Jacob Tsimerman believes that AI will be better than human mathematicians within 2 years.
We should not be surprised if AI solves the Millenium problems very soon and advances to a level that is barely comprehensible, or even incomprehensible, to the best humans.
We can expect this not by curve fitting to recent progress but by reasoning from first principles about where the progress has come from (synthetic data, non-human corpus) that has no obvious upper bound on capabilities.
The obvious and primary application will not be on showy millennium problems or long sought conjectures, but on optimizations and breakthroughs surrounding transformers.
We are doing fine so far with just Transformers, we have just crossed Trillion parameter models, we don't know if 100s of Trillion param model wont be better.
The breakthroughs would be mostly in optimizations for computing tokens, not as much in expanding parameter count (if anything we want to shrink that).
Does anyone know if Tsimerman talked about AI extinction risk at other places? Critch has worked long-term in the field (MIRI, CHAI) but I was surprised to see this colab.
May all world-class mathematicians take a look at the AI alignment problem such that humanity can have a better chance of passing through, and may some of them decide to not focus just on sexy parts like coming up with ways to kill us all, and may those who do so anyway at least come up with more interesting or plausible stories than the ones presented in this paper.
I think this says more that being very good at theoretical math does not at all translate into intelligence or subject matter experience about how humans will realistically handle potential armed conflict and potential risks of mass casualties, at the political/nation-state level.
Ed Zitron should write his newsletter with LaTeX and publish them as PDFs in arxiv. Seems to make the techbros automatically take it seriously. Maybe that's what it takes to pop the bubble.
Whether or not Ed is right, he doesn't take any criticism, but peer review is a fundamental tenet of actual science. Vs being a blowhard on the Internet and getting paid for it.
Peer review is actually a surprisingly modern concept. It is not actually a fundamental tenant of science. The scientific method notably doesn't actually include peer review in it's core.
Why would an article from 2019 be a counterargument to criticisms from 2022? Especially when the first sentence of that article is a prime example of Mearsheimers blind spot: that the Russians aren't actually realists, but instead are a kind of idealist focused on Russias dignity as a (percieved) great power. Almost like virtue ethics as opposed to the consequentionalist ethics of realism.
Chalmers: “It is natural to hope that there will be a materialist solution to the hard problem and a reductive explanation of consciousness, just as there have been reductive explanations of many other phenomena in many other domains. But consciousness seems to resist materialist explanation in a way that other phenomena do not.”
"A landscape of consciousness: Toward a taxonomy of explanations and implications"
No, I mean that the logical possibility itself is neither obvious nor intuitive. If you physically copied me atom by atom, every instinct in me says that the result has conscious experience. I'd need some serious convincing otherwise.
High oil prices are good for most of the oil industry, but the oil industry is only like 3% of the US economy.
On the other hand, elevated oil prices are bad for a very large chunk of the other 97% of the US economy, who rely on buying oil and oil derived products in order to operate their industry.
reply