The Struggle to Define 'AGI' - Controversial Terms in AI, Explained
Added 2024-09-09 16:00:08 +0000 UTCSurely now that companies like OpenAI, whose only goal is to create an AGI, are worth $100B+, we have a settled definition of 'AGI' itself? No? Or even a set of rival definitions, each of which are well-defined? Well, strap yourself in, we are gonna find out where the term started and show you that things aren't so simple...
Link for Off-line viewing and Download: https://drive.google.com/file/d/1FzHlT2rcJOn8IrCDGbOrNPKM2JqbntIR/view?usp=sharing
OpenAI AGI Definition https://openai.com/our-structure/
Alternative OpenAI Definition https://openai.com/research/
Origin of 'AGI' Term https://web.archive.org/web/20110529215447/http://www.foresight.org/Conferences/MNT05/Papers/Gubrud/
Wired: Sutskever ‘AGI as a Confusing Term’ https://www.wired.com/story/what-openai-really-wants/
AGI Movie - 'The Thinking Game' https://tribecafilm.com/films/thinking-game-2024
US Govt Reference https://ai.gov/wp-content/uploads/2023/10/NAIAC-Charter.pdf
UK Govt Reference https://assets.publishing.service.gov.uk/media/614db4d1e90e077a2cbdf3c4/National_AI_Strategy_PDFversion.pdf
Levels of AGI - Google DeepMind https://arxiv.org/pdf/2311.02462
Already AGI? Google VP https://www.noemamag.com/artificial-general-intelligence-is-already-here/
Debates on the nature of artificial general intelligence - Melanie Mitchell https://www.science.org/doi/10.1126/science.ado7069
OpenAI Denies AGI https://www.nytimes.com/2024/03/02/technology/elon-musk-openai-lawsuit-microsoft-research.html
MMMU Benchmark https://mmmu-benchmark.github.io/
Junior Lawyers Facing Competition https://arxiv.org/pdf/2401.16212v1
Comments
Great video! I feel like it would be helpful to separate out the concepts of intelligence, consciousness, and autonomy (and maybe reasoning, too). Then again, maybe adding more terms to define would just create more confusion.
S J G
2024-09-19 23:36:46 +0000 UTCGoogle.com Hello Phillip, I've gained early access to the full 4o1. How can I test it on Simple Bench? I'm a member of AI Insiders. Can you contact me privately at episcopalian_alien on discord?
r
2024-09-13 01:42:44 +0000 UTCHello Phillip, I've gained early access to the full 4o1. How can I test it on Simple Bench?
r
2024-09-13 01:27:40 +0000 UTCI think trying to define AGI is a fool’s errand. Defining human intelligence has been an elusive goal. When a group comes up with a working model of intelligence which has its uses another group comes up with a very different working model of intelligence which has its own uses. It’s 50 years since I first started learning about models of intelligence and there still isn’t a clear definition which is universally accepted . And even if i make a colloquial statement like Phillip is smarter than me it seems likely we can quickly become caught up in word salad if we push too hard to say exactly what we mean by that. I believe it was Humpty Dumpty who said something like words mean what I want them to mean and that is all. If we think of machine intelligence as a product like smartphones, then one company will say our product does a, b, and c while another company will say it’s product does a,b, g and f. Eventually, products will become more and more powerful in a general sense and subsume other product’s capabilities. Why does there have to be some artificially defined endpoint called AGI? It’s pretty clear to me that Anthropic and OpenAI and Grock, etc, are not being developed simply for the sake of it but to be marketable products which are financially viable. So it will be a game played out in the marketplace, regardless of some easily questionable conceptual definition of AGI.
Swoquix
2024-09-12 23:56:50 +0000 UTCI'm on tenterhooks waiting for your sober assessment of GPT-o1
Andréa Morris
2024-09-12 23:44:14 +0000 UTCLet’s hope it gets a good pass with simple bench
Sean Gallagher
2024-09-12 20:25:07 +0000 UTCI have access to o1-preview. V impressive so far
Sean Gallagher
2024-09-12 20:24:23 +0000 UTCAgreed on waiting 2 weeks to pass judgement. I'm also withholding all my latent hype for 2 weeks from now because if there's evidence shown of the scaling laws holding fast then I'll let myself belive that we can simply naively scale to AGI/ASI in the next ≈4 years and I'll start to seriously change around my life plans.
r
2024-09-11 17:50:06 +0000 UTCMy point is that I think definitions can be constructed. AGI just means a general intelligence. A human level AI is something that is as smart as the average human (this is difficult to judge, but could be ball parked). An ASI is a general AI that is smarter than any human. Unfortunately most define AIG more on the level rather than if it is general or not.
