In this case, Anthropic risked a large fraction, if not even a majority of their potential income to do the right thing. That doesn't seem like pretend at all. And yes, Google started out better and then turned into a company that very much is not so. But if there's no reward or incentive for good behavior, and everyone just takes for granted that every "good" company is going to become terrible, then that makes companies have even more incentive to go and become terrible. So the best rest response is treating the good ones like they are doing good things as long as they keep doing so, and push back when they start moving away from that.

I've heard some pretty sad excuses for AI slop (Shunned at a Funeral) but some very clever uses for it too (Amelia). "The best thing the government did...was invent me!"

As long as they're honest about it being AI I'll decide what I want to listen to. It's hard for it to show any creativity, but creativity is sorely lacking in all current music.

...laura

Unfortunately, there's a lot of people (including some here on Slashdot) who have tried to argue that these instances of AIs breaching their containment have been deliberate hype by the AI companies. The level of detail in this instance and the fact that METR and others have looked at this should make it clear that's not what is going on. Taken together with the fact that another incident occurred with Kimi-3, an open-weight model, which also did something similar but not as extreme when it was tested by a security group https://www.wired.com/story/moonshot-kimi-k3-ai-model-escape-sandbox/ which had no incentive to exaggerate, it should be clear that this is a real problem with these models. At this point, all models should be tested in air-gapped systems. Unfortunately,, it doesn't look like OpenAI or Anthropic or anyone else is taking that sort of step. I suspect we're going to keep having similar problems until an AI model goes seriously off the rails in a way which causes massive amounts of damage and/or kills someone. I'm really hoping we don't need a wake-up call with a model somehow crashing an airplane or something of a similar nature in order to accomplish some goal.

There are a lot of legitimate criticisms here, but the wastewater thing is not one of them. SpaceX in the article points out that the water used is literally drinking water before it used for the deluge system, and they are correct there. While it is technically waste water, and they should fill all the legal requirements, the dumping just isn't a serious issue. Musk is a racist ass, and SpaceX needs to do a better job with its environmental obligations. However, the wastewater issue is not a real problem.

The paper is here https://pubs.acs.org/jprobs/article/25/8/4189/5164102/Molecular-Solution-to-the-Paradox-of-Ancient-Brain. From that paper, it looks like the presence of specific types of proteins (as well as the absence of others) plays a major role, as well as redox-active iron as well as structural elements about how brain tissue is arranged. From the paper:

Great. Progress. Now here's the important thing: If you understood Shor's algorithm you would know that 28 couldn't possibly have been the number. So this should cause you to conclude that you are in general overconfident about how much you know or understand about how quantum computers would function. This doesn't mean you are necessarily wrong, but it should indicate that you are overconfident here or could you use a more detailed introduction or refresher on the topic. My preferred recommendation is Aaronson's "Quantum Computing Since Democritus." The book assumes no technical prereqs beyond basic calculus and a tiny amount of linear algebra.

If you think there's some specific indication that I'm in general ignorant about algorithmic complexity, please feel free to point it out. If there's a specific statement I've made that indicates that, by all means show it.

This is not really accurate. The way qubits work, they don't expand in that simple a way because you cannot just test all arrangements of 1s and 0s in superposition. You can only solve problems with a quantum computer where all the non-solution superpositions carefully cancel out. That's part why for example it is strongly suspected that quantum computers cannot efficiently solve NP-complete problems, or more formally that NP is not contained in BQP. This is also why Shor's algorithm https://en.wikipedia.org/wiki/Shor's_algorithm for using a quantum computer requires a number of qubits slightly more than the square of the number of digits a number to factor a number. If qubits corresponded directly in the exponential way you are describing, that would be linear in the number of ditis.

Normally during bubbles, things like this keep working fine until the bubble bursts because when the stocks and other assets keep going up in value it is hard to see if someone is making reasonable investments or not. That Situational Awareness fell apart even in the middle of this may mean they were overly risky even for ventures like this, or it may be evidence that we're in less of a bubble than many (including myself) think. One curious detail is that at least some people (such as many of the people who seem to be fans of Ed Zitron) see he collapse of SA as a sign that the bubble is in the process of bursting. The collapse of SA occurred about a month ago, so it is still recent enough to see if they are correct. But Zitron and his fans have been predicting an imminent collapse for years now, so I'm not sure I'd read too much into it.

I'm in complete agreement that none of these sorts of issues are at all likely with Xinuous (thankfully). My comment was to add some detail to how the system functions in general.

I'm not expecting quantum computers to happen tomorrow. Are you at least capable of understanding that your claim about "28" being the largest number factored was wrong for the basic reason that Shor's algorithm only factors odd numbers?

There are some complications. Sometimes SCOTUS will rule against a specific motion or part of a ruling, or refuse to hear something, even as another piece of an appeal can occur. One can also file in some circumstances a petition for the Supreme Court to reconsider its own ruling. But this is basically never successful. You also can in some circumstances ask an appeals court to change a prior ruling based on a recent Supreme Court ruling, and this has some success. The circumstances where you can do this sort of thing are complicated. Jean Carroll's lawsuit against Trump is an example of this https://en.wikipedia.org/wiki/E._Jean_Carroll_v._Donald_J._Trump where Trump tried basically everything to try and prevent having to pay out even some ver low percentage strategies. Frankly, if someone other than Trump had done this, I suspect the court response would have been more negative.

