They do have such sensors, and have for many years! Japan's shinkansen pioneered this tech in the 1960s (and has an essentially perfect safety record). It's just that by necessity, the sensing is done through the tracks, rather than directly by the train cars.
You might be imagining something like the automated emergency braking that's used in automobiles, where radar is used to detect an imminent collision. But high-speed trains have stopping distances measured in miles, and it's not feasible for one train to detect another at such distances (especially e.g. on parallel tracks or around curves).
So instead the train positions have to be monitored and relayed by sensors in the tracks, which necessitates a coordination system. And it seems that that system is what broke down. The automated system should have been (and presumably was) fail-safe, but instead of stopping trains until it was fixed, the operators chose to override it and use a much more fallible manual signaling system.
> Stuff like this just further enhances my fear of collective travelling.
I'm sure you're already aware that this is an irrational fear. As bad as this accident is, and as reprehensible as it was that it was allowed to happen, I'm pretty sure there have still been significantly fewer fatalities per passenger-km in Spain this year than there have been on US highways. It's just that we've become numb to the fact that 100 Americans die in car accidents every day.
I sort of agree but I think the issue is more subtle.
Languages change over time, and one of the well-documented "laws" of language change (using that term very fuzzily) is that people like to take irregular features of a language and generalize them to make them more regular. And there are many English verbs (particularly verbs expressing a change of state) which can be used transitively or intransitively, with equal validity:
> "I opened the door" -- "the door was opened" -- "the door opened"
> "the heat melted the ice" -- "the ice was melted" -- "the ice melted"
> "the impact cracked the glass" -- "the glass was cracked" -- "the glass cracked"
This generalization has been common, especially in jargon-y contexts, long before the rise of LLMs. A pedant might insist on saying "wait for the compiler to compile the code", and Merriam-Webster thinks "compile" can only be used transitively. But I think most developers would say "wait for the code to compile [i.e. be compiled]" is fine too. (see e.g. this XKCD comic from 2007: https://xkcd.com/303/)
In my view, this usage is completely fine as an occasional thing, especially when the social process of language change has normalized it over time for a particular verb. The issue is that LLMs drastically overuse it and pepper it everywhere. Just like "it's not X, it's Y", which is fine occasionally in human writing, but has become a dreadful cliche through overuse.
It's definitely context-dependent; it actually took me a minute to come up with a plausible example where I didn't immediately think of a counterexample of why it was a perfectly valid usage, lol.
Even as a grammar weirdo I'm not wholesale against dropping objects from verbs, especially when the object is strongly implied, or the original object becomes the new subject, like in "the heat melted the ice" > "the ice melted". But if you're writing for precision (and I'm a technical writer, so when I'm on the clock I certainly I am :P) there are times where those constructions can obscure useful details, even if they're not grammatically necessary. And there are other examples that I'd struggle to defend in any context, and frankly just give me the ick.
I was trying not to single out any actual sources I've seen, but here's one I saw the other day that made me irrationally mad, and that I think goes beyond the organic changes you mentioned:
> Talking also costs.
It costs? Costs what? Five dollars a word??
The rest of the passage made it obvious that the cost in question had to do with tokens (blatant LLMisms? in a blog post about LLMs? it's more likely than you think) (also ironic considering TFA's conclusion about this writing style being a way to reduce token usage), and maybe one day this will become a perfectly accepted usage and I'll be the old fogie moaning about how it should really say "has a cost" or whatever. But fuck, I hate it, and not just because it's an obvious tell that humans aren't steering the ship.
I think LLM-speak is heavily biased towards using as few words as possible, and as basic words as possible, even when that hurts overall clarity. To use another XKCD example, it often reminds me of "Up Goer Five": https://xkcd.com/1133/
I don't know precisely what to call it, but there is something about AI's word choice and phrasing that really stands out to me.
It's like it chooses "simple" words, but uses them in unclear or fuzzy or slightly abnormal ways that make the overall meaning more difficult to parse. On its own, each instance is defensible, but the aggregate result when it happens many times over is very tiring to read. It reminds me of XKCD's "Up Goer Five": https://xkcd.com/1133/
Like this:
> One waits for an expensive ad, then asks for a hand you can’t place.
It reads as though it started with a straightforward idea, and then was forced to express it with words of only one syllable.
> I don’t think you can get the SCT from a log without the log publishing it?
Well you can, but only if the log cooperates.
That is, the SCT is merely a promise to add the certificate to the log by some particular deadline. Nothing technically stops the log from disregarding that promise. But if it does fail to log the certificate, the SCT itself is proof of its misbehavior.
And it's not enough to subvert a single log, because browsers require SCTs from multiple distinct operators in order to consider a certificate valid.
The SCT may be proof that a log was supposed to be issued, but browsers don't submit every SCT they see (did you know "anonymizing" proxies send CT log companies a sampling of all your TLS traffic? fun stuff!), so in practice these logs aren't audited. As far as multiple operators, almost all of them are based in the US, so just pray the USG doesn't want to target you.
The difficulty with "alignment" is that you have to specify whose goals and values the model is supposed to be aligned with.
Maybe for you and me, this kind of behavior would count as misalignment. But there are certainly a subset people out there for whom the opposite is true. If they tell the agent "get me my refund by any means necessary", harassing bureaucrats and manufacturing excuses is exactly the behavior the want, because it's what they would do themselves if they had the time and patience.
