They definitely believe #2 and have been consistent about this since their founding (when they were a no name company nobody used). But HN tends to choose the most cartoonishly evil and conspiratorial takes on these matters because #2 is just boring
Even if they do believe that, they obviously want #1 too because that's how people work. You'd think when talking about a company called "Anthropic" that would be obvious.
The problem with #2 is why should we trust them? If they truly believe #2 then they shouldn't exist either.
China already makes GPUs that are good enough, even if they are slower and less efficient than the NVIDIA GPUs.
Their main problem is not the lower efficiency, but the limited capacity of production, caused by the fact that they cannot buy SOTA manufacturing equipment.
This will indeed slow them down for some time, but it is likely that sometime between 2030 and 2032 they will become able to use EUV lithography in mass production, eliminating the principal advantage of US companies like NVIDIA. After that, they will no longer be constrained by the capacity of production for GPUs and CPUs.
I consider rather doubtful that USA will possess advanced AGI before that.
Even with the smaller amount of GPUs that they can use, I do not see any sign that they will trail by much the AI research from USA.
Which METR report? The one about Anthropic's acceleration rate for improved benchmark results, that put the value at anywhere between 1.1 and 3x per year.
I don't know what your definition of "advanced AGI" is, but if it's the target for a final "win" to the AI race, I think you need to reassess what the current capabilities really are, we're still in the infancy or toddler stages of AGI at best. You should also take a look at how fast Chinese labs are closing the gap, with greater efficiency compared to US labs.
The US mindset of abundance and "scaling laws" is not going to have them win the race. It's not a sustainable strategy.
Opus 5.5 to Model 3 early -> +2.1
Model 3 Early to Model 4 -> +2.3
If Fable 5.2 is Model 3 rather than Model 3 early it reverses to +2.3 vs +1.7.
Based on the score it should be Model 3 rather than Model 3 Early, to me it's not entirely clear based on what it says about the tested Mythos 5.2 being a different checkpoint than the public one.
In any case, we aren't talking about a big leap in score unless further benchmarks narrow the CI, I remember that the Model 4 score supposedly came from fewer benchmarks.
This is unnecessarily cynical. The trajectory of the human condition has dramatically improved over the centuries. Zoom out a bit please, this sort of progress is why you aren't dying in a winter field from some preventable disease at the ripe age of 30, or why you get to see your children actually grow up.
FWIW METR isn't biased, they are probably the most trustworthy/respected organization in this space. You can read their reports and they are clearly not biased.
You are buying a few years but AGI embodied in robotics isn't that far out. It might come sooner than we think if the labs fully crack recursive self improvement.
Get into carpentry because you enjoy it, not because you think it'll be a lifeboat!
reply