You can scale them up and down at any time, they can work 24/7 (including holidays) with no overtime pay and no breaks, they need no corporate campuses, office space, HR personnel or travel budgets, you don't have to worry about key employees going on sick/maternity leave or taking time off the moment they're needed most, they won't assault a coworker, sue for discrimination or secretly turn out to be a pedophile and tarnish the reputation of your company, they won't leak internal documents to the press or rage quit because of new company policies, they won't even stop working when a pandemic stops most of the world from running.

I agree the most interesting thing to watch will be cost for a given score more than maximum possible score achieved (not that the latter won't be interesting by any means).

So ya, working on efficiency is important, but we're still pretty far away from AGI even ignoring efficiency. We need an actual breakthrough, which I believe will not be possible by simply scaling the transformer architecture.

Then let's say that OpenAI brute forced this without any meta-optimization of the hypothesized search component (they just set a compute budget). This is probably low hanging fruit and another 2x in compute reduction. ($850)

Then let's say that OpenAI was pushing really really hard for the numbers and was willing to burn cash and so didn't bother with serious thought around hardware aware distributed inference. This could be more than a 2x decrease in cost like we've seen deliver 10x reductions in cost via better attention mechanisms, but let's go with 2x for now. ($425).

So I think we've got about an 8x reduction in cost sitting there once Google steps up. This is probably 4-6 months of work flat out if they haven't already started down this path, but with what they've got with deep research, maybe it's sooner?

Then if "all" we get is hardware improvements we're down to what 10-14 years?

I'd hope we see more internal optimizations and improvements to the models. The idea that the big breakthrough being "don't spit out the first thought that pops into your head" seems obvious to everyone outside of the field, but guess what turns out it was a big improvement when the devs decided to add it.

The trend for power consumption of compute (Megaflops per watt) has generally tracked with Koomey’s law for a doubling every 1.57 years

Then you also have model performance improving with compression. For example, Llama 3.1’s 8B outperforming the original Llama 65B

If this turns out to be hard to optimize / improve then there will be a huge economic incentive for efficient ASICs. No freaking way we’ll be running on GPUs for 20-25 years, or even 2.

But sorry, blablabla, this shit is getting embarrassing.

> The question is now, can we close this "to human" gap

You won’t close this gap by throwing more compute at it. Anything in the sphere of creative thinking eludes most people in the history of the planet. People with PhDs in STEM end up working in IT sales not because they are good or capable of learning but because more than half of them can’t do squat shit, despite all that compute and all those algorithms in their brains.

I'm not so sure—what they're doing by just throwing more tokens at it is similar to "solving" the traveling salesman problem by just throwing tons of compute into a breadth first search. Sure, you can get better and better answers the more compute you throw at it (with diminishing returns), but is that really that surprising to anyone who's been following tree of thought models?

All it really seems to tell us is that the type of model that OpenAI has available is capable of solving many of the types of problems that ARC-AGI-PUB has set up given enough compute time. It says nothing about "intelligence" as the concept exists in most people's heads—it just means that a certain very artificial (and intentionally easy for humans) class of problem that wasn't computable is now computable if you're willing to pay an enormous sum to do it. A breakthrough of sorts, sure, but not a surprising one given what we've seen already.

In this case there is more reason to think these things are relevant outside of the direct context - these tests were specifically designed to see if AI can do general-thinking tasks. The benchmarks might be bad, but that's at least their purpose (unlike in Chess).

The other benchmarks are a good indication though.

The low compute was $17 per task. Speculate 172*$17 for the high compute is $2,924 per task, so I am also confused on the $3400 number.

The number for the high-compute one is ~172x the first one according to the article so ~=$2900

4o is cheaper than o1 mini so mini doesn't mean much for costs.

LLMs are far, _far_ below human on elementary problems, once you allow any variation and stop spoonfeeding perfectly phrased word problems. :)

https://machinelearning.apple.com/research/gsm-symbolic

https://arxiv.org/pdf/2410.05229

Paper came out in October, I don't think many have fully absorbed the implications.

It's hard to take any of the claims of "LLMs can do reasoning!" seriously, once you understand that simply changing what names are used in a 8th grade math word problem can have dramatic impact on the accuracy.

I would think this is a not so good bench. Author does not write logically, they write for entertainment.

What I'm saying is the fact that as models are getting better at reasoning they are also scoring better on ARC proves that it is measuring something relating to reasoning. And nobody else has come up with a comparable benchmark that is so easy for humans and so hard for LLMs. Even today, let alone five years ago when ARC was released. ARC was visionary.

While I appreciate the benchmark and its goals (not to mention the puzzles - I quite enjoy figuring them out), successfully passing this benchmark does not demonstrate or guarantee real world capabilities or performance. This is why I increasingly side-eye this field and its obsession with constantly passing benchmarks and then moving the goal posts to a newer, harder benchmark that claims to be a better simulation of human capabilities than the last one: it reeks of squandered capital and a lack of a viable/profitable product, at least to my sniff test. Rather than simply capitalize on their actual accomplishments (which LLMs are - natural language interaction is huge!), they're trying to prove to Capital that with a few (hundred) billion more in investments, they can make AGI out of this and replace all those expensive humans.

