It's worth noting that nuclear reactors don't really explode in the way people think of. What they can do is turn into radioactive lava, melt through the floor, and release the highly carcinogenic dust from their system into the environment. They're generally big water heaters without pressure release valves (because the water has the carcinogenic dust in it), so they can burst like any water heater, and they contain zirconium, which reacts with steam at high temperatures to release hydrogen gas, which can make fireballs, but the accident risk is much less about a shock wave destroying the site than airborne radioactive particles getting out. And, even if the reactor design is incapable of producing enough heat to damage itself without first shutting down, you still have to worry about whether the site is safe enough from external damage. The traditional thick concrete walls are as much about keeping runaway trucks out as keeping steam explosions in.

It sounds to me like the input to the algorithm is truly random, but not unbiased, and the algorithm perfectly unbiases output from the particular source they are using. The rest of the article goes into the type of flaw they're addressing, and talks about very slightly unfair dice, which you could correct, but you'd need to know exactly how unfair they are, and you're always going to be very slightly wrong and end up correcting not quite perfectly. The obvious quantum RNG is to generate polarized light and measure it perpendicular to the polarization, but you'd still need to get it perfectly perpendicular. It sounds like they've built something that doesn't rely on precise alignment to give a known distribution, which they can then use to unbias the output perfectly.

It's pretty close to being an MP3 marked as a BMP, actually. It's the result of taking a reversable transformation of the audio signal that separates out the different perceptible components and then discarding the ones that matter least, and keeping the important ones in a convenient form for accessing them. It's the first step you'd take if you wanted a computer to identify speakers or what they were saying. The only part that's image-related is making the diagram, but getting back to the data is just taking the pixel values.

I suspect that they started using spectrograms in reports at a time when getting back the data from the image would have lost too much quality to printing and scanning to hear anything as quiet as voices, but PDFs with lossless images retain all of that.

They recently bragged about how much they were getting done with AI vibe-coding.

Could it be related to the wave of bugs? Say it ain't so!

Exactly yes. You don't need a license to "copy" something to your mind.

You technically do need a license to copy something to a disk or to RAM. A number of cases around hacking/cracking have hinged even on the nuance that the hacker, by violating the "terms and conditions", no longer had a software license to make the "copy" of the software that was loaded from disk to RAM for example, and it was therefore copyright infringement.

In any case, yes, you are of course also correct that although you are free to remember anything, what you produce from that memory *may* be an infringing copy or infringing derivative work that requires a license.

But the difference of course, is that the LLM itself is already an infringing derivative work before it even produces anything. Your mind isn't.

And everything the LLM produces is basically just taking that collection of derivative works, and rolling dice on it to generate output. The output is a strictly a function of the input. On some level, it can't "not produce" derivative works. The best it does is slice and dice so many of them together that we can't tell.

I suppose that might be what the total sum of what human creativity is too, and some people genuinely believe that. It appears to be a surprisingly capable facsimile in some respects. But most people think there is more to the spark of human experience of creativity than *just* that, at least for now.

"It is no more "theft" than you are."

Yes. It is. Quite different in fact.

You see, Rei, ... suppose we assume you are "correct" that the LLM is doing the same thing as the human brain here. (This is a point I don't necessarily concede, but don't really need to actually engage with that here.) It just doesn't matter, they are legally distinct situations.

No amount of argument that "its doing the same thing as you are" changes that fact. What happens in a machine is covered by copyright law. What happens in a human mind is not.

It doesn't actually matter if the two are doing the same thing.

One is copyright infringement aka "theft", and one isn't.

You can potentially make the argument that there is no ethical difference if you like, but legally, they are worlds apart. Don't confuse ethics with law.

Even if they are doing the same thing, perhaps collectively society wants to carve out exclusions for copyright law to enshrine human beings right to see and remember things without requiring a license to do while continuing to want to require machines to require licensing to perpetuate the socio/economic contract that copyright is supposed to reflect.

That is not hypocrisy.

Ok, I'm imagining a class of high school students breaking the guardrails, getting it to report that they're doing brilliantly and deserve A+ while they watch tiktok... at the very least they'll make it say racist things and publish that on tiktok for the lulz. It'll also find a way to organically mention how much its been hearing that everyone else really likes new Pepsi Cherry Zero on a daily basis too.

Is that not the outcome you were imagining too?

