AGI Timelines
Added 2023-12-08 21:08:46 +0000 UTCDrawing on interviews with half a dozen top experts at companies like Nvidia, Google DeepMind and Microsoft, as well as a host of other experts in the field, I'll give you their - and my - timelines for artificial general intelligence. By the end of the video, I want you to have a much better sense of what scale is needed, and what the approximate consensus is.
Links:
6 Exclusive Interviews +
Metaculus Predictions: https://www.metaculus.com/questions/3479/date-weakly-general-ai-is-publicly-known/
Hassabis: https://www.independent.co.uk/tech/ai-deepmind-artificial-general-intelligence-b2332322.html
Carl Shulman: https://www.youtube.com/watch?v=_kRg-ZP1vQc&t=2151s
https://twitter.com/BostonDynamics/status/1618619858978996225
Nvidia B100: https://www.semianalysis.com/p/nvidias-plans-to-crush-competition
Yann LeCun: https://www.youtube.com/watch?v=EGDG3hgPNp8&t=5716s
https://twitter.com/ylecun/status/1731445805817409918
Synapses and Neural Networks: https://wp.nyu.edu/yungjurick/2020/03/15/debate-on-the-relationship-between-neural-network-and-the-brain/
Wait But Why: https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html
Jensen Huang: https://www.youtube.com/watch?v=Pkj-BLHs6dE
Amodei Clip: https://twitter.com/dwarkesh_sp/status/1688577515550597121
Avatars at the End: https://shunsukesaito.github.io/rgca/
Levels of AGI, Google DeepMind: https://arxiv.org/pdf/2311.02462.pdf
Comments
Thanks so much r, what a lovely comment
Philip
2024-01-18 14:58:16 +0000 UTCYou are wonderful Phillip. Thank you for your hard work 🙏🏿😔. Your channel is an island of good sense takes amid an ocean of AI hype nonsense.
r
2024-01-18 04:57:28 +0000 UTCThanks Philip, great video. One question you might like to address at some stage if you're talking to a neuroscientist: with Transformers, did AI perhaps stumble onto a process similar to that used by human brains?
Martin Percy
2023-12-30 13:03:26 +0000 UTCI agree the definition around AGI is somewhat off. I prefer to think of it as something that “thinks” for itself and makes its own decisions in a human-like way. Dangerous and grey I know, but growing up, that’s more the sense I got of what AGI is meant to be.
Shaun McDonogh
2023-12-15 06:28:45 +0000 UTCSome of the solutions of Supereminent is using weaker AI models to help stronger ones. https://openai.com/research/weak-to-strong-generalization. There is also code that has been released on Github as well. -> https://github.com/openai/weak-to-strong
Jon Kurishita
2023-12-15 03:47:00 +0000 UTCI’m maybe like 60% thinking that we can get some sort of hacky version of AGI with just GPT-4 (4.5?) level tech + regular engineering. Maybe equivalent to a human with certain special needs. Sidebar: I feel like the definition of AGI has moved to something like “as good as an average human in almost every task”, but that feels so ridiculous because by that definition I wouldn’t even be considered an AGI [or GI I suppose] due to being horrible at a huge number of things most people take for granted.
solarapparition
2023-12-15 00:18:47 +0000 UTCOkay so the Patreon-app has weird UX. I meant to comment this to your podcast! But the video was interesting, too :)
Rens
2023-12-14 19:01:00 +0000 UTCThanks Rens! Much more to come!
Philip
2023-12-14 16:59:10 +0000 UTCThis was amazing!!
Rens
2023-12-14 16:58:48 +0000 UTCHappy to have you here Markus and thank you
Philip
2023-12-14 07:43:30 +0000 UTCVery well done. Happy to be around for the ride!
Markus Heinsohn
2023-12-14 07:32:48 +0000 UTCThere is a subset of AI researchers who believe AGI will be impossible without embodiment. I'm a novice at philosophy, but I believe that viewpoint may have originated with the philosophical school of phenomenology and was later influenced by embodied cognition and enactivism. There is a spectrum of belief in embodiment with regard to AGI. On end of the spectrum, they think it could help develop AGI, and on the other end, they firmly believe AGI will be impossible without embodiment. If that's the case, it would make controlling AGI a lot easier. Just don't give it a body, and it will never happen :) The philosophy underlying embodiment developed before AI research, so the founders of that school never had to consider the dilemma of whether virtual embodiment would fulfill their requirement. The modern-day believers in embodiment have that issue to debate.
