Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!

 



Forgot your password?
typodupeerror
×

Comment Re: We do know how it works though (Score 3, Informative) 79

You're half right, they're not Markov chains. But OP isn't describing a Markov chain, he's talking about the transformers used in LLM, which use the output tokens from prior steps as an input to probabilistically generate the next token (based on what word is most likely next, given the entire context and training weights).

Comment Re: USB drives to blame (Score 2) 51

You generally wouldn't. Why would you need to upgrade it? The main reason for most systems is security vulnerabilities. That's not an issue if it's properly air gapped. You're certainly not going to trust something like Windows update or aptitude to update the system anyways (those are a *huge* security risk for state-level entities). If it really absolutely needs upgrades, you'd just pull the hard drive, or replace the entire system.

Comment Re: How is a stream of neutrinos generated? (Score 2) 112

These responses are a great example of why LLMs are absolutely godawful for producing factual information, because both are massively inaccurate or just plain wrong. 1) To produce neutrinos you use protons, not electrons (it might technically be possible to use electrons, idk I'm not that kind of physicist, but no one does). 2) you don't smash them together, you hit them against a target. 3) Neutrinos are neutral particles and can't be guided using electromagnetic fields. 4) Momentum conservation means the produced neutrinos have to travel in the *same* direction as the incoming protons, not the opposite.

Comment Re: Strange take (Score 1) 100

You're mistaking funding for income. The source of income is the end user, but funding mostly comes from investors. Income is (generally) what allow a company to sustain itself and keep running in an operating mode. Funding OTOH allows the company to setup and explore new avenues for income, which in this case means setting up mines/wells/etc, which is far more destructive (long term) for the environment than simply keeping existing projects running. Income can be used as a source of funding, in some cases, but it's a lot harder and scarcer, as investors usually demand the profits in return for their investment. The point of pushing for divestiture is mostly to limit the *expansion* of fossil fuel usage (which is sustained primarily by investors), and not to restrict or reduce current usage (which is sustained primarily by users, not investors). Obviously there are efforts to limit usage as well, but that's done through different means (like fuel efficiency, emissions standards, EV tax credits, etc ).

Comment Re:Was any existing encryption actually broken? (Score 1) 52

That said, one obvious concern in the other direction is that the encryption schemes which we are hoping to be resistant to quantum computing based attacks have had much less attention given to them (in part due to them simply being much younger), and thus we have less certainty that they are even classically good encryption. And we've had now multiple examples of supposedly quantum resistant algorithms being cracked by completely classical methods. See for example :https://cacm.acm.org/news/nist-post-quantum-cryptography-candidate-cracked/. So switching to these new algorithms may be creating new vulnerabilities to deal with a threat that has not yet substantially emerged.

Which is why no one is suggesting moving to a post-quantum algorithm alone. What Chrome is implementing is a hybrid key exchange, ML-KEM768+X25519 (the X25519 part is a standard elliptical curve cypher). Unless your implementation is absolutely terrible, you can't decrease security by layering on multiple encryption schemes, so even if ML-KEM is no more secure than ROT13, it still won't introduce any new vulnerability.

Comment Re:"Half of all people have below-average IQ" (Score 2) 84

That's mostly meant as a joke, but it is also true.

That would only be true if intelligences were evenly distributed across the curve, which they are not. But thanks for proving the point so far as it goes, anyway.

Actually it's true for any symmetric distribution. IQ happens to be normally distributed (by construction: the underlying raw scores are mapped to a normal distribution, though I suspect the underlying scores are themselves probably normal as well), which is very much symmetric. So, you know, maybe don't cast shade on other peoples' intelligence while making a false claim of your own?

Comment Re: This doesn't make a lot of sense (Score 3, Informative) 71

The mere existence of other app stores is not enough to say Google is not a monopoly, if those other stores have negligible market presence. Googles monopoly then became illegal when engaged in anticompetitive behavior with their monopoly (monopolies are not inherently illegal unless abused somehow). What *should* happen (and should have happened a decade ago) is Google being split up. But what they'll end up getting is a fairly minor slap on the wrist in a field they only marginally care about as a prop to their real business of ad sales.
Power

Fire Damages Russian-Occupied Nuclear Plant in Ukraine (theguardian.com) 249

The Guardian reports Sunday, Ukraine's president, Volodymyr Zelenskiy, highlighted that Russian forces appeared to have started a fire in one of the cooling towers of the Zaporizhzhia nuclear power plant that it has occupied since the early days of the war. "Radiation levels are within norm," Zelenskiy said before accusing Russia of using its control of the site, whose six reactors are in shutdown mode, "to blackmail Ukraine, all of Europe, and the world". A Ukrainian official in Nikopol, the nearest town across the river Dnipro from the nuclear plant, added that according to "unofficial information", the fire was caused by setting fire to "a large number of automobile tyres" in a cooling tower. Video and pictures showed smoke dramatically billowing from one of the towers, although experts said they are not in use while the reactor is in shutdown mode, prompting some to question whether it was a way of trying raise the stakes over Ukraine's incursion into Russia.
From the CBC: The Russian management of the facility said emergency workers had contained the fire and that there was no threat of it spreading further. "The fire did not affect the operation of the station," it said. The six reactors at the plant located close to the front line of the war in Ukraine are not in operation but the facility relies on external power to keep its nuclear material cool and prevent a catastrophic accident. Moscow and Kyiv have routinely accused each other of endangering safety around it.

Comment Re:Get her an apology letter from the Nobel commit (Score 1) 63

The Nobel prize to her supervisor wasn't for discovery of the kaon, it was for the development of the photographic technique used in its discovery. The second was also not for the discovery of the particle itself, but for showing it violated CP symmetry.

For that matter she wasn't even the first to discover the Kaon at all: that had happened 2 years earlier at the University of Manchester. Her observations were of an alternative decay pathway. Significant, yes, and that would eventually lead to investigations that would yield a Nobel prize, but simply making those observations (using a technique developed by someone else, in someone else's lab, and under someone else's supervision) might be worthy of a PhD, but wouldn't ever have even been under consideration for a Nobel prize.

Comment Re: Copyright (Score 2) 53

No, training an AI literally involves making a copy of the training data set. If you walk through any guide on how to get started training your own neural network the second step (after you install Tensorflow et al) is literally "go download a copy of the training data set". Even if you argue that the process of training is somehow similar to a human listening to music (which it's really not: the word "training" is used by analogy with how humans learn, the fundamental process is almost completely different), they would need to have bought a copy of every song in the data set for every single thread they used in their training process, as each of those threads will have its own copy of the training data.

Comment Re: Copyright (Score 2) 53

No, they very much have a chance for winning (much as it pains me to agree with the music industry on anything). Copyright, as the name implies, gives the owner the right to control how and when copies of the work are made. Training an "AI" requires making copies of the original work. So unless the rights holder gives permission, using copyrighted data for AI training is a violation of copyright. The "AI" developers will certainly try to argue some kind of fair use exception, but that's a pretty poorly defined area even at the best of times (and commercial uses that compete with the original work are definitely not the best of times).

Slashdot Top Deals

Experience varies directly with equipment ruined.

Working...