Mike D
2024-09-11 16:53:29 +0000 UTCI have thought of that. One could argue that chat gpt is narrow because it is "only" predicting the next token. It seems to me (and I might be wrong on this) that generalness does rise from this because generality is backed into the training data. That is a good question on the AI system made up of narrow AIs; I would call the system a general intelligence even if the parts are made from narrow AI's. Maybe a good example of this might be our own brain. We presumably have parts of our brains that specialize, but the whole thing together makes our general intelligence.
Mike D
2024-09-11 16:47:29 +0000 UTCSimple Bench is crucial. Current benchmarks are like navigating with faulty instruments that don't accurately measure progress or direction. It's surprising that leading AI companies haven't developed benchmarks that better reflect reasoning and AGI advancement. Perhaps they have, but for internal use only. With past "grand challenges" such as Apollo and the LHC, the instrumentation had to be developed before the other project elements. It's illogical that a multi-trillion $ AI effort would not place extreme priority on reflective benchmarks. Without those, steering toward the ultimate goal is not efficient or maybe not even possible.
Joe Marler
2024-09-11 12:07:00 +0000 UTCAs someone who runs a company trying to build software platforms on top of the current models, I know that the term AGI / ASI is not helpful for our work. The current set of benchmarks is also not beneficial due to the lack of business relevance and the large number of errors in their data sets (as you have pointed out). I have not seen your benchmarks, but it would be beneficial if someone would come up with a benchmark that consists of a large set of well-defined SOPs (Standard Operating procedures). These are the documents and processes that organisations use to run their operations. They are used to define job roles and train users. If an AI system could work through simple, individual SOPs and then collections of increasingly complex SOPs, this would be a valid indication of their capability as opposed to their ability to answer exam questions or solve random logic puzzles.
Clwyd Probert
2024-09-11 11:53:56 +0000 UTCPhilip - Maybe we should timestamp our NAI/AGI/ASI meaning each year going forward and when next year 2025/september comes, we can compare how the definition has changed. I feel as AI matures, so will our understanding of the meaning of intelligence both human and machine. I like your statement "as a moving target" . Finally we maybe should say AGI2024 rather than AGI because I feel AGI2027 meaning will be very different when trying to map the history of AGI.
Jon Kurishita
2024-09-11 11:43:39 +0000 UTCWell with that definition chat gpt is still narrow because it can only do one thing - generate text. Now imagine 3 AIs, one can only generate text, secone one can only generate pictures, third can only route a prompt to one of the other too. You have three narrow AIs but overall as a system they can now do 2 things. Is that general now? Or still narrow?
Sam
2024-09-11 11:38:37 +0000 UTCInteresting point. If it is prematurely defined could lend an argument for heavy-handed approaches
Philip
2024-09-11 11:13:23 +0000 UTCThe ultimate vibe
Philip
2024-09-11 11:12:49 +0000 UTCMight bring it up in a podcast one day! And on progress, let's see what the next two weeks hold...
Philip
2024-09-11 11:12:05 +0000 UTCA meal of thought, for sure. Yes, I am super-ambitious for Simple Bench now, not just for grading current models but expanding into a tool that can truly indicate a new era has arrived, if models ever reach 90%+. Not there yet, lots of work to do.
Philip
2024-09-11 11:11:32 +0000 UTCReally great points. Simple is testing those 'implicit' gaps and exposing them. Thank you Joe.
Philip
2024-09-11 11:10:07 +0000 UTCGreat points Sam, and I like your definition, albeit a hard one to pin down. I think I agree with it as a dividing line.
Philip
2024-09-11 11:09:29 +0000 UTCExactly. It's just unfortunate that this definition made it into a contract like that. MS will never want to define it as AGI, and Open AI presumably will do so when they think it's beneficial for them. Very strange situation they are all in. I wonder if it will even matter eventually, have a feeling not.
Daniel Henderson
2024-09-10 18:04:47 +0000 UTCNarrow AI is an AI that can do only one thing. It can play GO really well, or classify an image. As far as I am aware of there hasn't been any argument on the definition of narrow AI.
Mike D
2024-09-10 14:52:42 +0000 UTCIt's especially interesting that Altman seems to push the viewpoint that AI is a tool on his more recent interviews. So they keep working on a tool - but their mission is to create "something more than a tool"? Really shows they have no idea either :)
Sam
2024-09-10 14:26:18 +0000 UTCYes but how do you define narrow? Philip's argument was that it's hard to draw the exact line where narrow suddenly becomes general. Can you say LLMs are general because they can discuss various topics? (that's general). Someone could argue that they still require words / text, so they are narrow in that sense. Even if you add other modalities (image,sound,video), someone could argue that that's still a limited set of modalities and thus it's narrow.
Sam
2024-09-10 14:23:09 +0000 UTCWow Philip, the idea that you can't separate humans from their tools is so profound, thank you! If AGI is itself a tool, we may need to compare AGI performance vs human+AGI hybrid. Does the true AGI come to be when AGI vs AGI+human doesn't produce drastically different performance? Or is that already an ASI at that point? Very interesting food for thought!
