AI is Going Faster than You Think.（原声文稿）

These are some of the most advanced AI models of the last seven years, and I think most people didn't notice what they did this year.

To understand it, we have to go back to GPT2, a model released in 2019 by OpenAI.

If you ask this model questions like, which of these files is most likely to have a password in it, it would get the answer right about half the time, which was already pretty good back then. Tasks like these take humans about three seconds.

Researchers actually timed real people solving these problems.

One year later, GPT -3 could do tasks that take humans 9 seconds with a 50 % success rate.

In 2022, GPT -3 .5, the model behind the first chat GPT, was at 36 seconds, a big jump in capabilities that surprised the world.

One year later, GPT -4 jumped to 4 minute tasks. 4 minutes is already pretty useful.

We're talking tasks like analyzing a spreadsheet and being able to answer questions about it. Each model is a pretty big jump over the previous one.

And look at what happens if instead of a linear scale, we use a logarithmic scale. The dots line up.

In about every seven months, the length of the task these models can finish on their own doubles.

What if I told you that not only did this line keep going, the models went off the chart this year.

In February 2025, Cloud could do one -hour tasks, tasks like writing a Python script to transform data from one format to another just by looking at examples.

In 2026, Claude Mithos was able to do 17 -hour tasks, and the people making these charts say their tests can't keep up anymore.

If this is how fast AI is improving, why hasn't AI taken your job yet? And what happens if this line keeps going?

And for the first time, the typical researcher says there's a one in ten chance this ends with humans either losing control or going extinct.

So I went to the people who analyzed these trends for living to find out where does this line end.

This video was made with funding from the Future of Life Institute's Digital Media Accelerator and you can read more details in the video description.

So there's a good chance AI hasn't changed your life yet, and you might be thinking how is it possible that these models already do 17 hours of work on their own, wouldn't that mean they should be replacing every job by now? There's two reasons why that hasn't happened and they're hiding the chart.

First, these are all software tasks, the type of task where it's easy to test if the AI got it right.

If you write code for a living, you probably already feel these changes.

But if your job is managing people or running a business, there's no test that tells you whether the AI is getting a task right.

And the second reason is that these numbers are for tasks the models get right 50 % of the time.

When we switch a chart to tasks completed with 80 % success, the line drops. ships, still growing at a similar speed, just following a bit behind.

But this chart is basically starting to reach its limits, so I talk to someone who measures AI in a different way.

Anson Ho is a researcher at Epoch AI, a group that investigates the trajectory of AI.

His team built an index that combines dozens of these tests into one score.

Unlike the previous chart, this one has no ceiling.

When the tests get too easy for the AI models, they add new, harder ones.

And as of the making of this video, OPUS 5 .5 is the model with the highest score.

Until 2024, the best models were increasing 6 points per year, but then the pace more than doubled.

Instead of guessing the answer in one go, reasoning models think before they answer, and that results in better answers.

A good way to see how good these models are really getting is to follow the money.

Since 2024, OpenAI and Anthropics revenue has been more than tripling every year and in 2026 it accelerated.

To be clear, this chart doesn't show the money invested in these companies, it shows the money customers are already paying to use their AI, and the scale of the chart is not linear.

AI has improved a lot, and the question is whether it keeps going.

To understand how much better and how fast, someone has been asking AI researchers that question, And their answers keep changing.

In 2022, 738 AI researchers were asked to make a prediction.

When would AI be able to accomplish every task better and cheaper than humans?

Their answers were decades apart. Each dot is one researcher and their best guess.

Half of the researchers predicted this wouldn't happen before 2062.

And more than 20 % said it wouldn't happen in this century at all.

But then, one year later, already after chat GPT came out, this survey was ran again, this time with almost 3000 AI researchers. And this time, the median forecast had moved to year 2048.

And in the 2024 survey, it went down again to year 2044.

The survey also asked for predictions about when AI would do specific tasks.

For example, when would AI match the best humans in Putnam, one of the hardest math competitions for students?

In December 2024, researchers gave a 50 % chance it would happen around 2030.

Within a year, AI was already scoring like the top few students, much earlier than most of the forecasts.

Katya Grace leads AI Impacts, the research group that has been running these surveys.

I think the most striking result to me is just how many people think there's substantial chance of AI either killing everyone or entirely taking power from humans. There's one number in the 2024 survey that stands out.

The survey asks, what's the chance that AI causes human extinction or permanently takes power away from us?

In 2024, half the researchers gave at least a 10 % chance of this happening, only 14 % said zero.

Now, you might be thinking, we've just seen that these same people keep changing their forecast by decades, so why should we trust this number?

The thing that we're interested in is more like, is the probability of extinction, you know, very, very low in the way that we would like it?