Joe Marler
2023-12-10 21:12:57 +0000 UTCFantastic point, I think you are right. Simulated embodiment would be the obvious first instantiation, this is what Nvidia are working so hard on. In effect, the first full-fledged AGI might appear to us like a video game character...
Philip
2023-12-10 10:24:34 +0000 UTC. Does embodiment help AI learn cause and effect better? I would say we could simulate many of the embodiment features of humans in virtually worlds and SIM-type games. You can already see Unreal5 starting to integrate this. Nvidia also is trying to do such simulation with Omniverse Ecosystem and capabilities. We may see the first AGI-like agents in these worlds FIRST before we see it in the real physical world.
Jon Kurishita
2023-12-10 08:08:00 +0000 UTCPerhaps that is correct. Sheer compute might be enough to spark learning and reasoning beyond the data. GPT-4 seems to have given us a glimpse of that. I feel positive that we will live to see the truth of it regardless. I bet you £5 it takes another 3 breakthroughs though lol.
Shaun McDonogh
2023-12-09 16:33:42 +0000 UTCYah, Ilya leaving would almost certainly slow down the superalignment team. And I think it's worth noting that Jan Leike expects to have an internal GPT by mid-2026 capable of "roughly human-level automated alignment research". That sounds very AGI-like. He said: "The roughly human-level automated alignment researcher is this instrumental goal that we are pursuing in order to figure out how to align superintelligence because we don’t yet know how to do that. ... If you want to work back from the four years, I think, basically, probably, in three years you would want to be mostly done with your automated alignment researcher, assuming the capabilities are there. If they’re not there, then, our project might take longer, but for the best reasons." axrp.net/episode/2023/07/27/episode-24-superalignment-jan-leike.html Also of note from that interview: "...we would want them [human-level automated alignment researchers] to figure out how to better align the next iteration of itself that then can work on this problem with even more brain power and then make more progress and tackle a wider range of approaches. And so you kind of bootstrap your way up to eventually having a system that can do very different research that will then allow us to align superintelligence." So I think he's saying the team's plan for year 4 is a controlled, year-long explosion of both alignment and intelligence - with the "alignment explosion" staying one step ahead of the intelligence explosion until superalignment is solved by mid-2027. Presumably their internal GPT would bootstrap itself well above human-level intelligence to accomplish this which is WILD. Does anyone have a different interpretation, though?
Brian Crabtree
2023-12-09 16:23:57 +0000 UTCThing is, with sheer scale of compute, we might not even need breakthroughs (like Mamba perhaps?) to get to AGI...
Philip
2023-12-09 13:37:20 +0000 UTCWith Sutskever leaving most likely, wonder if that superalignment team is slowed down...
Philip
2023-12-09 10:38:09 +0000 UTCGreat TED interview with Shane Legg released on AGI that you should include in the list above: https://youtu.be/kMUdrUP-QCs?si=WV7UgyvCNl6BKg29
Sean Betts
2023-12-09 10:35:51 +0000 UTCFeels like the timeline to AGI is quadratic to the number of recursive actionable breakthroughs (for example transformers). With each recursive breakthrough leading to the next breakthrough, maybe the question is how many breakthroughs lead to AGI? I would guess 3. With the timespan between each breakthrough being shorter due to the self recursive nature you pointed out after featuring Jensen. Also mind and body…feels like chicken and egg. Does embodiment help AI learn cause and effect better…no idea but fascinating.
Shaun McDonogh
2023-12-09 07:57:45 +0000 UTCDoes OpenAI's 4-year Superalignment goal of July 2027 inform your timelines? I think Jan Leike said the plan is... Year 1-3: Solve alignment of human-level AI safety researchers. Year 4: Spin up millions of them to solve Superalignment. Also, given the huge influx of talent and compute into AI research, shouldn't we expect someone to figure out a successor to the transformer before 2028? Seems reasonable to expect humanity will collectively "push" on AGI architectures more in 2024 than it did in the past 10 years combined.
Brian Crabtree
2023-12-09 05:41:11 +0000 UTCYann Le Cun has really brought his AGI expectations forwards it feels? He used to be very belittling of LLMs
Rakesh Murria
2023-12-08 22:21:21 +0000 UTCExcited for this - I knew you’d be a Wait Buy Why fan!
Sean Betts
2023-12-08 21:21:21 +0000 UTC
