Apple AI boffins puncture AGI hype as reasoning models flail on complex planning If you are betting on AGI – artificial general intelligence, the point at which AI models rival human cognition – showing up next year, you may want to adjust your timeline. Apple AI researchers have found that the "thinking" ability of so-called "large reasoning models" collapses when things get complicated. The authors' …
But they naysayers will say loudly that
THIS IS APPLE who can't get their Apple Intelligence to work (and SIRI after all these years)
so the sensible original statement will get ignored.
I hope this "AI with everything" hype goes away soon. But, it is 'hip', 'cool', 'with it' etc etc so it won't.
I don't think it's hip or cool, I think it is a desperate last resort for Venture Capitalists who have been throwing money at the blockchain, the metaverse and anything else they imagine might offer exponential profit growth over the last ten years and not seen much return.
At this point they're all-in on AI and although it is less of an obvious bubble than NFTS were - there probably is a baby in this bathwater somewhere - there is no way it can ever be as profitable as it would need to be to justify the investment. They have to keep hyping it because papers like this one are going to keep coming and if they don't get out in time they're going to get their fingers burnt in at least a 2000 way even if we don't go full 2008.
Expect to see more nonsensical hype and more mechanical turks as the road ahead gets bumpier, and in the meantime hope your pension isn't too heavily invested in AI.
We don't really know what "intelligence" is, but, ultimately, there are only two possibilities. Either it is nothing more than emergent behavior from big statistics with a lot of recursion, or it is more than that.
In the first case, we might get AGI comparatively soon. It might take some new development, such as embodied AI, or it might take finding just the right architecture. But the path would be roughly correct and ought to eventually get there.
In the second case, we are not getting AGI any time soon, and all the effort spent on LLMs can only run in circles while getting nowhere. If we're very lucky, we might learn why it's not the answer.
Both possibilities are philosophically interesting, and I'm looking forward to the outcome.
Being related to someone who has been studying AI for decades, I know for certain that what is currently called "AI" isn't, and never will be, intelligent. While it does a decent job of mimicking a human response, it doesn't really learn, and definitely cannot "understand" anything. It's glorified pattern-matching, an advanced version of Eliza that's been fed a lot of input.
No, real AI isn't simple to do, we don't have it, and we likely won't anytime in our lifetimes.
I think there's a third possibility: that intelligence is a behaviour that emerges from statistics, etc... but that LLMs, while a step towards it, aren't a suitable architecture and so nothing close to AGI will arise "comparatively soon"; it will need another architectural break through.
And, frankly, LLMs are enough to be getting on with. If their "intelligence" is capped that will give us time to adjust before anything close to AGI emerges..
'We don't really know what "intelligence" is'
When I first heard the hype about LLM's, I thought it would probably challenge peoples idea of what intelligence is.
My take is that intelligence has evolved over *many* generations of trial and error for the purpose of operating the creatures that possess it, such that they may feed themselves and reproduce in the face of competition for common resources. Artificial intelligence has no purpose, so what's the impetus for it to be intelligent? The fact that so many supposedly intelligent people think that LLM's are intelligent leads me to doubt the capabilities of real human intelligence, which did not evolve for the modern world.
"such that they may feed themselves and reproduce in the face of competition for common resources"
It's a bit more than that. Reading and commenting here achieves neither of those goals. So either that is not intelligence at work (a plausibly arguable PoV) or intelligence acquires a whole lot of additional goal for its own pleasure.
We evolved intelligence in order to communicate complex ideas to work together in groups, make and use tools, gain knowledge and experience, plan ahead, understand cause and effect etc. It worked for cavemen, and it still works for us today. Reading and commenting here is exercising your intelligence because you don't need to use it to plan your next crop planting or mammoth hunt, and doing nothing with it is dull.
brainwrong,
We didn't evolve intelligence to achieve any kind of a goal. There wasn't some committee in a cave somewhere who went on UggBay to order up a crate of intelligence pills.
We evolved intelligence, and it enabled us to successfully compete with other species and eventually invent digital watches.
Wheresas dolphins evolved intelligence in order to lark about in the water having a good time, and eating a lot of fish.
Or even, possibly not. Maybe the Neanderthals were more intelligent than us, but got outcompeted anyway? Who knows. There's a lot of luck going on in evolution as well, so I suspect its foolish to try to come up with any definitive conclusions. Unless the alien Xenu gave us intelligence, in order to build him (her/it?) better pyramids?
The site lesswrong.com has some excellent essays and dialogues about the nature of thought and intelligence as it relates to LLM/AI. It’s well moderated and thoughtful in a way that’s hard to find online anymore.
There are two broad camps. One sees thought and intelligence as purely technical and subject to modeling. The other sees them as more along the lines you’re touching on.
At the risk of sounding old, AGI is incapable of synergistic reasoning. It may recognize that candies on a birthday cake is an important tradition, but doesn’t recognize why a single candle in a croissant is equally valid or why you don’t put 70 candles on the cake.
I’m in broad agreement with your take on it. AI has no inherent purpose, thus cannot reason unreasonably like humans can. I reckon it’ll collapse in on itself or kill us all.
There's something odd about us. Stuff goes on in our head that is unique on this planet. The level of self-awareness is extreme in many of us and the ability of some to make phenomenal leaps in imagination, abstraction and discovery. We build complex models in our head. On the other hand I have come to the conclusion that the majority just think what the TV tells them to think; NPCs?