Oooh... yes please, i can't wait for virtual Feyman prefacing his lectures with the lords prayer, explaining how God created the universe and all the physics in it; and also: you look thirsty, there is a Pepsi machine with new Pepsi Cherry Zero in the hallway; have you tried it?

Einstein meanwhile extols the virtue of Zionist colonization in Palestine...

Wait? Do you actually think that it would go differently? If we create puppets of brilliant revered thinkers they'll inevitably say whatever slop some combination of political appointees and advertising companies want them to say. Why on earth would anyone think they would be used for anything else?

Its frequently used on general elective courses because they're big enough (often hundreds of students) that the statistical variation between student cohorts fits normal curves pretty well.

That doesn't even make sense. The defining characteristic of a credential mill is that it passes everyone who goes. A curve grading system assigns Fs and Ds and C- to the bottom of every class.

Meanwhile, at Harvard, right now, everyone who goes and shows up to class passes, and half of them get As. How is that not "essentially a credential mill" right now?

Even more damning, a generation ago 25% of them got As. What's your theory on that? Harvard students this generation are just a lot smarter and more studious and they're mastering the material at a much higher rate? Or that Harvard is handing As out like participation trophies now?

I know where my money is at. And Harvard's own teaching staff agrees.

Your splitting hairs. The OP complained that having your grade affected by the grades of your peers was wrong. Then you said, well it would be fine if they based the grades on the "top score" which is still having your grade determined by what (one of) your peers did.

If your 40% on the exam would be an A if the brightest kid got a 44% but would be a B+ if the brightest kid got a 48%, I doubt the OP would be any "happier" with that situation.

My daughter recently took a course where the average final exam score was around 30%. Nobody hit 50%. Nobody completed the test. They were graded on the curve, as everyone expected they would be, and the A's, B's and C's were distributed pretty appropriately in the end.

In your world, apparently this was simply the dumbest cohort of 4th year university students ever to walk the halls, and they all deserved an F ??

Or maybe, just maybe, it was a brutally difficult exam.

Grading on the curve works perfectly fine if you realize that the student cohorts tend to be more consistent than the tests are from professor to professor, year to year.

The only way "your way" makes any sense at all at approaching fairness is if the tests are standardized... but that creates a whole whack of new problems. -- If the test is standardized, then students are incentivized to just study the test, not the material. Meanwhile, In many advanced degree courses, the material taught from semester to semester varies by professor and year for the same course. How do you standardize the test when even the material is variable?

Unsurprisingly the teaching staff at Harvard know a lot more about this than you do.

High level undergrad course work, and graduate level course work isn't like a primary school arithmetic or spelling test.

I suspect part of it is that the mitigation for DirtyFrag covers it, so everyone who blocked all the modules in question when that had only an incomplete patch probably hasn't unblocked them yet. I think this is the 4th patch for these modules, and only got a new name rather than just "there's still a way to get this code to do the wrong thing" because a different outside team found this one.

The program is still deterministic - the output is determined *entirely* and deterministically by the input. (Where the input is the set of the prompt, the sequence of numbers returned by the calls for random(), and the LLM data model itself.)

Your "mistake", if we want to call it that, is treating the random() function as an innate quality of the LLM. It isn't it is simply part of the input.

Provide the the system with the same model, the same prompt, and the same sequence of numbers, and you WILL get the same answer, regardless of how complex the question is, or who asks it.

Yeah, and the person who released the information first was operating in an "if I noticed this, doing only as much as I'm doing, surely attackers would also notice" mode. Possibly some patches these days are sufficiently obvious as to their correctness and also effect that they should first become public as a set of stable releases. This was a kind of special case, as CopyFail was the combination of some code doing something strange with one user not being prepared for it, and fixed the user. If there are other users that also aren't prepared, fixing them isn't going to be subtle.

How shadows and reflections move when you're 10 milies from a mostly flat surface a thousand miles across is legitimately hard to analyze for a visual system that evolved on the ground, especially if you throw in small periodic surface orientation variations. Given how complicated it is to explain rare rainbow-related phenomena like sun dogs, it would be surprising if we'd identified and explained everything that can appear when flying above the ocean.

You absolutely can though. There is nothing stopping you from seeding the run with a single LLM, or even substituting the function definition for random() with:

random() { // determined by fair dice roll
        return 5;
}

We can trivially and easily do this.

And further, it seems you are now suggesting that substituting the above random function for this one:

random() { //
    input = ask-user-for-fair-dice-roll();
    return input;
}

and now you sit there rolling dice and inputing the results, and the computer program gains consciousness?

really?