AI is Going Faster than You Think. — Transcript
Full transcript
- 0:00These are some of the most advanced AI models of the last seven years.
- 0:04And I think most people didn't notice what they did this year.
- 0:07To understand it, we have to go back to GPT-2, a model released in 2019 by OpenAI.
- 0:12If you asked this model questions like:
- 0:14Which of these files is most likely to have a password in it?
- 0:17It would get the answer right about half the time, which was already pretty good back then.
- 0:22Tasks like these take humans about 3 seconds.
- 0:25Researchers actually timed real people solving these problems.
- 0:291 year later, GPT-3 could do tasks that take humans 9 seconds, with a 50% success rate.
- 0:36In 2022, GPT-3.5, the model behind the first ChatGPT, was at 36 seconds.
- 0:42A big jump in capabilities that surprised the world.
- 0:451 year later, GPT-4 jumped to 4 minute tasks.
- 0:494 minutes is already pretty useful.
- 0:51We're talking tasks like analyzing a spreadsheet and being able to answer questions about it.
- 0:56Each model is a pretty big jump over the previous one.
- 0:59And look at what happens if instead of a linear scale we use a logarithmic scale.
- 1:04The dots line up.
- 1:06And about every 7 months, the length of the tasks these models can finish on their own doubles.
- 1:12What if I told you that not only did this line keep going,
- 1:15the models went off the chart this year?
- 1:19In February 2025, Claude could do 1 hour tasks.
- 1:23Tasks like writing a Python script to transform data from
- 1:26one format to another just by looking at examples.
- 1:31In 2026, Claude Mythos was able to do 17 hour tasks.
- 1:35And the people making these charts say their tests can't keep up anymore.
- 1:39If this is how fast AI is improving, why hasn't AI taken your job yet?
- 1:43And what happens if this line keeps going?
- 1:46Probably the biggest misconception is AI will look like today but a little bit better.
- 1:50Like no, it's going to be a lot better.
- 1:52And for the first time, the typical researcher says there's a 1 in 10
- 1:55chance this ends with humans either losing control or going extinct.
- 1:59I think it's still kind of shocking.
- 2:02So I went to the people who analyze these trends for a living to find out:
- 2:06Where does this line end?
- 2:12This video was made with funding from the Future of Life Institute's Digital Media
- 2:15Accelerator and you can read more details in the video description.
- 2:18So there's a good chance AI hasn't changed your life yet.
- 2:21And you might be thinking: how is it possible that
- 2:23these models already do 17 hours of work on their own?
- 2:26Wouldn't that mean they should be replacing every job by now?
- 2:29There's two reasons why that hasn't happened and they're hiding in the chart.
- 2:32First, these are all software tasks.
- 2:34The type of task where it's easy to test if the AI got it right.
- 2:37If you write code for a living you probably already feel these changes.
- 2:41But if your job is managing people or running a business,
- 2:44there's no test that tells you whether the AI is getting a task right.
- 2:47And the second reason is that these numbers are for tasks the models get right 50% of the time.
- 2:52When we switch the chart to tasks completed with 80% success, the line drops.
- 2:57Still growing at a similar speed, just falling a bit behind.
- 3:00But this chart is basically starting to reach its limits,
- 3:02so I talked to someone who measures AI in a different way.
- 3:05Anson Ho is a researcher at Epoch AI, a group that investigates the trajectory of AI.
- 3:10His team built an index that combines dozens of these tests into one score.
- 3:15Unlike the previous chart, this one has no ceiling:
- 3:17when the tests get too easy for the AI models, they add new harder ones.
- 3:22And as of the making of this video, Opus 5.5 is the model with the highest score.
- 3:26Until 2024, the best models were increasing 6 points per year.
- 3:31But then, the pace more than doubled.
- 3:33So when we see this jump, it corresponds pretty
- 3:35closely to when we had the rise of reasoning models like o1 and o3.
- 3:41Instead of guessing the answer in one go, reasoning models think before they answer.
- 3:45And that results in better answers.
- 3:48A good way to see how good these models are really getting is to follow the money.
- 3:52Actually, the trend that stands out the most to
- 3:53me is probably the revenues of AI companies over time.
- 3:57Since 2024, OpenAI and Anthropic's revenue
- 4:00has been more than tripling every year and in 2026 it accelerated.
- 4:04And it's just like a much clearer signal I think,
- 4:07like harder to fake signal compared to any other trend in like benchmarking and so on.
