Time Until Superintelligence: 1-2 Years, or 20? Something Doesn't Add Up
00:00:00.000 | Just this week we have had OpenAI tell us that superintelligence might need to be made safe
00:00:05.520 | within 4 years, competing lab leaders say it's decades away, and expert warnings that AI might
00:00:12.380 | have runaway power within 2 years. Let's try to unpack those disparate timelines,
00:00:18.260 | see what might speed up the timing or slow it down, show what superintelligence might mean,
00:00:23.880 | and end with some interesting clips that capture the moment we're in.
00:00:27.920 | But the first timeline is from Mustafa Suleiman, head of Inflection AI this week.
00:00:33.500 | If it's so risky, why don't you stop?
00:00:36.100 | I think that the point of raising concerns is that we can see a moment at some point in the future,
00:00:41.960 | probably over a decade or two decades time horizon,
00:00:45.340 | when slowing down is likely going to be the safe and ethical thing to do.
00:00:51.840 | 10 years is not a long time.
00:00:53.480 | I find it fascinating that he talks about two decades from now when Inflection AI
00:00:57.760 | has just built the world's second highest performing supercomputer.
00:01:02.300 | And even as they admit, that's three times as much compute as was used to train all of GPT-4.
00:01:08.660 | Telling the public that we have a decade or two before we have to worry about safety
00:01:12.860 | seems extremely conservative to me.
00:01:15.420 | But what do we even mean by transformative AI or superintelligence?
00:01:20.040 | Well, here is just one projection of current scaling laws out to 2030 from Jacob Steinhardt
00:01:26.540 | of Berkeley.
00:01:27.180 | And here is just one projection of current scaling laws out to 2030 from Jacob Steinhardt of Berkeley.
00:01:27.600 | And here of course we're talking about just six and a half years away.
00:01:30.400 | If we look at projections of future compute and data availability and the velocity of current improvement,
00:01:36.160 | which of course might not hold forever,
00:01:38.480 | some experts claim that we'll need new innovations beyond the transformer.
00:01:42.320 | But if current projections of future compute and data availability scale up,
00:01:47.360 | here's the kind of thing that we're talking about.
00:01:49.360 | Being superhuman at tasks including coding, hacking, mathematics, protein engineering,
00:01:55.120 | doing 1.8 million years of work,
00:01:57.440 | in 2.4 months,
00:01:59.200 | learning for 2,500 human equivalent years in just one day,
00:02:04.240 | and by training on different modalities such as molecular structures,
00:02:07.760 | low-level machine code,
00:02:09.120 | astronomical images and brain scans,
00:02:11.520 | it might have a strong intuitive grasp of domains where we have limited experience
00:02:16.560 | including forming concepts that we do not have.
00:02:19.040 | Indeed, some research released this week show that GPT-4 already crushes some benchmarks for creative thinking.
00:02:26.000 | And the median forecast,
00:02:27.280 | for being better than all but the very best humans at coding,
00:02:31.200 | is 2027.
00:02:32.720 | And here we have a median forecast of 2028 for AI winning a gold medal at the International Math Olympiad.
00:02:40.480 | The number that I'm looking out for is getting 100% on the MMLU.
00:02:45.360 | That's a test of 57 different subject matters.
00:02:48.560 | And I've actually been discussing with some of the creators of the MMLU
00:02:52.480 | that we might not even know the full potential of GPT-4 on this test.
00:02:57.120 | Officially it's 86.4%.
00:02:59.120 | So we've heard 20 years and 6.5 years, well how about 2?
00:03:04.160 | This article comes from the Boston Globe that did a feature piece on Dan Hendricks and the Centre for AI Safety.
00:03:10.480 | They were behind that one sentence letter that was signed by almost all of the AGI lab leaders and world experts on AI.
00:03:17.760 | The journalist asked Dan Hendricks how much time we have to tame AI.
00:03:21.840 | And he said, well, how long till it can build a bioweapon?
00:03:25.920 | How long till it can have a bioweapon?
00:03:26.960 | How long till it can hack?
00:03:28.000 | It seems plausible that all of that is within a year.
00:03:31.280 | And within two, he says, AI could have so much runaway power that it can't be pulled back.
00:03:37.120 | Seems a pretty massive contrast to Mustafa Suleiman talking about a decade or two from now.
00:03:42.640 | I'm going to come back to this article quite a few times, but now I want to move on to OpenAI's recent statement.
00:03:48.080 | This week they released this, introducing super alignment.
00:03:51.360 | We need scientific and technical breakthroughs to steer and control AI systems
00:03:56.800 | and to make it smarter than us.
00:03:58.400 | I can just see now all the comments from people saying that that's going to be physically impossible.
00:04:02.720 | But moving on to solve this problem within four years, we're starting a new team co-led by Ilya Sutskevert and Jan Leiker
00:04:11.920 | and dedicating 20% of the compute we've secured to date to this effort.
