9,025 karma · joined April 27, 2016
In Schroedinger cat experiment, the evolution is not thermalization but rather entangling the whole cat with atom state. I don't think we're really puzzled about it now, since we know that macroscopic objects can be in superposition. So cat, theoretically, can be in superposition of alive and dead until measured. I don't think there is any contradiction or confusion here physics-wise.
Unitary evolution can destroy superposition of the small subsystem by introducing negligible amount of superposition into large system.
The other thing you talk about is true. If you have a system in quantum superposition, to measure it you need to entangle it with some macroscopic system. Macroscopic systems decohere very fast. This concept is well explained in excellent John Preskill's quantum computing lecture notes [1]
Note that the system for which wavefunction collapses is not in equilibrium with the measurements equipment (otherwise it wouldn't be measured), so the claims of the paper do not apply to this case.
[1] https://web.archive.org/web/20231005202201/http://theory.cal...
The usual critique of academia is that people are expected to publish a lot (which creates wrong incentives) and that nobody really accesses quality of the publications (which gives opportunities to cheat). At the same time there is huge PhD overproduction, which creates competition.
I don't think the problem is that people are working on wrong problems, the problem is how they work on problem of their choice and why.
I also don't think there is a simple fix to this problem.
PhD is not paid by a student even in US, it's all scholarships.
Also I'd say that what you describe is quite different from what many people believe is the problem with research, either academic or industrial.
Academic research has a problem of low-quality outputs, but it's not necessarily "hard to apply" kind of problems. There is no "engineering department" that comes with problems to people in academia, and there is no inherent expectation of the research to be immediately applicable or applicable at all. A lot of high quality research is not immediately applicable and is not engineering-oriented, in fact. And this kind of research is still worth automatization, just like it is worth doing while we can't automatize it. There is of course a feedback between experiment and theory, but I would say it's the thing that's broken in the academia.
As for industrial research, which I guess you referred to in first place, I have less experience with it. But the goal of industrial research is not publishing papers, so it's not in their KPI, and they won't be drinking champaign for long if they are publishing junk papers instead of doing what they're hired for. Note that not any research done in industry is "industrial research" by this classification, and companies do academic research occasionally (Bell Labs or Google Brain for example). So if industrial research lab is created to solve engineering problems and doesn't do it properly, then, well, any sane company will fire them. It doesn't have wrong incentives problem like academic research, because its incentive is to make money for the parent company.
Now back to AI. Assuming theoretical research is easier for AI, the bottleneck is going to be human doing experiments. Obviously, AI output should be better than human output for it to be tested by experimentalists. What I was saying is that even worse, since the volume of AI output is higher, and the usual heuristics (name of author, personal discussions, etc.) won't work for filtering AI output, it should be better than high-quality human work and not just average human work. I'll be happy if this can be achieved, but we are very far from there.
> You scored 28/100. The best language model, gpt-4, scored 32/100. The unigram model, which just picks the most common word without reading the prompt, scored 28/100.
Assuming complexity averages out on N=100, small test with LLM score above ~5 is "easy"
Unless the testing will be done by AI, I doubt human scientist will bother with testing tons of incomprehensible models, unless the accuracy of these models will be exceptionally high
If you think it will be helpful to others to know about specific failure, put it in a blogpost or even on arxiv. Or talk about it at conference (for CS, workshop).
Also, if we use publications as a measure of scientists success, and we do, is a scientist with a lot of null results really successful?
[1] https://reliefweb.int/attachments/c6421cb7-c936-4145-af54-b7...
I've rerun test locally -- no fail. I've changed the seed to one that was used in failing run -- nothing.
I add a loop to the code to repeat text hundred times -- still nothing. I run the test in bash loop hundred time -- 3 fails. So this already hints on some internal problems. I fixed every possible source of randomness and verified that all the inputs are identical between the runs -- still fails only once in a while. I started building MWE, but the function involved in reproduction is fairly complicated. I'm left with a hundreds lines of Jax code which fails in couple of percent of cases.
I look at the output of the compiler, and it is identical between failing and successful runs. So the problem is in the compiled code. The compiled code is ~1000 lines of HLO (not much better than assembly). Unfortunately HLO tooling is both unfamiliar to me and not well fit to this case (or at least I couldn't figure it out). So I start manually bisecting the code. I'm finally left with ~30 lines of HLO. It fails even less often (1% maybe), but at least it runs fast. It also seems to fail in exactly the same way (i.e., there is single incorrect output that I've between 3 fails). Now that's something maintainers can be hoped to look at.
It turned out that matrices with same content but different layout were deduplicated, leading to, in my case, transposed matrix being replaced by non-transposed one. The hash used for storage did take layout into account so the bug appeared only if two entries ended up in the same bucket (~3% of times). The fix was an obvious one liner [1].
[1] https://github.com/openxla/xla/commit/76e7353599d914546f9b30...
[1] https://mlcommons.org/benchmarks/training/
[2] https://cloud.google.com/blog/products/compute/introducing-t...
Standardized testing is a good predictor. In my country people are admitted based on standardized testing only (country wide subject exams and sat-like test) or results of university-adjacent "preparation courses" and it works fairly well. The affirmative action is realized directly as a bonus points to your scores based on the background, which reduces the effect of ethical gaming.
Energy scales for new physics can be enormous for all we know, I do not see any reason to expect that new physics hides "around the corner" (in a 1.5 orders of magnitude FCC or muon collider could give us). Muon collider is at least looking in a different place, and it's definitely more interesting than FCC for me, but I'm still not 100% sure it is the right thing to do, given the limited resources.
> intent to minimize civilian casualties
Intent is very hard to argue about. Especially when talking about proportionality which is not rigorously defined. Numbers are easier to argue about. It is reasonable to expect that civilian casualty rate will be higher in cases where there is no intent to minimize civilian casualties compared to cases where it is; comparing numbers with conflict in similar conditions where you believe there was such intent is a good starting point
> enable any sort of progression into a longer holding peace in that area
Getting rid of terrorist government who promised to repeat 7th of October looks like necessary condition for peace.
Can you give the numbers for comparison? What are acceptable rate in your opinion?
Being indistinguishable from civilians (blending in) is exactly using civilian as a shield -- the opponent can't identify you and can't kill everyone who looks like you
> They can still minimize civilian death
Arguably Israel does it with fairly low civilian: combatant death ratio