Intelligence is the probabilities vs real world experience. The probabilities give you a range of outcomes, and real world experience tells you what the most likely outcome is. Ai can do probabilities, but needs your real world experience to put it to good use.
Quantum mechanics is also a game of probabilities. You have to make a real world measurement to verify your best guess. In the case of the internet there are no more places of truth left to verify against.
You are playing a game of Schrödinger's cat without any measurements taken therefor, you cannot trust the AI to be truthful, or even accurate.
The current problem with AI and/or AGI is not one of architecture or understanding what intelligence really is !!!
The problem is one of 'Truth' ...
The people who are working on creating 'AI' are being 'economical' with the 'Truth'.
'AI' can masquerade as 'Intelligent' BUT it is not 'TRUE' by any realistic measure.
It may help to 'Sell' whatever 'AI' is BUT it is NOT truely 'Intelligent' and never can be ... there is something missing.
Sell 'AI' as a flawed agent to assist in knowledge search BUT don't hide the fact that it still 'Lies' [Hallucinations] and unless the knowledge-set it is based on is carefully curated, you still need to take the answers given with a LARGE pinch of salt !!!
Greed is driving the current 'AI' mania ... eventually the funds to keep going WILL run out !!!
A new and different approach is needed ... current 'AI' is a dead-end !!!
That is the 'Truth' !!!
:)
Human lie and helluciate all the time. That has nothing to do with whether these models are intelligent or not.
Lies and helluciations are a problems because they are for profit companies working towards commercial applications where companies don't want their product lying to customers, unless it a company producing sex robots, then you know they will need them to lie.
I love that the boffins developed their own independent benchmark (as in their previous GSM-Symbolic benchmark) to prevent "benchmark data contamination". Too many AI benchmarks are affected by this sort of contamination, like the o1 Pro YT eval on the William Lowell Putnam Mathematics Competition.
The European Commission's Joint Research Center paper (linked under "aren't all that great to begin with") nicely notes for example that (from Narayanan and Kapoor [2023a]) GPT4 "could regularly solve benchmark problems classified as easy - as long as the problems had been added before 5th September 2021. For problems added later, GPT4 could not get a single question right, suggesting that the model had memorised questions and answers".
Similarly, using the "Black Box test" method, Microsoft folks found that 7 multilingual benchmarks had been swallowed up whole by 7 LLMs (cf. Contamination Report for Multilingual Benchmarks) and another group of boffins Investigating Data Contamination in Modern Benchmarks for LLMs used "Testset Slot Guessing" to find LLMs outputting verbatim the missing (right or wrong) multiple-choice options in MMLU benchmark data more than 50% of the time. Louis Hunt posted about similar contamination, with GSM8K data, on his linkedin page.
Very nice article!
The utterances of Apple AI Boffins reminded me of the astronomers in Galileo's day who were well aware that the Ptolemaic Earth centred cosmology that had the imprimatur of the Church wasn't exactly supported by observation. They would have been understandably circumspect in advertising this knowledge not wanting to become a kebab/giro at the next Bruno barbeque.
Equally if you were an AI Boffins you would hardly undermine the tracks your particular gravy train runs on.
Much of the deception of contemporary AI flows from the ineffability of human intelligence, thought and consciousness etc. We (some of us anyway) can experience these (and more) while being able to state some the properties we understand to be essential to each we really cannot provide anything like a comprehensive definition of intelligence etc or even whether one is an illusion or byproduct created by the others. Philosophers still get to dine out on these inponderables - that is if they can get their cutlery sorted.
Anyone who believes that the ability to apply a consistent system deductive logic to reasoning is a prerequisite for intelligence neglects the observation that the vast majority of people of perfectly normal intelligence would fail using that criterion.
Last century there were enough daft AI researchers that were convinced you could create intelligence using Prolog (or another implementation of causal form logic) and expert systems. Expert systems were useful and are routinely used today although often with a little AI badge engineering.
I recently completed an "AI" online form for which the applicant had to fullfill an number of quite specific criteria - a task which screams "decision tree" - this web form clearly had a last century expert system behind it driving the JavaScript/CSS - worked efficiently and effectively but in no sense was intelligent.
AI/LLM will eventually find its niche but will be nowhere as pervasive as would be required, given the resources so far invested, to break even let alone make a profit. People still grow tulips but aren't generally billionaires. ;)
It took experts to realise this? Anyone using "AI" can see they are not actually intelligent and fail more than succeed in complex analytical tasks. We can argue why but the outcome is self evident. It may change though. Maybe it's because humans form generalised networks on a lot of data that we manage to distil down into something useful generally. But it did take quite a few evolutionary generations of brain development going back hundreds of millions of years. AI doesn't appear to learn on the job either. I can't discern much improvement when working on a project from the AI, I find I have to adapt to it and improve my prompts.
is a program like Reason! (the program developed by Gordon Way and company in Dirk Gently's Holistic Detective Agency). You give it input facts, and the conclusion you want to reach. It then creates a series of plausible-sounding steps to support your conclusion. I mean: who wants an AI program that says your conclusion is wrong? I am sure LLMs could be up to this task already.
I'll be going. The one with "The Long Dark Teatime of the Soul" in the pocket
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
