And I think the effort here can be compared in scale to the Manhattan or Apollo projects, but there is also the potential for a huge backlash to the hype that was built up and created what is arguably a bubble, so this is a race against the clock.
I also think he is wrong about the markets reaction, markets are inherently good integrators and bad predictors, we should not expect to learn anything about the future by looking at stocks movements.
A priori, it was not obvious that those clearly stated problems had solutions within our grasp (see fusion) but at least we knew what the big picture looks like.
With AI, we don't have that, and never really had that. We've just been gradually making incremental improvements to AI itself, and exponential improvements in the amount of raw compute we can through at it. We know that we are reaching fundamental limits on transistor density so compute power will plateau unless we find a different paradigm for improvement; and those are all currently in the same position as fusion in terms of engineering.
That doesn't mean the approach isn't an evolutionary dead end, like every other so far, in the search for AGI. In fact, I suspect that is the most likely case.
The fact that something simple like the Transformer architecture can do so much will spark so many ideas (and investment!) that it's hard to imagine that AGI will not happen eventually.
They will need to be so different that any talk implying current LLMs eventually replaced humans will be like saying trees eventually replaced horses because the first cars were wooden.
> And I think the effort here can be compared in scale to the Manhattan or Apollo projects, but there is also the potential for a huge backlash to the hype that was built up and created what is arguably a bubble, so this is a race against the clock.
It's not useful to blindly compare scale. We're not approaching AI like the Manhattan or Apollo projects, we're approaching this like we did crypto, and ads, and other tech.
That's not to say nothing useful will come out of it, I think very amazing things will come out of it and already have... but none of them will resemble mass replacement of skilled workers.
We're already so focused on productization and typical tech distractions that this is nothing like those efforts.
(In fact thinking a bit more, I'd say this is like the Space Shuttle. We didn't try to make the best spacecraft for scientific exploration and hope later on it'd be profitable in other ways... instead we immediately saddled it with serving what the Air Force/DoD wanted and ended up doing everything worse.)
> I also think he is wrong about the markets reaction, markets are inherently good integrators and bad predictors, we should not expect to learn anything about the future by looking at stocks movements.
I agree, so it's wrong about the over half of punchline too.
unless you consider people who write clickbait blogs to be skilled workers, in which case the damage is already done.
> AI is exceptional for coding! [high-compute scaffold around multiple instances / undisclosed IOI model / AlphaEvolve]
> AI is awesome for coding! [Gpt-5 Pro]
> AI is somewhat awesome for coding! ["gpt-5" with verbosity "high" and effort "high"]
> AI is a pretty good at coding! [ChatGPT 5 Thinking through a Pro subscription with Juice of 128]
> AI is mediocre at coding! [ChatGPT 5 Thinking through a Plus subscription with a Juice of 64]
> AI sucks at coding! [ChatGPT 5 auto routing]
Whether they turn out right or wrong, they undoubtedly cheered on the prospect of millions of people suffering just so they can sound good at the family dinner.
I wouldn’t want to work for or with these people.
I think we might see AI being much, much more effective with embodiment.
Will it do as good a job a competent adult? Probably not. Will it do it as well as the average 6 year old kid? Yeah, probably.
But given enough properly loaded dishwashers to work from, I think you might be surprised how effective VLA/VLB models can be. We just need a few hundred thousand man hours of dishwasher loading for training data.
Plumbing is a requires the effective application of engineering knowledge, and I don’t think unsupervised transformer models are going to do that well.
As a large language model developed by OpenAI I am unable to fulfill that request.
There've been next to no advancements relative to what's needed to redefine our economic systems by replacing the majority of skilled workers.
-
Productionizing test-time compute covers 80% of what we've gotten in the last 6-8 months. Advancements in distillation and quantization cover the 20% of the rest... neither unlocks some path to mass unemployment.
What we're doing is like 10x'ing your vertical leap when your goal is to land on the moon: 10x is very impressive and you're going to dominate some stuff in ways no one ever thought possible.
But you can 100x it and it's still not getting you to the moon.