ML Newsnew | past | best | rss | submit about
What the F*ck Is Artificial General Intelligence? (arxiv.org)
59 points by SweetSoftPillow on Sep 29, 2025 | hide | past | pdf | 45 comments on HN

In plain words: An overview essay sorts through competing definitions of general intelligence, settling on intelligence as the ability to adapt and AGI as an artificial scientist, then groups the tools and strategies used to build it. It finds that simply scaling up pattern-matching systems leads today, but real AGI will need a mix, with data and energy now the main limits.

Abstract

Artificial general intelligence (AGI) is an established field of research. Yet some have questioned if the term still has meaning. AGI has been subject to so much hype and speculation it has become something of a Rorschach test. Melanie Mitchell argues the debate will only be settled through long term, scientific investigation. To that end here is a short, accessible and provocative overview of AGI. I compare definitions of intelligence, settling on intelligence in terms of adaptation and AGI as an artificial scientist. Taking my cue from Sutton's Bitter Lesson I describe two foundational tools used to build adaptive systems: search and approximation. I compare pros, cons, hybrids and architectures like o3, AlphaGo, AERA, NARS and Hyperon. I then discuss overall meta-approaches to making systems behave more intelligently. I divide them into scale-maxing, simp-maxing, w-maxing based on the Bitter Lesson, Ockham's and Bennett's Razors. These maximise resources, simplicity of form, and the weakness of constraints on functionality. I discuss examples including AIXI, the free energy principle and The Embiggening of language models. I conclude that though scale-maxed approximation dominates, AGI will be a fusion of tools and meta-approaches. The Embiggening was enabled by improvements in hardware. Now the bottlenecks are sample and energy efficiency.

Michael Timothy Bennett
arXiv:2503.23923 · cs.AI · submitted Mar 31, 2025 · updated Jul 18, 2025
abstract · pdf · html · Preprint; paper accepted to and forthcoming in Springer Nature LNCS as part of the 2025 AGI proceedings; 10 pages;

add comment on HN
Also discussed: Apr 2025 (1 point, 0 comments)
ks2048 on Sep 29, 2025 [–]
I'm not sure if there's anything interesting here, but I did notice the author was interviewed on the podcast Machine Learning Street Talk about this paper,

https://www.youtube.com/watch?v=K18Gmp2oXIM&t=3s

reply
getnormality on Sep 30, 2025 [–]
In statistics, sample efficiency means you can precisely estimate a specified parameter like the mean with few samples. In AI, it seems to mean that the AI can learn how to do unspecified, very general stuff without much data. Like the underlying truth about the world and how to reach one's goals within it is just some giant parameter vector that we need to infer more or less efficiently from "sampled" sensory data.
reply
EliRivers on Sep 29, 2025 [–]
Picture a machine endowed with human intellect. In its most simplistic form, that is Artificial General Intelligence (AGI)

Artificial human intelligence. Not what I'd call general, but I guess so long as we make it clear that by "general" we don't actually mean general, fine. I'd really expect actual general intelligence to do a lot better than human, in ways we can't understand any more than ants can comprehend us.

reply
UltraSane on Sep 30, 2025 | parent [–]
Humans are the best/only example of General Intelligence we have.
reply
comeonbro on Sep 29, 2025 [–]
> simp-maxxing

Might want to write this out in full lol I thought this in particular was going to be a much more entertaining point.

reply
zahlman on Sep 29, 2025 | parent [–]
To be fair, it is spelled with a single 'x' in the paper.
reply
nis0s on Sep 29, 2025 [–]
Per my view, it fulfills the following criteria:

1) Few-shot to zero-shot training for achieving a useful ability on a given new problem.

2) Self-determining optimal paths to fine-tuning at inference time based on minimal instructions or examples.

3) Having the capacity to self-correct, maybe by building or confirming heuristics.

All of these concern an intern, for example, who is given a new, unseen task and can figure out the rest without handholding.

reply
SeanLuke on Sep 29, 2025 [–]
My answer: while 99% of the AI community was busy working on Weak AI, that is, developing systems that could perform tasks that humans can do notionally because of our Big Brains, a tiny fraction of people promoted Hard AI, that is, AI as a philosophical recreation of Lt. Commander Data.

Hard AI has long had a well-deserved jet black reputation as a flakey field filled with armchair philosophers, hucksters, impressarios, and Loebner followers who don't understand the Turing Test. It eventually got so bad that the entire field decided to rebrand itself as "Artificial General Intelligence". But it's the same duck.

reply
cogman10 on Sep 29, 2025 | parent [–]
The only difference is the same hucksters are trying to sell the notion that LLMs are or will become AGI through some sort of magic trick or with just one more input.
reply
adastra22 on Sep 29, 2025 | parent [–]
“Strong AI” is the traditional term to compare with “Weak AI.”
reply
SeanLuke on Oct 1, 2025 | parent [–]
My bad. Of course it is. Had a brain fart there.
reply
mwkaufma on Sep 29, 2025 [–]
A term in search of a definition, clearly.
reply
jongjong on Sep 29, 2025 [–]
It's been a moving goalpost but I think the point where people will be forced to acknowledge it is when fully autonomous agents are outcompeting most humans in most areas.