They've built the most advanced prediction engines ever conceived, and insist they're best used to replace labor. I'm not sure how they reached that conclusion, but considering even their own models refute this use case for LLMs, I doubt their execution ability on that lofty promise.

If the hypothesis is that LLMs are the “computer” that drives the AGI then of course the benchmark is relevant in testing for AGI.

I don’t think you understand the benchmark and its motivation. ARC AGI benchmark problems are extremely easy and simple for humans. But LLMs fail spectacularly at them. Why they fail is irrelevant, the fact they fail though means that we don’t have AGI.

https://youtu.be/SKBG1sqdyIU?t=4m40s

    | Name                                 | Semi-private eval | Public eval |
    |--------------------------------------|-------------------|-------------|
    | Jeremy Berman                        | 53.6%             | 58.5%       |
    | Akyürek et al.                       | 47.5%             | 62.8%       |
    | Ryan Greenblatt                      | 43%               | 42%         |
    | OpenAI o1-preview (pass@1)           | 18%               | 21%         |
    | Anthropic Claude 3.5 Sonnet (pass@1) | 14%               | 21%         |
    | OpenAI GPT-4o (pass@1)               | 5%                | 9%          |
    | Google Gemini 1.5 (pass@1)           | 4.5%              | 8%          |

  o3 (coming soon)  75.7% 82.8%
  o1-preview        18%   21%
  Claude 3.5 Sonnet 14%   21%
  GPT-4o            5%    9%
  Gemini 1.5        4.5%  8%

[1]: https://arcprize.org/2024-results

The only effect smarter models will have is that intelligent people will have to use less of their brain to do their work. As has always been the case, the medium is the message, and climate change is one of the most difficult and worst problems of our time.

If this gets software people to quit en-masse and start working in energy, biology, ecology and preservation? Then it has succeeded.

The Average human is a lot dumber than people on hackernews and reddit seem to realize, shit the people on mturk are likely smarter than the AVERAGE person

There are blind spots, doesn't take away from 'general'.

In other words, it's possible humans can reason better than o3, but cannot articulate that reasoning as well through text - only in our heads, or through some alternative medium.

From a post elsewhere the scores on ARC-AGI-PUB are approx average human 64%, o3 87%. https://news.ycombinator.com/item?id=42474659

Though also elsewhere, o3 seems very expensive to operate. You could probably hire a PhD researcher for cheaper.

How does a giant pile of linear algebra not meet that definition?

At most you can argue that there isn't a useful bounded loss on every possible input, but it turns out that humans don't achieve useful bounded loss on identifying arbitrary sets of pixels as a cat or whatever, either. Most problems NNs are aimed at are qualitative or probabilistic where provable bounds are less useful than Nth-percentile performance on real-world data.

But, to my mind, something of the form "Train a neural network with an architecture generally like [blah], with a training method+data like [bleh], and save the result. Then, when inputs are received, run them through the NN in such-and-such way." would constitute an algorithm.

When a NN is trained, it produces a set of parameters that basically define an algorithm to do inference with: it's a very big one though.

We also call that a NN (the joy of natural language).

When someone is "disinterested enough" to publish though, note the obvious way to launch a new fund or advisor with a good track record: crank out a pile of them, run them one or two years, discard the many losers and publish the one or two top winners. I.E. first you should be suspicious of why it's being published, then of how selected that result is.

Their post has stem grad at nearly 100%

I guess you can say the same thing for the Turing Test. Simple chat bots beat it ages ago in specific settings, but the bar is much higher now that the average person is familiar with their limitations.

If/once we have an AGI, it will probably take weeks to months to really convince ourselves that it is one.

Also, it depends a great deal on what we define as AGI and whether they need to be a strict superset of typical human intelligence. o3's intelligence is probably superhuman in some aspects but inferior in others. We can find many humans who exhibit such tendencies as well. We'd probably say they think differently but would still call them generally intelligent.

ChatGPT has had for some time the feature of storing memories about its conversations with users. And you can use function calling to make this more generic.

I think drawing the boundary at “model + scaffolding” is more interesting.

For one, I speculate OpenAI is using a very basic agent harness to get the results they've published on SWEBench. I believe there is a fair amount of headroom to improve results above what they published, using the same models.

For two, some of the instances, even in SWEBench-Verified, require a bit of "going above and beyond" to get right. One example is an instance where the user states that a TypeError isn't properly handled. The developer who fixed it handled the TypeError but also handled a ValueError, and the golden test checks for both. I don't know how many instances fall in this category, but I suspect its more than on a simpler benchmark like MATH.

Regardless the critical aspect is valid, AGI would be something like Cortana from Halo.

If you disagree with me, state why instead of opting to downvote me

That said, this is all predicated on o3 or similar actually having achieved human level reasoning. That's yet to be fully proven. We'll see!

Your secrecy comment is really intriguing actually. And morbid lol.

Yeah, that makes this result a lot less impressive for me.

> An acid test is a qualitative chemical or metallurgical assay utilizing acid. Historically, it often involved the use of a robust acid to distinguish gold from base metals. Figuratively, the term represents any definitive test for attributes, such as gauging a person's character or evaluating a product's performance.

Specifically here, they're using the figurative sense of "definitive test".