I've travelled all over the world. I've familiarized myself with many places, and met many people. And I'm a Canadian, although I’m privileged to reside once again in the States.
And here's something I've noticed, and it’s a key element of America's continuing greatness:
You bloody Americans value success, and you believe in its existence.
This is something that doesn't really happen anywhere else in the world. Even in other free democracies—the United Kingdom; Finland, Sweden, and Norway; Australia, New Zealand and Canada; Germany, France, and the Netherlands (great countries all)—a counterproductive cynicism too often reigns.
Success is equated with exploitation.
Ambition is looked upon with contempt.
This happens sometimes in the United States too—particularly among the miserable progressives, who confuse their resentment, ingratitude and unearned skepticism with wisdom.
But in your great country, by and large, striving is admired and success celebrated.
This means that more people strive and succeed in the US than anywhere else. And it's increasingly obvious. You remain stunningly more innovative and productive than any people anywhere else on the planet.
And so I say, as all should who are fortunate enough to live in the western world, let alone America:
Thank God for the United States.
Thank God for the wisdom of its founders.
Thank God for its faith in the free market and in the natural rights of man.
Happy birthday, you damn Yankees and Southerners.
Long may your admirable country dominate the world.
Long may your freedom and hope provide an example to those suffering everywhere at the hands of their malevolent states.
May your two and a half centuries of unparallelled success be just the beginning.
Your country is the light of the world, and the city on the hill.
The GPU-kernel item is the one I’d watch. Once a model can write the kernels that make models faster, the tool starts improving its own substrate, and that loop compounds. The catch is verification: a kernel that’s subtly wrong or slightly slower is expensive to notice, and “looks right” is not “is faster.”
the risk i'd actually worry about is the checker becoming the attack surface. optimize hard enough against a verification harness and the model just finds the try/except that makes the test green. goodhart's law, basically.
This tech tale reminds me of Dune and that world’s banishment of AI after the Butlerian Jihad. Hopefully it will not come to a galactic battle. I wonder if, as some theologies tell us, that angels are assigned to planets and nations, species, people, and professions, if they might also become assigned to AI. It would be nice to think so.
From the Orange Catholic Bible of Dune - “Thou shall not make a machine in the likeness of a human mind.”
The GPU kernel result is the part worth sitting with. Once AI starts writing the low-level code that makes AI faster, you get a compounding loop that most automation timelines quietly underprice. But there's a distinction I keep bumping into running AI operations daily: generating a kernel and trusting a kernel are different problems. A model can produce something that benchmarks beautifully and still fails on the edge case you didn't think to test. The bottleneck stops being generation and becomes verification. That's the bit the automation narrative tends to skip. The interesting question isn't whether AI can write the kernel, it's who signs off on it, and whether that sign-off scales as fast as the generation does. In my experience it doesn't, which is why the real gains come from constraints and approval gates, not raw capability. The analog computation thread is the sleeper here too. If the automation story is about squeezing more out of the same silicon, a shift in the substrate itself resets the assumptions everyone's forecasting on. Curious whether you see verification as the thing that paces all of this, or just a temporary drag.
Honestly the score is less interesting than where the agents fail: recovering hidden state, tracking many items, handling instructions that conflict. That's a decent description of a normal Tuesday at an office job. The freelance work where AI hit 16.1% is the opposite kind of task, fully specified and self-contained, so the gap between those two results tells you which work gets handed off first.
Always one question - amazing info even for a non techie listener - I do always wonder if we in some ways test the systems “unfairly” - I know we look for them to do complex jobs - but seems like huge variety of jobs all by same system - could you create a high level marketing ad ready to roll out ? I feel like humans, systems should be tested based on some expertise and so on - don’t need a “healthcare system” to know and create songs or marketing …. Is that a way to think or is a the I can do it ALL no matter what the true test ??
I wish I was 1 year old instead of 81 years old.
Quote of the Day:
To the Americans:
I've travelled all over the world. I've familiarized myself with many places, and met many people. And I'm a Canadian, although I’m privileged to reside once again in the States.