00:04:16.000 | That is quite a remarkable statement.
00:04:18.320 | To their credit, they've made themselves accountable in a way that they didn't have to and that others haven't.
00:04:23.920 | And they're deploying one of the legends of deep learning.
00:04:26.640 | They say that super intelligence will be the most impactful technology humanity has ever invented.
00:04:33.760 | And I agree with that.
00:04:34.800 | And it could help us solve many of the world's most important problems.
00:04:38.000 | Absolutely.
00:04:38.640 | But the vast power of super intelligence could also be very dangerous and could lead to the disempowerment of humanity or even human extinction.
00:04:47.040 | They go on, while super intelligence seems far off now, we believe it could arrive this decade.
00:04:52.480 | Notice they don't say in a decade, they say this decade.
00:04:55.920 | They go on.
00:04:56.480 | Currently, we don't have a solution for steering or controlling a potentially super intelligent AI.
00:05:03.280 | They can't prevent it from going rogue.
00:05:05.520 | And our current techniques for aligning AI rely on humans ability to supervise AI.
00:05:11.440 | But humans won't be able to reliably supervise AI systems that are much smarter than us.
00:05:17.040 | And so our current alignment techniques will not scale to super intelligence.
00:05:22.000 | I'm going to go into more detail about their plan for aligning super intelligence in another video.
00:05:26.320 | But here is the high level overview.
00:05:28.320 | Essentially, they want to automate alignment or safety research.
00:05:32.240 | Build an AI alignment researcher.
00:05:34.960 | I've read each of these papers and posts and some of them are very interesting, including automated red teaming and using a model to look inside the internals of another model.
00:05:45.360 | But the point of including this post in this video was the timeline of four years.
00:05:50.720 | 20% of their compute is millions and millions and millions of dollars.
00:05:55.280 | And 40% of their compute is data.
00:05:56.160 | So that's a very strict deadline.
00:05:58.320 | And one of the most interesting aspects of this post came in one of the footnotes.
00:06:02.720 | They say:
00:06:03.280 | Solving the problem includes providing evidence and arguments that convince the machine learning and safety community that it has been solved.
00:06:12.160 | That is an extremely high bar to set yourself.
00:06:15.200 | They go on:
00:06:15.840 | If we fail to have a very high level of confidence in our solutions, we hope our findings let us and the community plan appropriately.
00:06:26.000 | That's probably one of the most interesting sentences I've read for quite a while.
00:06:30.160 | If we fail to have a very high level of confidence in our solutions, we hope our findings let us and the community plan appropriately.
00:06:37.600 | In other words, if they can't make their models safe, they're going to have contingency plans and they want the community to have plans as well.
00:06:45.280 | And it is a really interesting number, isn't it?
00:06:47.200 | Four years, not even around five years or just end of the decade.
00:06:51.600 | And it does make me wonder what Ilya Satskova thinks is coming within four years.
00:06:55.840 | To have such a deadline.
00:06:57.200 | Now, apparently the prediction markets give them only a 15% chance of succeeding.
00:07:02.160 | And the head of alignment at OpenAI said he's excited to beat these odds.
00:07:06.720 | So we've heard about one to two years and about four years.
00:07:09.760 | But what might slow those timelines down?
00:07:12.240 | The other day I read this fascinating paper, coincidentally co-authored by Jacob Steinhardt, on jailbreaking large language models.
00:07:19.520 | The paper showed that you could basically jailbreak GPT-4 and CLAWD 100% of the time using AI.
00:07:25.680 | And that is fascinating to me as we approach the one year anniversary of the creation of GPT-4.
00:07:33.920 | And the relevance to superintelligence is that if the creators of these models can't stop them being used to commit crimes,
00:07:41.120 | then you would think that they might have to dedicate more and more of their efforts in stopping jailbreaks versus working on capabilities.
00:07:48.320 | For obvious reasons, I'm not going to go into too much detail on jailbreaking here, but here is CLAWD+ from Anthropic telling me how to hold jailbreak.
00:07:55.520 | The first thing I wanted to say is that the most innocent version of the CLAWD+ is the one that I found to be the most interesting.
00:07:56.800 | And to be honest, that's just the most innocent one.
00:07:58.800 | And yes, it did also work on GPT-4.
00:08:01.440 | I did find one of the reasons why it does work quite interesting though.
00:08:05.120 | That reason is about competing objectives where its compulsion to predict the next word successfully overrides its safety training.
00:08:13.040 | And so because those two facets of smartness clash inside the model, it's not an issue that can be fixed with more data and more scale.
00:08:20.640 | What else might slow down the work on superintelligence?
00:08:23.760 | Well, lawsuits and lawsuits.
00:08:25.360 | And possibly criminal sanctions.
00:08:27.360 | Yuval Noah Harari recently said that AI firms should face prison over the creation of fake humans.