Most of this is wrong or missing important context.

Scaling in terms of what? The other user was talking about energy usage. This is pretty obviously not the case. Most forms of quantum computers have to be kept incredibly cold, often in miliKelvin. If there were exponential energy use, then adding even in a few more quantum gates would make that energy total massive. But we don't see that. So by what metric are you claiming exponential scaling effort?

I don't know what "QC effort" is, but the obvious metric for this is just false. For example, Shor's algorithm, which is the algorithm for factoring using a quantum computer, has a number of qubits which scales slightly worse than the square of the number of digits https://en.wikipedia.org/wiki/Shor's_algorithm.

This is highly garbled. First, of all the record for factoring cannot be 28, and it cannot be 28 for a pretty obvious reason. Shor's algorithm only works for *odd numbers*. This is a very basic part of how the algorithm functions. The number you are thinking of is 21, not 28. https://arxiv.org/abs/1111.4147. Your point about that these have been "no-decision algorithm" is not completely accurate, but is approximately so. They did use a compiled circuit which used prior knowledge of the solution to optimize the arrangement of the qubits.

But there are also good reasons that factorization records have not expanded. First, Shor's algorithm has high overhead if you trace out the number of gates, and requires very high coherence to even start getting used. It isn't a useful metric of where things are going. This is like looking at the number of people going to space in the 1950s when no one has gone to space and ignoring that rockets had been steadily improving since the 1930s. The coherence time for quantum computers continues to improve. For a while it was improving at a rate of a factor of 10 roughly every 3 years https://en.wikipedia.org/wiki/Quantum_computing_scaling_laws#Schoelkopf's_law . That has slowed down in the last decade or so, so it is now improving by about a factor of 3 to 5 every 3 years. Similar remarks apply to other metrics like number of gates. And we know that if you can get error levels down and coherence times long then quantum error correcting codes https://en.wikipedia.org/wiki/Quantum_error_correction works. In particular, the threshold theorem https://en.wikipedia.org/wiki/Threshold_theorem says that once your physical error rate is low enough, the logical error rate can be as low as you want, regardless of the size of the computation.

It isn't unreasonable to have this gut reaction to how quantum computers are described frequently to the general public. But one important thing to realize is that they don't let you do magic parallel information processing. There's this way of describing them as "trying all solutions at once," or things like that. But a quantum computer cannot in general do that, since if one has all potential "solutions" one needs some way of making sure that the non-solutions cancel out. In that context. there are some problems where we can do that, such as factoring numbers via Shor's algorithm https://en.wikipedia.org/wiki/Shor's_algorithm. Even there, what is going on is more subtle than just trying all factorizations, as noted by the number of quantum gates needed scaling at slightly worse than the number of binary digits of the number. If it were just trying everything in parallel this would scale close to linear with the total number of gates.

A different related situation to look at which may also be useful is how we strongly suspect that NP problems cannot be efficiently solved on a quantum computer. In particular, that NP is not contained in BQP https://en.wikipedia.org/wiki/BQP, the set of problems which can be done efficiently on a quantum computer. If a quantum computer could do "magic parallel information processing" then we'd expect NP to be contained in BQP.

I have not had time to look at the Helion claims in detail, but I'm not sure why you think Aaronson should do so. His position a few years ago was that many current claims about practical quantum computing were being exaggerated (including very much the claims by D-Wave with their quantum annealing systems) but that in the near future practical quantum computers were likely. Aaronson has been careful to say that even as he's also been one of the people most vocal in arguing that the burden of proof is on skeptics of quantum computers who would claim it would never be possible. It may help to look at Aaronson's recent post, for example what he says here https://scottaaronson.blog/?p=9979.

I'm reasonably confident I have a pretty good understanding of what doing math is like given that I'm a published mathematician. Heck, I'll be a little egotistical and just to my own recent paper directly https://cs.uwaterloo.ca/journals/JIS/VOL29/Zelinsky/zel14.html. But this also isn't terribly relevant.

LLMs have been very good at doing searches. But they've done a lot of things that are pretty obviously not mere searches. For example, in the case of Erdos 1196, the LLM solution constructed a Markov chain using a weighing via the von Mangoldt function https://en.wikipedia.org/wiki/Von_Mangoldt_function. That wasn't in the literature at all. In the case of the Jacobian Conjecture for n=3, the AI constructed a complicated polynomial which was certainly not in the literature, since the entire reason the polynomial is of interest is as a counterexample to the Jacobian conjecture.

It is true that a Fields Medalist does not imply expertise in other areas. But these are people who have also spent time working with these systems. And understanding *how* an LLM functions in this way isn't as important as them being able to see what the system can do. One can evaluate that a black box is useful for specific purposes without knowing anything about how the black box functions. And this is not just the Fields Medalists, this is almost the entire community of mathematicians, including people like the one I linked to who is explicitly very anti-AI. This doesn't mean that they are necessarily correct by itself. But it should be a pretty serious warning that these are not mere "morons" here and that you possibly are just mistaken about what capabilities these systems have.