- 4:11To be clear, this chart doesn't show the money invested in these companies,
- 4:14it shows the money customers are already paying to use their AI.
- 4:17And the scale of the chart is not linear.
- 4:20AI has improved a lot and the question is whether it keeps going.
- 4:23We look at the trends, and the trends have held for a long time and like there's a decent chance
- 4:27it'll continue to hold, especially given how much investment is going into these AI systems.
- 4:31AI in 2030 will not be AI in 2026 plus a little bit.
- 4:35It's going to be a lot better.
- 4:37To understand how much better and how fast,
- 4:39someone has been asking AI researchers that question.
- 4:42And their answers keep changing.
- 4:44In 2022, 738 AI researchers were asked to make a prediction.
- 4:49When would AI be able to accomplish every task better and cheaper than humans?
- 4:54Their answers were decades apart.
- 4:56Each dot is one researcher and their best guess.
- 4:59Half of the researchers predicted this wouldn't happen before 2062.
- 5:03And more than 20% said it wouldn't happen in this century at all.
- 5:07But then one year later, already after ChatGPT came out, this survey was run again,
- 5:12this time with almost 3000 AI researchers.
- 5:14And this time, the median forecast had moved to year 2048.
- 5:18And in the 2024 survey, it went down again to year 2044.
- 5:23The survey also asked for predictions about when AI would do specific tasks.
- 5:27For example, when would AI match the best humans in the Putnam,
- 5:30one of the hardest maths competitions for students?
- 5:33In December 2024, researchers gave a 50% chance it would happen around 2030.
- 5:38Within a year, AI was already scoring like the top few students,
- 5:42much earlier than most of the forecasts.
- 5:44Katja Grace leads AI Impacts, the research group that has been running these surveys.
- 5:48I think the most striking result to me is just how many people think there's
- 5:53like a substantial chance of AI either killing everyone or entirely taking power from humans.
- 6:01There's one number in the 2024 survey that stands out.
- 6:05The survey asked: What's the chance that AI causes human extinction,
- 6:09or permanently takes power away from us?
- 6:11In 2024, half the researchers gave at least a 10% chance of this happening.
- 6:17Only 14% said zero.
- 6:19Now, you might be thinking:
- 6:20We've just seen that these same people keep changing their forecasts by decades.
- 6:24So why should we trust this number?
- 6:26The thing that we're interested in is more like, is the probability of extinction,
- 6:30you know, very, very low in the way that we would like it?
- 6:33If you talk to people who've been working on new drugs and they're about to try the
- 6:37drugs on humans, and you're like, what do you think the chances this will kill the humans?
- 6:41And they were like, 10%?
- 6:45That would be like not sufficiently reassuring.
- 6:47Well, you might have thought that the people building it had like a really
- 6:50good assurance that it was going to be safe.
- 6:52To understand why experts are worried about human extinction,
- 6:55we have to understand the risks of this technology.
- 6:58The International AI Safety Report lists 3 ways things could go wrong.
- 7:02The first one is that AI is misused.
- 7:0683% of researchers were seriously concerned about AI spreading false information, like deepfakes.
- 7:12No other risk in the survey scored this high.
- 7:14But these are not the risks with the highest stakes.
- 7:1672% of the researchers were seriously concerned that AI
- 7:20could help dangerous groups make powerful tools.
- 7:23It can be used for cybercrime.
- 7:25And things that used to take a government weapons programme,
- 7:27like building a biological weapon, could become something a small group can do.
- 7:32And 69% of the researchers were seriously concerned that
- 7:35authoritarian governments could use AI to control their people.
- 7:39Last year, OpenAI banned accounts suspected of links to the Chinese government that had asked
- 7:43ChatGPT to design tools for monitoring social media and tracking Uyghurs.
- 7:48Even if we keep AI away from bad actors, there's a second risk:
- 7:52the system itself might not be able to handle it.
- 7:55There's a concern that if AI does more of the work then the money could go to whoever owns the AI.
- 8:0169% of researchers were seriously concerned AI could make inequality worse.
- 8:06And then there's the final risk: what if we lose control of this technology?
- 8:11These models aren't programmed line by line.
- 8:14They're trained.
- 8:15We give them a task, reward them when the result looks right and they work
- 8:18out the rest on their own through a process called reinforcement learning.
- 8:22But we don't really know what else these models are learning in the
- 8:24process or what they're actually optimizing for.
