or not, only of a sequence being probable.

And that probability cannot work as a standin for truthfulness, because the LM doesn't produce improbable sequences to begin with...it's output will always be the most (within heat settings) probable sequence. The LM simply has no way of knowing whether the sequence it just predicted is grounded in reality or not.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#88

Post by usrbinbash »

nomel wrote: post #49
Animats wrote: post #11

We need a way to make tight little specialist models that don't hallucinate and reliably report when they don't know. Trying to cram all of the web into a LLM is a dead end.

> that don't hallucinate “Hallucination” is part of thought. Solving a new problem requires hallucinating new, non existing, possible outcomes and solutions, to find one that will work. It seems that eliminating the ability to interpolate and extrapolate (hallucinations) would make intelligence impossible. It would eliminate creativity, tying together new concepts, creation, etc. Is the goal AI, or a nice database fr…

The problem is, what we call "hallucinating" in LMs isn't a way of creative thinking and coming up with novel solutions. It also has nothing to do with "interpolate and extrapolate".

It's simply when the predicted probable sequence isn't grounded in reality.

When I ask an LLM to summarize the great water wars of 1999, and how the Trade Union was ultimately defeated by the Antarctic Coalitions hovercraft-fleet under Vice Admiral Zagalow, it isn't "extrapolating" from knowledge of history, it is simply inventing a load of bollocks. But that bollocks will be dressed in fine language and probably mixed in with plausible-sounding references that have a somewhat-logical-sounding relation to the training data.

The problem is, the LM doesn't and cannot know when it produces bollocks.

All it can care about is if the sequences produced are probable according to it's model.

Re: Many in the AI field think the bigger-is-better approach is running out of road

#89

Post by fooker »

danielvaughn wrote: post #64
rco8786 wrote: post #19

Earlier quoted context omitted.

Bingo. I've been beating this drum since the initial GPT-3 awe.. The future of AI is bespoke, purpose-driven models trained on a combination of public and (importantly) proprietary data. Data is still king.

I’d be really interested to see an AI model built off of Sci Hub data.

Facebook announced this a month or so before chatgpt took off, and got lambasted everywhere because of hallucinations.
Post reply on HN