So long as half of people are employed or in business, these people will insist that it's not AGI yet.

Until AI can fully replace you in your job, it's going to continue to feel like a tool.

reply
slfnflctd on Sep 29, 2025 | parent [–]
Robotics are also a big one.

Given a useful-enough general purpose body (with multiple appendage options), one of the most significant applications of whatever we end up calling AGI should be finally seeing most of our household chores properly roboticized.

When I can actually give plain language descriptions of 'simple' manual tasks around the house to a machine the same way I would to, say, a human 4th grader, and not have to spend more time helping it get through the task than it would take me to do it myself, that is when I will feel we have turned the corner.

I still am not at all convinced I will see this within the next few decades I probably have left.

reply
lazide on Sep 29, 2025 | parent [–]
The military would pay 1000x what a household would for the same capability, and they are nowhere near the ability to do that. Which should tell you all you need to know.
reply
adastra22 on Sep 29, 2025 | parent [–]
Without denigrating the importance of robotics at all (it is important), I don’t see the connection.
reply
ACCount37 on Sep 29, 2025 | parent [–]
I wonder if all the grad students that struggle to find jobs now and all the cheap workers in India who were laid off are "feeling the AGI" then.
reply
jonny_eh on Sep 29, 2025 [–]
Please fix the title in HN to match the actual paper's superior title: "What the F*ck Is Artificial General Intelligence?"
reply
dang on Sep 29, 2025 | parent [–]
We don't have an issue with profanity on HN but we do take out clickbait.

Edit: ok you guys, I take the point and have put the original title back. More at https://news.ycombinator.com/item?id=45430354.

reply
adastra22 on Sep 29, 2025 | parent [–]
Replace it with “what the cuss”?
reply
dang on Sep 29, 2025 | parent [–]
The word 'fuck' isn't the issue. The issue is that "What the fuck is AGI", as a title, doesn't add anything besides sensationalism to "What is AGI".
reply
UltraSane on Sep 30, 2025 | parent [–]
It communicates that the paper will probably be a lot less "stuffy" than the typical fancy science PDF
reply
dang on Sep 30, 2025 | parent [–]
I agree with blooalien - that's a great point. To me it doesn't feel quite enough to overcome the baity/provocative effects, but since several commenters have made good points about this, I we might as well put the original title back.

I've kept "f*ck" in the title since that's in the original and arguably adds some subtlety in this case. Normally we'd replace it with the real word since we don't like bowdlerisms.

reply
blooalien on Sep 30, 2025 | parent [–]
> "It communicates that the paper will probably be a lot less "stuffy" than the typical fancy science PDF"

You pose an excellent point... I tend to agree.

reply
adastra22 on Sep 30, 2025 | parent [–]
I don’t know. They typically read entirely differently to me, in the sense that what I would expect to see after clicking the link is different.

I admit though the in this case “What is AGI?” better matches expectation to reality. Before I noticed the domain, “What the f*ck is AGI?” would have led me to expect more of a technical blog post with a playful presentation rather than the review article it actually is.

reply
mbgerring on Sep 29, 2025 [–]
From what I can see, Artificial General Intelligence is a drug-fueled millenarian cult, and attempts to define it that don't consider this angle will fail.
reply
nativeit on Sep 29, 2025 | parent [–]
This feels like we’re approaching consensus. https://news.ycombinator.com/item?id=45418763
reply
robotcookies on Sep 30, 2025 [–]
It is intelligence created by design rather than by natural selection.
reply
heresie-dabord on Sep 30, 2025 | parent [–]
The limitation of your definition is that any intelligence that is untrained will have a high rate of failure.

So, an intelligence may have evolved in geological time or in laboratorical time, but the ability of the intelligence to learn to think and solve problems will distinguish it from the high rate of general failure.

reply
nativeit on Sep 29, 2025 [–]
[flagged]
reply
dang on Sep 29, 2025 | parent [–]
Please don't fulminate. This is in the site guidelines: https://news.ycombinator.com/newsguidelines.html.
reply
emil-lp on Sep 29, 2025 | parent [–]
Stuart Russell said AGI is coming and that we will get 45 trillion dollars from them.

That's what I'm waiting for.

(He didn't specify when or how the money will get here, but I'm betting that I'll get my fair share.)

reply
YesThatTom2 on Sep 29, 2025 | parent [–]
I (and I’m being serious) assumed AGI would break into the world’s financial institutions and steal the 45 trillion.
reply
tim333 on Sep 30, 2025 | parent [–]
Stuart was saying 15,000 tn dollars here https://youtu.be/z4M6vN31Vc0?t=1420

You cheque will be in the post shortly.

reply
adastra22 on Sep 29, 2025 | parent [–]
Hyperinflation?
reply
realityfactchex on Sep 29, 2025 [–]
It would mean actually reasoning, not just applying stats to look like reasoning.
reply
drdeca on Sep 30, 2025 | parent [–]
What do you mean by “just applying stats”?
reply
About | RSS | RSS (all) | HN arXiv