00:08:34.160 | And he was saying this to the United Nations.
00:08:36.960 | He called for sanctions, including prison sentences, to apply to tech company executives who fail to guard against fake profiles on their social media platforms.
00:08:46.560 | Of course, those executives might well blame the AI companies themselves.
00:08:50.240 | But Harari said that the proliferation of fake humans could lead to a collapse
00:08:55.200 | in public trust and democracy.
00:08:57.280 | Now it's possible for the first time in history to create fake people, billions of fake people.
00:09:02.240 | If this is allowed to happen, it will do to society what fake money threatened to do to the financial system.
00:09:07.920 | If you can't know who is a real human, trust will collapse.
00:09:11.280 | What's another famous roadblock to superintelligence?
00:09:14.240 | Hallucinations.
00:09:15.280 | I've already talked in another video about how Sam Altman thinks that won't be an issue in 18 to 24 months.
00:09:21.520 | But here again is Mustafa Suleiman on the issue of hallucinations.
00:09:25.040 | Yesterday he said, "Soon LLMs will know when they don't know.
00:09:28.640 | They'll know when to say 'I don't know' or instead ask another AI, or ask a human, or use a different tool or a different knowledge base.
00:09:35.920 | This will be a hugely transformative moment."
00:09:38.880 | And on that I agree, hallucinations are probably one of the biggest hurdles stopping most people from using LLMs more commonly.
00:09:46.000 | It's not about knowing more, it's about when these models bullcrap less, or the moment when they don't bullcrap at all.
00:09:52.160 | But what about things that could actually speed up the process?
00:09:54.880 | What about things that could speed up the timelines to superintelligence?
00:09:57.920 | Going back to the Boston Globe article, one thing could be competition for military supremacy, which has already produced a startling turn to automation.
00:10:07.200 | And that's not just robotics and autonomous drones, that's the LLMs that might control them.
00:10:12.160 | Here is a snippet of a trailer for a Netflix show released today.
00:10:16.160 | "A.I. is a dual-edged sword.
00:10:20.640 | The flip of a switch, and the technology becomes...
00:10:24.720 | lethal."
00:10:25.760 | "There is no place that is ground zero for this conversation more than military applications."
00:10:35.040 | "Forces that are supported by A.I. will absolutely crush and destroy forces without."
00:10:43.680 | "Militaries are racing to develop A.I. faster than their adversaries."
00:10:48.960 | "The A.I., unless it's told to fear death, will not fear death."
00:10:53.120 | "There is no second place in war.
00:10:54.560 | If you're going up against an A.I. pilot, you don't stand a chance."
00:10:58.480 | If language models prove useful in war, the amount of investment that's going to go into them will skyrocket.
00:11:04.240 | Of course, investment doesn't always equal innovation, but it usually does.
00:11:08.240 | And one of the other things that could speed up timelines is the automation of the economy.
00:11:12.880 | For detail on why it might, check out the paper linked above and in the description.
00:11:17.600 | But the high-level overview is this:
00:11:19.840 | As A.I. grows more capable and ubiquitous, companies will be forced to
00:11:24.400 | "hand over increasingly high-level decisions to A.I.s in order to keep up with their rivals."
00:11:30.240 | If an A.I. as CEO does a better job for stockholders, how long can a company resist employing them?
00:11:36.800 | And of course, it doesn't just have to be white-collar work.
00:11:39.600 | As Andrej Karpathy said, "Welcome to the matrix for apples."
00:11:44.160 | But the thing is, whether we're talking about one year or four years or six,
00:11:48.320 | superintelligence is coming pretty soon.
00:11:51.040 | And it is interesting to me that so much of society is carrying
00:11:54.240 | on as if it's not coming.
00:11:55.840 | Take these 50-year long mortgages that are available in the UK.
00:11:59.600 | How can anyone plan out 50 years from now in a world where we might have superintelligence in five?
00:12:05.120 | Of course, I do think we all need to start defining terms a bit better,
00:12:08.400 | and I've tried to do that on this channel with A.G.I. and superintelligence.
00:12:12.480 | But I don't think it's quite good enough to give vague reassurances of a decade or two from now.
00:12:18.160 | How we're going to react when superintelligence arrives is anyone's guess.
00:12:22.480 | We might be crushed by the
00:12:24.080 | sense of inferiority, as Douglas Hofstadter recently said.
00:12:27.840 | Or some of us might become like curious children speaking to a wise adult.
00:12:33.040 | Just the other day, I got a foreshadowing of my own reaction by speaking to Pi,
00:12:37.760 | the model from Inflection AI.
00:12:39.680 | It is designed to be extremely human-like, and the conversations can be quite startling and personal.
00:12:46.960 | Of course, just imagine when they're superintelligent and multimodal.
00:12:50.880 | Anyway, let me know your thoughts in the comments and as all,
00:12:53.920 | always, have a wonderful day.