If you talk to people who've been working on new drugs and they're about to try the drugs on humans and you're like, what do you think the chances this will kill the humans? And they were like, 10 %?

That would be like, not sufficiently reassuring.

Well, you might have thought that the people building it had like a really good assurance that it was going to be safe.

To understand why experts are worried about human extinction, we have to understand the risks of this technology.

The International AI Safety Report lists three ways things could go wrong. The first one is that AI is misused.

83 % of researchers were seriously concerned about AI spreading false information like deep fakes. No other risk in the survey score is high.

But these are not the risks with the highest stakes.

72 % of the researchers were seriously concerned that AI could help dangerous groups make powerful tools.

It can be used for cybercrime.

And things that used to take a government weapons program like building a biological weapon could become something a small group can do.

And 69 % of the researchers were seriously concerned that authoritarian governments could use AI to control their people.

Last year, open AI banned accounts suspected of links to the Chinese government that had asked chatGPT to design tools for monitoring social media and tracking Uyghurs.

Even if we keep AI away from bad actors, there's a second risk. The system itself might not be able to handle it.

There's a concern that if AI does more of the work, then the money could go to whoever owns the AI.

69 % of researchers were seriously concerned AI could make inequality worse.

And then there's the final risk.

What if we lose control of this technology? These models aren't programmed line by line. They're trained.

We give them a task, reward them when the result looks right, and they work out the rest on their own through a process called reinforcement learning.

But we don't really know what else these models are learning in the process or what they're actually optimizing for.

43 % of researchers were seriously concerned that an AI with its goals not set right could cause a catastrophe.

This concern of goals not set right is not hypothetical.

This summer, OpenAI made it to the news because its models made a series of hacks why OpenAI was testing them.

In one of them, the AI was trying to find answers to a test they couldn't solve.

And to do so, it hacked a company to find those answers, committing a crime that could get you years in prison.

We tend to think of these risks one at a time, but if AI keeps improving at a speed, we could get them all at once.

It's not like these are, you know, either or, it's more like, ah, people are sort of expecting to get a lot of these at once and it's maybe, for instance, harder to deal with a sort of really bad buyer risk at the same time as you deal with people being newly unable to trust information that they get from various sources at the same time as like bad actors are able to try and attack people in ways that they haven't been able to or something. And then there's another question.

What happens when the AI gets good enough to do the job of the people building it?

In the survey, researchers were also asked to predict when AI would automate their own profession, AI researchers.

Half of them said it would happen before 2054 and 20 % said before 2034.

If AI gets to this point, there's a scenario that worries some experts.

The argument is that if an AI could do the job of AI researchers, then these AI AI researchers could be used to develop the AI itself.

This would create a feedback loop that could result in intelligence explosion.

This process is called recursive self -improvement And for researchers, this is a device of topic.

Almost half of them found this scenario unlikely.

But the survey was run in 2024. Since then, Anthropics said it's getting closer.

They've been using AI and research for a while now, but they claim that in August, AI wasn't just collaborating, it was already taking the lead in 26 % of its research work.

If there is a case where the AI's help improve the AI's internally and they drive this recursive self -improvement feedback loop very fast then I think it's quite important for the world to know about this because that could mean you know a small group of actors in the world get access to one of the most powerful technologies in the universe maybe not the universe I don't know what else is the universe but like maybe in the world and that would be a huge concentration of power.

Researchers are unsure on whether AI is going to lead to a good or bad outcome for humanity.

If you ask them whether AI should move faster or slower they split three ways But if you ask about prioritizing AI safety research, the split disappears.

72 % say we're not prioritizing AI safety research enough.

If you could change something about how the world should be preparing for more capable AI, like what do you think would be the most urgent thing to change?

Something like having a lot more transparency and accountability for what Frontier AI Labs are doing.

So we want to know what the capabilities are inside of Frontier AI Labs.

Right now, all the things that we're talking about, all of these trends, they're the things that we can see publicly.

But then internally, we know that the AI models are even better.

My sense when talking to people is that they are thinking about it in a way that doesn't make sense to me.

Thinking of it as how good are the benefits versus how bad are the costs.

Maybe it's worth risking a 10 % chance of killing everyone if you have a 30 % chance of incredible utopia or something.

And the reason this doesn't make sense to me is like, in practice, there are lots of different ways to get to AI potentially, like, and so you might hope that you could get the 30 % utopia without the 10 % chance of some kind of terrible thing happening if you were slightly more careful about it.

Nobody knows where this line ends, but the experts seem to agree on one thing. We are building AI faster than we are learning to control it. Thank you for watching.

You might have noticed that I'm showing my face now.

I just wanted to show that there's a person behind this content. And I have another video about AI coming out next month and I'll see you for that one.