I doubt that anything resembling genuine AGI is within reach of current AI tools
mathstodon.xyz
mathstodon.xyz
It’s the result that consumers are interested in, not the mechanics of how it’s achieved. Software engineers are often extraordinarily bad at seeing the difference because they’re so interested in the implementation details.
A machine that magically replaces several hours of her manual work? As far as she’s concerned, it’s a specialized maid that doesn’t eat at her table and never gets sick.
I used to play with a Maytag machine machine motor. It had a single cylinder, ran on gasoline, and had a kick-start. It was from, IIRC, 1926.
The exhaust would have been plumbed to the outdoors, but other than that the expectation was that there would be a gas-fired engine running in the house while the washing was done.
https://www.nfpa.org/education-and-research/research/nfpa-re...
In the table from the Pdf link failure to clean was the only category that resulted in deaths.
In both cases, automation of what was previously human labor is very early and they’ve seen almost nothing yet.
I agree that in the year 2225 people are not going to consider basic LLMs artificial intelligences, just like we don’t consider a washing machine a maid replacement anymore.
They were not called maids nor personified.
I could even see the humour in "washer-bot" and "dryer-bot" if they did anything notably more complex. But we don't need/want appliances to become more complex than is necessary. We usually just call such things programmable.
I can accept calling our new, over-hyped, hallucinating overlords chatbots. But to be fair to the technology, it is we chatty humans doing all the hyping and hallucinating.
The market capitalisation for this sector is sickly feverish — all we have done is to have built a significantly better ELIZA [1]. Not a HIGGINS and certainly not AGI. If this results in the construction of new nuclear power facilities, maybe we can do the latter with significant improvement too. (I hope.)
My toaster and oven will never be bots to me. Although my current vehicle is better than earlier generations, it contains plenty of bad code and it spews telemetry. It should not be trusted with any important task.
Aside from the obviously humorous content the rest is useless allegory (I want a recipe not a story and need some code, not personal affection for software engineering) and no true scotsman (no true adherent of my native language would call it a robotic maid!)
As social creatures humans are pretty repetitive.
Washing is a useful word to describe what that machine does. Our current setup is like if washing machines were called "badness removers," and there was a widespread belief that we were only a few years out from a new model of washing machine being able to cure diseases.
Given that, I consider it quite possible that we'll reach a point where even more people will consider LLMs having reached or surpassed AGI, while others still only consider it "sufficiently advanced autocomplete".
I buy that there's disagreement on what intelligence means in the enthusiast space, but "thinks like people" is pretty clearly the general understanding of the word, and the one that tech companies are hoping to leverage.
Doubly so if the AGI writes software for itself to accomplish a task it decided to do.
Once someone has software like that, not a dog that is sicced on a task, but a bloodhound that seeks out novelty and accomplishment for its own personal curiosity or to test its capabilities, then you have a good chance of convincing me that AGI has been achieved.
Until then, we have fancy autocomplete.
I don't understand this mindset that because someone stuck the label "AI" on it, consumers are suddenly unable to think for themselves. AI as a marketing label has been used for dozens of years, yet only now is it taking off like crazy. The word hasn't change - what it's actually capable of doing has.
Yikes. I’m guessing you’ve never lost anyone to “alternative” medical treatments.
LLMS can appear intelligent until they, often, say things no intelligent being would. Then they appear profoundly stupid.
Washing machines wash reliably. LLMs do not.
A machine will be intelligent when instead of producing false output it responds with “I don’t know” and can be trusted.
AI (supervised).
But more seriously, this is ELIZA with network effects. Credulous multitudes chatting with a system that they believe is sentient.
The conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.
The only AI explainer youll need: https://kemendo.com/Understand-AI.html
- You would prompt "Ok AGI, read through the last 26978894356 research papers on cancer and tell me what are some unexplored angles" and it would tell you
- You would prompt "Show me the last 10 emails on Sam Altman's inbox" and it would actually show you
- You would prompt "Give me a list of people who have murdered someone in the USA and havent been caught yet" and it would give you a list of suspects that fit the profile
You really dont want AGI
ANI Artificial Narrow Intelligence
AGI Artificial General Intelligence
ASI Artificial Super Intelligence
https://aeon.co/essays/generative-ai-has-access-to-a-small-s...
I guess ultimately what is intelligence? We compact our memories, forget things, and try repeatedly. Our inputs are a bit more diverse but ultimately we autocomplete our lives. Hmm… maybe we’ve already achieved this.
The only question remaining is what is the end point of AGI capability.
What’s the final IQ we’ll hit, and more importantly why will it end there?
