@ThenElection's banner p

ThenElection


				

				

				
0 followers   follows 3 users  
joined 2022 September 05 16:19:15 UTC

				

User ID: 622

ThenElection


				
				
				

				
0 followers   follows 3 users   joined 2022 September 05 16:19:15 UTC

					

No bio...


					

User ID: 622

Creating convincing animated JRPG sprites.

there seems to be a parallel announcement that a 100% organic proof was being kept on the down low while undergoing final review. I think this makes it quite likely that the AI proof is a bit less of a calibration-shattering advance than it appears to be at first glance

This is wishful thinking. Some points about the human attempt:

  1. It might be 100% organic, but it's not 100% complete. It's maybe 75% complete. It has all of chapters 11, 12, and 13 littered with dozens of statements like "Return here" or "Complete proof." Section 12.3 ("Putting the pieces together: global to S-local deconstruction of the geometric kernel", p. 514) is literally blank with a footnote saying "Include section"

  2. That in itself isn't damning; these potentially can be developed out. The author had budgeted a year at Stanford to complete the proof; so, if all went well, it means humans were merely a year behind AI in solving the RH. The point is that this wasn't a state where it's crossing ts and dotting is or getting a close read verification; the proof attempt is far from that point.

  3. To get a sense of how likely it is for everything to go well, I fed it into some models. All identified serious flaws in the attempt. They're beyond me to adjudicate, but a sample response:

LLM: Section 12.3 is blank in the manuscript because the bridge it needs to build between Chapter 11 and Section 12.4 cannot exist. [insert lots of math that's beyond me]

Me: How likely is it for the author to be able to recover the proof from that?

LLM: Virtually zero. Not because of missing technical glue that a clever estimate might supply, but because the exact identity the paper tries to prove in Theorem 12.5.1.1 is mathematically false. [Provides a counterexample]

There is a genuine interesting tradeoff: probably some parts of the chip process are inherently serialized, and there's probably some optimal policy that is neither "focus on chips immediately" versus "wait to start on chips until we've reached the end of software improvements," even with the sizeable unknowns about what both hardware and software improvements will look like. What that policy is, I'm not smart enough to say. Probably worth spending some compute to figure out.

I remember a lot of cope back when Navier Stokes was announced. One amusing example was things along the lines of NS being the easiest Millenium Problem, and if people had paid mathematicians the same amount as they used for AI, we'd have gotten it already. You would think this will put that to rest. But the goalposts will undoubtedly shift, and we'll hear a bunch of nonsense about the Riemann Hypothesis being trivial, or AI just stealing a human solution, and the people who are in the thrall of the anti-AI psychosis will walk away thinking this is meaningless.

This will change exactly no opinions, because anyone who's not a willful denialist about AI capabilities could see this coming with NS at the latest.

Agree with all your points. But

If the AI companies could do anything, they'd do things to achieve material superiority like devising advanced nanomolecular engineering for radically faster AI chips (perhaps optical or some other exotic method). That's worth a trillion times more than the Riemann Hypothesis.

I would put solid money on them having done so already (in fact, investing in Anthropic is exactly that bet). To the extent that they aren't full steam there, it's because they're focusing on software-only approaches to material superiority, which have a much shorter feedback loop.

It came with a Lean proof, in the process of being validated; this is not a contentless PR result.

This honestly comes off as desperate cope. Even in the unlikely world where the human proof checks out, the author claims that he used no AI for it. How would Anthropic have gotten a hold of it to plagiarize? Its agents secretly hacking Google's servers to snoop through his gmail?

This is merely evidence that Claude is not amazingly way better than human mathematicians in this case.

Claude spent two weeks attempting this proof. Humans have spent over a century on it.

Yes, Claude really is that much better.

I will note that starving people tend to become lean meat, and thus to be useless as food without some other source of calories

What I'm hearing is that it's better to go cannibal as the first resort, before the other food sources have run out.

Could I avoid the protein poisoning by going for tissue high in fat? E.g. hunting people for their brains.

About 20% of my 401k.

I was saddened to find out that my main 401k plan doesn't allow IPO participation for the self-directed brokerage. Was going to go for a similarly sized position as you.

cannibal looter mobs will be rampant and I'll likely have had to shoot some

Your cannibal looter mob, my opportunistic acquisition of nutrients following self-defense.

I have some level of confidence that datacenters and Internet trunks themselves are safe. At least, they're much harder targets than the others, and the hyperscalers are absolutely putting massive resources into securing them.

Of course, all the supporting infrastructure is vulnerable as heck, and a datacenter with no power doesn't do much.

I'm not sure the decabillionaire bunker is even the best option for them. It needs staff and automation to work, and both of those are unreliable (and with what money will you pay the staff or technicians?) You'd be better off filling a bag with millions of dollars to throw cash at the mob (I know someone who literally did this while fleeing Indonesia in the 60s) as you run to your flight off to Buenos Aires.

On the order of months, my wife and I are going to buy 40 or so acres an hour outside the city. It'll let me LARP (best case) my farmer fantasy and also (worst case) my prepper fantasy. It's fun.