- 8:2743% of researchers were seriously concerned
- 8:30that an AI with its goals not set right could cause a catastrophe.
- 8:34This concern of goals not set right is not hypothetical.
- 8:38This summer, OpenAI made it to the news because its models made
- 8:41a series of hacks while OpenAI was testing them.
- 8:43In one of them, the AI was trying to find answers to a test it couldn't solve.
- 8:47And to do so, it hacked a company to find those answers,
- 8:50committing a crime that could get you years in prison.
- 8:53We tend to think of these risks one at a time.
- 8:56But if AI keeps improving at this speed, we could get them all at once.
- 9:00It's not like these are, you know, either or.
- 9:03It's more like people are sort of expecting to get a lot of these at once.
- 9:07And it's maybe, for instance, harder to deal with a sort of really bad bio risk at the same time as
- 9:14you deal with people being newly unable to trust information that they get from various sources.
- 9:20At the same time as like bad actors are able to try and attack people in
- 9:26ways that they haven't been able to or something.
- 9:28And then there's another question:
- 9:30what happens when the AI gets good enough to do the job of the people building it?
- 9:34In the survey, researchers were also asked to predict
- 9:36when AI would automate their own profession, AI researchers.
- 9:40Half of them said it would happen before 2054.
- 9:43And 20% said before 2034.
- 9:46If AI gets to this point, there's a scenario that worries some experts.
- 9:50The argument is that if an AI can do the job of AI researchers,
- 9:54then these AI AI researchers could be used to develop the AI itself.
- 9:58This would create a feedback loop that could result in an intelligence explosion.
- 10:04This process is called recursive self-improvement and for researchers this is a divisive topic.
- 10:09Almost half of them found this scenario unlikely.
- 10:12But the survey was run in 2024.
- 10:15Since then, Anthropic said it's getting closer.
- 10:18They've been using AI in research for a while now, but they claimed that in August
- 10:22AI wasn't just collaborating, it was already taking the lead in 26% of its research work.
- 10:28If there is a case where the AIs help improve the AIs internally,
- 10:32and they drive this recursive self-improvement feedback loop very fast, then I think it's quite
- 10:37important for the world to know about this because that could mean a small group of
- 10:41actors in the world get access to one of the most powerful technologies in the universe.
- 10:46I mean, not the universe, I don't know, what else is the universe?
- 10:48But maybe in the world.
- 10:49And that would be a huge concentration of power.
- 10:52Researchers are unsure on whether AI is going to lead to a good or bad outcome for humanity.
- 10:57If you ask them whether AI should move faster or slower, they split three ways.
- 11:01But if you ask about prioritizing AI safety research, the split disappears.
- 11:0672% say we're not prioritizing AI Safety research enough.
- 11:11If you could change something about how the world should be preparing for more capable AI,
- 11:15like what do you think would be the most urgent thing to change?
- 11:19Something like having a lot more transparency
- 11:22and accountability for what frontier AI labs are doing.
- 11:25So we want to know what the capabilities are inside of frontier AI labs.
- 11:30Like right now, all the things that we're talking about, like all of these trends,
- 11:33like they're the things that we can see publicly.
- 11:35But then internally we know that the AI models are even better.
- 11:38My sense when talking to people is that they
- 11:41are thinking about it in a way that doesn't make sense to me.
- 11:45Like thinking of it as a, well, how good are the benefits versus how bad are the costs?
- 11:49Maybe it's worth risking a 10% chance of killing
- 11:51everyone if you have a 30% chance of incredible utopia or something.
- 11:57And the reason this doesn't make sense to me is like, in practice,
- 12:00there are lots of different ways to get to AI potentially.
- 12:02And so you might hope that you could get the 30% utopia without the 10% chance of
- 12:06some kind of terrible thing happening if you were slightly more careful about it.
- 12:11Nobody knows where this line ends.
- 12:13But the experts seem to agree on one thing:
- 12:15We are building AI faster than we are learning to control it.
- 12:19Thank you for watching.
- 12:21You might have noticed that I'm showing my face now.
- 12:23I just wanted to show that there's a person behind this content.
- 12:26I have another video about AI coming out next month, and I'll see you for that one.
About this transcript
This page contains the full transcript of AI is Going Faster than You Think. by Memeable Data, generated from the public captions YouTube serves with the video. The transcript has 2,394 words across 209 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.