Think about the kinds of people who will escalate petty grievances with their neighbors by weaponizing the police, HOAs, public opinion, etc. Or see this news article about a German tourist who visited NYC for a week and ended up filing three lawsuits, including one against a taco restaurant for having salsa that was too spicy, and one against Wal-Mart because he couldn't get their Wi-Fi to work:
For some reason, the author chose to imitate a "strikethrough" text effect by following each letter with "U+0336 COMBINING LONG STROKE OVERLAY", rather than just using the HTML element.
They survive format changes. If copied and used as plain text, the meaning does not suddenly change from a joke to an ominous "interrogation consultation room".
It certainly doesn't survive everywhere. It fails or renders differently in various applications I've just tested such as pdf renderers, tts, messaging, different browsers, terminals, and at least the one email client I tested it in.
When failing it appears either as "interrogation-" or some form of "interrog̶a̶t̶i̶o̶n̶-" or as "i-n-t-e-r-r-o-g-a-t-i-o-n-"
Unlucky? Crappy tools? I just pasted it into several places on my Android phone, including browsers and renaming a file, and it rendered correctly each time.
Yes, but for one thing, floating-point math is non-associative so even if you use the exact same random seed, you can get different results on different implementation kernels.
For another, the dependence on a random seed is just one example of how you can get different behavior. You could just as well get different results due to e.g. a tiny difference in prompt wording or punctuation or whitespace.
If a model passes some test on 70% of random seeds, and fails on the other 30%, then knowing the rough probability of success is much more useful than knowing one random sample of the outcome.
And empirically this kind of probabilistic success is quite common, because many LLM benchmarks report "pass@k" success rates that are much higher than the "pass@1" success rate on a single attempt.
I don’t understand what “floating-point math is non-associative” has to do with the random selection of possible result token set after the soft max operation. The weights of the model are set and don’t change during the many matrix operations(which is deterministic regardless of the floating point encoding scheme).
The reason GPUs are used for NN computation is that they are massively parallel. You have hundreds or thousands of cores that perform the same operation on different parts of the network.
Tiny variations in the computational environment, such as slightly different set of processes running on the system as well as tiny differences in core temperature or voltage - among many other factors, can cause computations to complete in non deterministic order relative to each other.
This meaning one run eventually calculates a+b+c, whereas another does a+c+b, yielding slightly different results. Those cascade down the chain, resulting, finally, in a different set of produced logits.
Notably, similar effects can and do affect pure CPU calculations on multi core machines, and as a matter of fact even on a single core depending on the architecture when considering eg pipelining, branch prediction etc.
Thus, real world matrix operations are not, in fact, deterministic.
Now, one could argue that all of these are negligible, but that is an empirical claim dependent on the problem at hand.
It depends on the execution order. Model computations can be scheduled differently depending on what else is going on. You need extra effort to guarantee the same order of execution. For example if addition saturates at 8, then 4+5-2 is non-associative.
I believe the answer to your question is that the structure is "connected" to form a solid piece of material, but like any real material it isn't perfectly rigid. So it can vibrate, kind of like a tuning fork, and the thinner parts can act as springs. See the diagrams here: https://pmc.ncbi.nlm.nih.gov/articles/PMC9571586/
If you wanted to use such a device as a simple oscillator, you just set it vibrating and measure the oscillation frequency. Since the material isn't perfectly elastic, the vibrations will eventually be damped and die out (I guess this is what you mean by "friction"), so you use a feedback mechanism to keep the oscillation going.
A gyroscope works the same way, except that you use feedback to maintain an oscillation in one axis, and measure how much of that oscillation couples to a different axis. Whatever effect the damping has on the final measurement, I would assume it can be easily calibrated out.
I don't really know how they physically manufacture these devices without damaging the tiny, delicate structures. It's kind of mindblowing that they're so incredibly cheap.
I'm reminded of Vernor Vinge's novel Rainbows End, which is set in 2025. It features a cyberattack against a rogue AI, which involves revoking a root CA certificate belonging to Credit Suisse, causing massive collateral damage to the European economy and infrastructure.
Vinge didn't quite get to see his rogue AI prediction become (sort of) reality, right on schedule, but he did at least outlive Credit Suisse.
Maybe, but in any case far fewer than will rely on a fixed page. I think very clearly it'd be a net-pain lowering move for example.com to randomize the results a bit, if they felt like it.
~~I feel like that would probably meaningfully increase the amount of compute necessary to serve the website. As it were it’s likely static html, but any randomness would have to rely on a server side operation for each request and couldn’t be cached.~~
I went and checked and in fact it is already serving a script s.js so I guess you’re right!
You might be imagining something like the automated emergency braking that's used in automobiles, where radar is used to detect an imminent collision. But high-speed trains have stopping distances measured in miles, and it's not feasible for one train to detect another at such distances (especially e.g. on parallel tracks or around curves).
So instead the train positions have to be monitored and relayed by sensors in the tracks, which necessitates a coordination system. And it seems that that system is what broke down. The automated system should have been (and presumably was) fail-safe, but instead of stopping trains until it was fixed, the operators chose to override it and use a much more fallible manual signaling system.
> Stuff like this just further enhances my fear of collective travelling.
I'm sure you're already aware that this is an irrational fear. As bad as this accident is, and as reprehensible as it was that it was allowed to happen, I'm pretty sure there have still been significantly fewer fatalities per passenger-km in Spain this year than there have been on US highways. It's just that we've become numb to the fact that 100 Americans die in car accidents every day.
reply