In 2025 Claude had pass rate of 2% in FrontierMath. In 2026 Claude had 88%. Perhaps we should just wait a few years and see if the pass rates stop increasing?

Source:
https://epoch.ai/frontiermath/...

Grades increase extrinsic motivation, but they often hurt intrinsic motivation. I've been thinking that we should replace grades with a single percentage number per subject. This number would tell that you have learned for example 12% of math. This percentage would be calculated automatically based on your answers on controlled environment (tests or class room work). Behind this number would be actual data that shows exactly the things you have not yet learned, so teaching could be targeted there. But as tests would sometimes test past knowledge also, the number might drop also.

The benefit of this single number would be that it would be very simple way to tell how much knowledge a person has and long term learning would provide better scores than studying just for the test.

"within a century (by 2030), rapid technological progress would create so much wealth and productivity that people would only need to work a 15-hour workweek"

You have to wait 4 more years before you can claim that he was wrong. And note that he didn't say that people would work 15 hour weeks, he said that would need to. I actually could do that, I could work for 15 hours only and live with that, but I have decided to work more and retire early.

Actual quote was "by 2025, AI will allow developers to complete a week’s worth of work in four days".

There are plenty of cases where AI has been used to train another AI, usually to make lighter and cheaper AI, which is in some cases better than original. For example AlphaZero was trained this way. There are some examples where humans have done similar things, one European young mathematician comes to mind for example.

No. The real issue is the double underscore. If double underscore had not been allowed in the username, this mistake would have never happened. This is an usability error in the software.

There are plenty of things they could do.
1. You can hire people to generate content. This has been often done with medical LLM training.
2. You can generate content. This can be done when training LLM to generate code or math. Outside of LLM, it has also been done in medical imaging training.
3. You can make smarter LLM that doesn't need that much training data (this is what Google Deepmind aims to do, not OpenAI, but it is an alternative for them also)

One time might be enough. Even if all AI development would stop now, as in these are the models we got and no new models are created, we could still do a ton of work with the existing models. Much more than what we currently can. We are constantly learning new ways to get better results from lesser models so AI performance would continue to improve, even if the models themselves would not improve.

Like, think of a stupid student who always gets an F in a test. What if you would let him do the test with a text book? With the text book, he might get a D. The student didn't get any smarter, nor did he learn anything, but with a small trick, the output was improved.

And even if LLM would be hype, the medical classifiers certainly are not. The usage of classifiers in real medical work is constantly increasing and we are getting results that show that they actually work. It is not hype, it is just poorly communicated to the general public.

That is not true. I did actually read the old encyclopedia in my parent's book shelf, well not the whole book, but the technology parts there as I was very interested in it. Since then I have found much better source for the information and that has been youtube. Yes, youtube is mostly unnecessary things, but as long as you can find it, it contains incredibly detailed information on niche things in a format that is much more easy to understand than it books. Books usually leave a lot of things out and same can be true for videos also, but there are also uncut videos where you can see the whole process and that is extremely valuable if you want to learn something really well.

Go visit your friends. pick a board game or D&D and play with them. Or go outside and play some basketball or anything with them. It is almost free. There are a lot of things you can do with little or no money.

While you are correct, there isn't actually any reason why it should be like that (other than human behavior). We could arrange everything so that everyone just works less, spends less and has more free time. People won't do this willingly, so you would have to use either law or taxes to force people to divide work. I am not sure if this is a smart thing to do, but it is an option.

I agree with you. AI will produce random junk. That is why you need the guard rails that will check and verify the results and throw away the junk. You create a loop where most of the stuff is thrown away and only the good parts are kept.

Let me give you a simple example. Lets say that you want to solve "2+2" math problem, this is how you do it:
1. Give question to the AI. Ask it also to give only the answer as output and nothing else.
2. Check if the result is 4, if it is not, throw the result away, start a new session and go back to step 1.
3. If you got the correct answer, exit the loop.

If that was too simple, here is a paper about more complex problem:
https://arxiv.org/html/2511.09...

Like I said, this limits the use cases, but there are still use cases, especially in math where you can use this.

Just because there are limitations, doesn't mean that it is worthless. You can even use it to solve unsolved math problems:
https://www.technologyreview.c...

> It most definitely cannot "make flawless work without any mistakes, million times in a row"

Perfect means zero errors.

Here is the paper:
"Solving a Million-Step LLM Task with Zero Errors"
https://arxiv.org/html/2511.09...

I disagree with you. AI can make higher quality work than a human. In fact, AI can make flawless work without any mistakes, million times in a row. But it doesn't usually happen unless the AI is specifically guided to do so.

The key difference is that AI (chat bot) doesn't know when it makes a mistake, nor does it know when it does poor quality work. You can add guard rails for the AI that will detect those mistakes and guide the AI to try again, and this is in fact used in many use cases of the AI, but the downside is that you have to write those guard rails by hand and for each situation and it is usually quite hard to do perfectly unless the problem is really simple.

As long as you understand that limitation and you have a good use case, you can setup AI to make high quality or even perfect work for you.