Sam
2024-09-10 14:16:31 +0000 UTCPhilip: as always, your reasoning is clear, incisive and cuts to the core. For me, "AGI" is when leading AI systems can consistently outperform 90th percentile human experts across diverse fields. My experience is current AI falls far short of this (even in a specific field) and conventional benchmarks are misleading indicators. Much human knowledge is tacit, behind paywalls, or otherwise inaccessible to building a comprehensive corpus. This is a major challenge to achieving AGI. Reasoning must also improve. This makes the AGI timeline uncertain.
Joe Marler
2024-09-10 12:32:59 +0000 UTCI could very well be wrong here, but it seems to me that all this confusion is just a byproduct of something much larger. We hear all this talk about AGI and "can you feel the AGI" and see all the predictions that are being made about its arrival without knowing how to quantify what it is, but I believe that what we are 'feeling' is the threat of the unknown caused by the idea that something other than a biological human could be put into the same category of 'sentience' as us. How do we know that someone truly Groks something? While growing up I learned the fundamentals of how something worked but was given different scenarios from my 'training' to prevent over-fitting. Instead of having to learn every word in the English language, and being completely confused when faced with a new word, I was taught the fundamental building blocks that made up the words so that when faced with a randomized problem, I was forced to use the pieces I had to figure it out. This is how I see AGI, and why I think it's the wrong battle to fight. We can't define consciousness, but insist that we can define digital consciousness. I don't believe we have made 'AVA' yet, but we probably will at some point, and I think we need to have a real discussion about what separates us from the bare metal. This is why I think that the SIMPLE bench is such an important benchmark, because it is not testing a models ability to regurgitate knowledge it gained from training, but learning how to reason. Without that it is no better than the robots manufacturing cars in warehouses, simply repeating the same movement it was trained to do, over and over again. Anyhow, just some food for thought. Let me know what you think.
Jonathan Kirk
2024-09-10 02:26:03 +0000 UTCDemis Hassabis Said On Ezra Klein's New York Times Podcast That He Plans On Leading An Expedition To Alpha Centuri. You've done two videos in a row exploring the skeptical side of AI can you make your next video about focusing on what's exciting and the positives about what's happening in AI? Like, Have You Seriously Considered Your Post-ASI Game plan?
r
2024-09-09 23:11:33 +0000 UTCTo put it simply, we should define AGI as "all AI systems that are not narrow".
Mike D
2024-09-09 21:55:17 +0000 UTCI think AI Scientist are overly complicating this. I agree with the one scientist who says we already have AGI. An AI that isn't narrow should by definition be general. A more useful approach is looking at levels of AI. For instance, it'd be possible to have an AGI that is smart as a mouse, one that is as smart as a cat, or one that is as smart as the "average" human, etc.
Mike D
2024-09-09 21:45:44 +0000 UTCIt seems Microsoft has strong incentive to keep moving the goal post, or even suppressing the advent of AGI that can pass the Turing test of public opinion. Unfortunately, we now have profit incentives clouding our judgment of whether AGI is true intelligence or just another tool.
ismschism
2024-09-09 18:04:58 +0000 UTCI feel like agi should be more about it's ability to learn, like you should be able to teach it to do any task
Shawn Conard
2024-09-09 17:58:31 +0000 UTCMakes you wonder why we even care about such a term. Why did Open AI decide to put such a vague term in such an important contract? Why do we have to define anything? For me, AI is a tool like anything else. Until you can simply create an instance of an AI and say "this is your job, do it", and it just does it and fixes all its own problems or enlists the help of others like humans do and can manipulate others to further its own goals like humans do, then it is just a tool. Giving it an acronym and trying to define it seems pretty pointless. I know Open AI said AGI is their mission but tbh, if they don't know what that means then their mission was never set in stone and all they wanted to do was to work on AI and make it as good as possible.
Daniel Henderson
2024-09-09 16:54:15 +0000 UTCIt's uncanny, it's resonant, it's vibrant, it's AGI! There is a sense of familiarity in the word so much so that people "feel" it, and will predict when it 'happens', despite not being able to say what it is. Maybe it's been within us the whole time? Maybe it's a subjective experience or innate qualia? :) AGI feels like a "vibe" rather than science or art, and maybe its beauty is that no definition and any definition applies, making it accessible to anyone to engage in conversations. Def a Vibe :)
Enrico Ros
2024-09-09 16:42:53 +0000 UTCI'm worried that we may be arriving at a situation where if the big AI labs can't come up with a working definition for AGI, eventually governments/agencies/some other non-expert or misaligned organization with the appropriate authority may step in and define it for us. And it's not hard to imagine the negative consequences of that.
Lucas Smith
2024-09-09 16:35:37 +0000 UTC