Power limits? Hardware bandwidth limit? Storage limits? the AI creation math scales to infinity so that’s not an issue.
Source data limits? Most likely. We should have recorded more. We should have recorded more.
No. You are misrepresenting the test's purpose, the argument made around it and the results people have gotten. Turing was explicit that the question was ill-posed in the first place, and proposed a test of useful capability. But even then, hypothetical imagining of what a "passing" agent's responses might look like, was radically different from what we get today. And the supposed "passes" we've seen recently are highly suspect.
Do current AI tools genuinely pose such risks?
For starters, I think we can rightly ask what it means to say "genuine artificial general intelligence", as opposed to just "artificial general intelligence". Actually, I think it's fair to ask what "genuine artificial" $ANYTHING would be.
I suspect that what he means is something like "artificial intelligence, but that works just like human intelligence". Something like that seems to be what a lot of people are saying when they talk about AI and make claims like "that's not real AI". But for myself, I reject the notion that we need "genuine artificial general intelligence" that works like human intelligence in order to say we have artificial general intelligence. Human intelligence is a nice existence proof that some sort of "general intelligence" is possible, and a nice example to model after, but the marquee sign does say artificial at the end of the day.
Beyond that... I know, I know - it's the oldest cliche in the world, but I will fall back on it because it's still valid, no matter how trite. We don't say "airplanes don't really fly" because they don't use the exact same mechanism as birds. And I don't see any reason to say that an AI system isn't "really intelligent" if it doesn't use the same mechanism as human.
Now maybe I'm wrong and Terry meant something altogether different, and all of this is moot. But it felt worth writing this out, because I feel like a lot of commenters on this subject engage in a line of thinking like what is described above, and I think it's a poor way of viewing the issue no matter who is doing it.
I think he means "something that can discover new areas of mathematics".
That does seem awfully specific though, in the context of talking about "general" intelligence. But I suppose it could rightly be argued that any intelligence capable of "discovering new areas of mathematics" would inherently need to be fairly general.
It's one of a large set of attributes you would expect in something called "AGI."
Superintelligence is smarter than Terrence Tao, or any other human.
SuperIntelligence (or ASI), OTOH, has - so far as I can recall - always been even more loosely specified, and translates roughly to "an intelligence beyond any human intelligence".
Another term you might hear, although not as frequently, is "Universal Artificial Intelligence". This comes mostly from the work of Marcus Hutter[1] and means something approximately like "an intelligence that can solve any problem that can, in principle, be solved".
How many software engineers with a good math education can do this?
So in Tao’s statement I interpret “genuine” not as an adverb modifying the “artificial” adjective but as an attributive adjective modifying the noun “intelligence”, describing its quality… “genuine intelligence that is non-biological in nature”
That's definitely possible. But it seems redundant to phrase it that way. That is to say, the goal (the end goal anyway) of the AI enterprise has always been, at least as I've always understood it, to make "genuine intelligence that is non-biological in nature". That said, Terry is a mathematician, not an "AI person" so maybe it makes more sense when you look at it from that perspective. I've been immersed in AI stuff for 35+ years, so I may have developed a bit of myopia in some regards.
The point above is valid. I'd like to deconstruct the concept of intelligence even more. What humans are able to do is a relatively artificial collection of skills a physical and social organism needs. The so highly valued intelligence around math etc. is a corner case of those abilities.
There's no reason to think that human mathematical intelligence is unique by its structure, an isolated well-defined skill. Artificial systems are likely to be able to do much more, maybe not exactly the same peak ability, but adjacent ones, many of which will be superhuman and augmentative to what humans do. This will likely include "new math" in some sense too.
The problem and what most people intuitively understand is that this compression is not enough. There is something more going on because people can come up with novel ideas/solutions and whats more important they can judge and figure out if the solution will work. So even if the core of the idea is “compressed” or “mixed” from past knowledge there is some other process going on that leads to the important part of invention-progress.
That is why people hate the term AI because it is just partial capability of “inteligence” or it might even be complete illusion of inteligence that is nowhere close what people would expect.
What about reinforcement learning? RL models don't train on an existing dataset, they try their own solutions and learn from feedback.
RL models can definitely "invent" new things. Here's an example where they design novel molecules that bind with a protein: https://academic.oup.com/bioinformatics/article/39/4/btad157...
Counterpoint: ChatGPT came up with the new idiom "The confetti has left the cannon"
Useful = great. We've made incredible progress in the past 3-5 years.
The people who are disappointed have their standards and expectations set at "science fiction".