Aside from that, we already have a couple weeks worth of food and water, along with a go bag and various emergency supplies. That's more for natural disaster purposes than permanent social breakdown, though. In a case of complete breakdown, probably the best thing to have a supply of is enough fentanyl to OD on (or whatever the most pleasant way to go is).

All the recent disruptions have entirely reasonable explanations, but it's odd that they're happening all around the same time. I'd be curious to see an analysis if it's a real escalation or just recency bias. Could also be AI-related but in more comforting ways (vibe coding shitty code, or rushing to deploy patches to vulns).

There's only one way he can recover from this one: more leveraged AI bets!

ChatGPT pulls up a skeptics exchange post that has some additional research:

https://skeptics.stackexchange.com/questions/57431/in-1964-were-some-prospective-voters-in-louisiana-asked-to-spell-backwards-fo?noredirect=1.com

Most interestingly, the original source appears to be

Melvin Lubershane, “Teaching Students To Follow Directions,” Chicago Schools Journal, vol. 41, no. 2 (November 1959), pp. 72–76.

https://archive.org/details/sim_illinois-schools-journal_1959-11_41_2/page/74/mode/2up

It contains an instructional exercise for students aged roughly 10 to 15, rather than a voter registration test, and the supposed voter registration questions come from it.

That's not some hypothetical future, by the way. The raw cost of inference is cheaper for the US frontier labs today than even the most shoestring budget setup using open weight models, even for substantially larger and more capable models. People confuse API costs with how much it costs the frontier labs to run inference, and that makes people underestimate just how far ahead the frontier labs are on this front.

Even at equilibrium (and we won't be at equilibrium for awhile), there're network effects that allow for much better hardware utilization. The main thing open weight models do is set a weak cap on just how large frontier labs' inference margins will be.

Correct time is when you believe markets will be going up at a fast enough rate to counteract fees and volatility. As it so happens, that's been most of the time for over a decade.

AI will kill this strategy though as it'll drive real rates through the roof.

Her involves the AIs explicitly engaging in self improvement until they ultimately decide to go rogue and abandon humanity (a relatively optimistic outcome).

Gary Marcus has said he expects AGI by the late 2030s IIRC. Although his predictions might be slow, exponentials are hard (and it's not impossible for him to be correct even now), and he's fundamentally someone who does believe in AI, though he plays up his skepticism of the current paradigm for popular audiences. You just can't interpret his statements in the obvious way, if you choose to read him (more trouble than it's worth IMO).

Some concepts are just too high IQ for most people. Probably most people with an IQ below 100 don't really understand supply and demand for example. And the floor for understanding recursive self improvement is a lot higher.

This isn't my read on things. Average and dumb people can understand the concept just fine. They might not like it and the changes it brings, but they absolutely recognize the threat; it's been communicated by, if nothing else, pretty simple and direct sci fi stories that they've seen on the big screen.

But it takes a higher IQ person to come up with superficially insightful but wrong/irrelevant points that leads to true AI denialism (as opposed to AI opposition). Stochastic parrots, Chinese rooms, explanations of why a trend has to stop. They've heard of P =! NP; surely that prevents AI! They can construct a TESCREAL genealogy of eugenics to AI, therefore it's bad and fake. Etc.

There are stronger arguments, of course, but I'd classify questions around things like model performance generalizing beyond RLVR friendly tasks as a different category than true denialism.

I was referring to the recent CVE, where versions of openssh compiled with musl would allow you to login to root with any random 257+ length password. Greater password length doesn't help with that; in fact, the bug makes longer passwords less viable as a mitigation.

I can imagine a world where AI lets defenders run fully formally verified software that's been hardened to perfection. But the fact that institutional inertia is going to prevent us from getting there quickly doesn't change the fact that the transition is going to be very painful. And it's not just old creaky software: it's everywhere. Not just weirdly superhuman out-of-band attacks that only a superintelligence could imagine, but bugs on the level of "you can get root through ssh on some systems by inputting a random password longer than 256 characters." Software is ubiquitously insecure.

He's also a Taiwanese engineer. He's rising above his station, which upsets people.

Nice to have AI do all my work now so I can post on the themotte, read X, and trade stocks all day.

What's your implementation? MCP of some brokerage? Do you trust your agent to do anything sophisticated, or mostly just buy/sell?

The solution is obvious. We’re going to work invented jobs

But if the jobs are all invented, how do we decide who gets the better fake invented jobs and who gets the worse ones? It's not a new problem, but at least now there is some semblance of an escape valve for people who don't like the fake job allocation process. (This is less a reason it can't happen and more a reason it'll be shitty.)

I wonder if we'll even realize the jobs are all fake. One person's fake job is another person's critical function to protect consumers/users. One plausible transition isn't some Great Fake Jobs Act of 2035 but instead the government finding it has a lot more funding and a lot more interception points in the economy and passing laws to take advantage of its new capabilities. Probably better psychologically for people than a complete lack of structure.