And here's something I've noticed, and it’s a key element of America's continuing greatness:
You bloody Americans value success, and you believe in its existence.
This is something that doesn't really happen anywhere else in the world. Even in other free democracies—the United Kingdom; Finland, Sweden, and Norway; Australia, New Zealand and Canada; Germany, France, and the Netherlands (great countries all)—a counterproductive cynicism too often reigns.
Success is equated with exploitation.
Ambition is looked upon with contempt.
This happens sometimes in the United States too—particularly among the miserable progressives, who confuse their resentment, ingratitude and unearned skepticism with wisdom.
But in your great country, by and large, striving is admired and success celebrated.
This means that more people strive and succeed in the US than anywhere else. And it's increasingly obvious. You remain stunningly more innovative and productive than any people anywhere else on the planet.
And so I say, as all should who are fortunate enough to live in the western world, let alone America:
Thank God for the United States.
Thank God for the wisdom of its founders.
Thank God for its faith in the free market and in the natural rights of man.
Happy birthday, you damn Yankees and Southerners.
Long may your admirable country dominate the world.
Long may your freedom and hope provide an example to those suffering everywhere at the hands of their malevolent states.
May your two and a half centuries of unparallelled success be just the beginning.
Your country is the light of the world, and the city on the hill.
Thank God for the USA.
Happy 250th.
Dr. Jordan B. Peterson
The GPU-kernel item is the one I’d watch. Once a model can write the kernels that make models faster, the tool starts improving its own substrate, and that loop compounds. The catch is verification: a kernel that’s subtly wrong or slightly slower is expensive to notice, and “looks right” is not “is faster.”
the risk i'd actually worry about is the checker becoming the attack surface. optimize hard enough against a verification harness and the model just finds the try/except that makes the test green. goodhart's law, basically.
This tech tale reminds me of Dune and that world’s banishment of AI after the Butlerian Jihad. Hopefully it will not come to a galactic battle. I wonder if, as some theologies tell us, that angels are assigned to planets and nations, species, people, and professions, if they might also become assigned to AI. It would be nice to think so.
From the Orange Catholic Bible of Dune - “Thou shall not make a machine in the likeness of a human mind.”
Tech tale rhymes with world in Stephenson's "Anathem"
Today one Cyber LAWYER was animately proposing we should not allow AI to make another AI machines
But now Fable is writing kernel codes. It will all accelerate from now on
The GPU kernel result is the part worth sitting with. Once AI starts writing the low-level code that makes AI faster, you get a compounding loop that most automation timelines quietly underprice. But there's a distinction I keep bumping into running AI operations daily: generating a kernel and trusting a kernel are different problems. A model can produce something that benchmarks beautifully and still fails on the edge case you didn't think to test. The bottleneck stops being generation and becomes verification. That's the bit the automation narrative tends to skip. The interesting question isn't whether AI can write the kernel, it's who signs off on it, and whether that sign-off scales as fast as the generation does. In my experience it doesn't, which is why the real gains come from constraints and approval gates, not raw capability. The analog computation thread is the sleeper here too. If the automation story is about squeezing more out of the same silicon, a shift in the substrate itself resets the assumptions everyone's forecasting on. Curious whether you see verification as the thing that paces all of this, or just a temporary drag.
Honestly the score is less interesting than where the agents fail: recovering hidden state, tracking many items, handling instructions that conflict. That's a decent description of a normal Tuesday at an office job. The freelance work where AI hit 16.1% is the opposite kind of task, fully specified and self-contained, so the gap between those two results tells you which work gets handed off first.
Hi what's up jack , Why do you banned Iranian accounts in Claude AI?
wait, if Fable can actually handle the kernel writing part, does that mean the barrier for custom hardware is basically gone now? [language: en]
Always one question - amazing info even for a non techie listener - I do always wonder if we in some ways test the systems “unfairly” - I know we look for them to do complex jobs - but seems like huge variety of jobs all by same system - could you create a high level marketing ad ready to roll out ? I feel like humans, systems should be tested based on some expertise and so on - don’t need a “healthcare system” to know and create songs or marketing …. Is that a way to think or is a the I can do it ALL no matter what the true test ??