From what I've seen, in response to that, goalposts are then often moved in the way that requires least updating of somebody's political, societal, metaphysical etc. worldview. (This also includes updates in favor of "this will definitely achieve AGI soon", fwiw.)
That's certainly not coming back.
It's not a real thing. You do not remember the goal posts ever being there.
Turing put forth a thought experiment in the early days of some discussions about "artificial" thinking machines on a very philisophical level.
Add to that, nobody who claims to have "passed" the turing test has ever done an actual example of that thought experiment, which is about taking two respondents and finding out which is human. It is NOT talking to a single respondent and deciding whether they are an LLM or not.
It also has never been considered a valid "test" of "intelligence" as it was obvious from the very very beginning that tricking a person wasn't really meaningful, as most people can be tricked by even simple systems.
ELIZA was the end of any thought around "The turing test", as it was able to "trick" tons of people and show how useless the turing thought experience was. Anyone who claims ELIZA is intelligent would be very silly.
I don't think it's fair to deride people who are disappointed in LLMs for not being AGI when many very prominent proponents have been claiming they are or soon will be exactly that.
But now, we have LLMs that can reliably beat video games like Pokemon, without any specialized training for playing video games. And those same LLMs can write code, do math, write poetry, be language tutors, find optimal flight routes from one city to another during the busy Christmas season, etc.
How does that not fit the definition of "General Intelligence"? It's literally as capable as a high school student for almost any general task you throw it at.
I'm not sure the party that "they" is referring to here, since arc-agi-3 dataset isn't released yet and labs probably have not begun targeting it. For arc-agi-2, possibly just synthetic data might have been enough to saturate the benchmark, since most frontier models do well on it yet we haven't seen any corresponding jump in multimodal skill use, with maybe the exception of "nano banana".
>lend itself well to token based “reasoning”
One could perhaps do reasoning/COT with vision tokens instead of just text tokens. Or reasoning in latent space which I guess might be even better. There have been papers on both, but I don't know if it's an approach that scales. Regardless gemini 3 / nano banana have had big gains on visual and spatial reasoning, so they must have done something to get multimodality with cross-domain transfer in a way that 4o/gpt-image wasn't able to.
For arc-agi-3, the missing pieces seem to be both "temporal reasoning" and efficient in-context learning. If they can train for this, it'd have benefits for things like tool-calling as well, which is why it's an exciting benchmark.
No; that was one, extremely limited example of a broader idea. If I point out that your machine is not a general calculator because it gives the wrong answer for six times nine, and then you fix the result it gives in that case, you have not refuted me. If I now find that the answer is incorrect in some other case, I am not "moving goalposts" by pointing it out.
(But also, what lxgr said.)
> But now, we have LLMs that can reliably beat video games like Pokemon, without any specialized training for playing video games. And those same LLMs can write code, do math, write poetry, be language tutors, find optimal flight routes from one city to another during the busy Christmas season, etc.
The AI systems that do most of these things are not "LLMs".
> It's literally as capable as a high school student for almost any general task you throw it at.
And yet embarrassing deficiencies are found all the time ("how many r's in strawberry", getting duped by straightforward problems dressed up to resemble classic riddles but without the actual gotcha, etc.).
Uh, every single example that I listed except for the 'playing video games' example is something that I regularly use frontier models to do for myself. I have ChatGPT and Gemini help me find flight routes, tutor me in Spanish (Gemini 3 is really good at this), write poetry and code, solve professional math problems (usually related to finance and trading), help me fix technical issues with my phone and laptop, etc etc.
If you say to yourself, "hey this thing is a general intelligence, I should try to throw it at problems I have generally", you'll find yourself astonished at the range of tasks with which it can outperform you.
LLMs are at most one component of the systems you refer to. Reasoning models and agents are something larger.
> If you say to yourself, "hey this thing is a general intelligence, I should try to throw it at problems I have generally", you'll find yourself astonished at the range of tasks with which it can outperform you.
Where AI has been thrust at me (search engines and YouTube video and chat summaries) it has been for the sort of thing where I'd expect it to excel, yet I've been underwhelmed. The one time I consciously invoked the "AI assist" on a search query (to do the sort of thing I might otherwise try on Wolfram Alpha) it committed a basic logical error. The project READMEs that Show HN has exposed me to this year have been almost unfailingly abominable. (Curiously, I'm actually okay with AI art a significant amount of the time.)
But none of that experience is even a hundredth as annoying as the constant insinuation from AI proponents that any and all opposition is in some way motivated by ego protection.
Okay, enough eggnog and posting.
It also seems orders of magnitude less resource efficient than higher-level approaches.