LLM-based “AI” is able to use its vast corpus of inputs and calculate the most statistically likely output in a given situation. It is probabilistic, and when you are dealing with probabilities in a situation where certainties, not probabilities, matter, you’re going to get dinged on credibility massively when your LLM-based “AI” gets the probabilities wrong at best, or in this case, claims a line of code generates a vulnerability when it is, in fact, a code comment.
LLMs are text-prediction engines. They are not Artificial Intelligence, and shouldn’t not be treated in any form or fashion as if they possess intelligence. What bothers me about this entire situation is that presumably the folks that relied on the LLM-based “AI” to generate these vulnerabilities knew (or should have known) enough about their tool to know this would happen, but did not.
Now, we all pay the consequence, to the tune of hundreds of thousands if not millions of dollars of wasted productivity from teams that have to deal with the resulting fall-out of this usage of “AI”.
A human must verify everything an LLM presents as fact. Everything. If you don’t, we all pay the price. LLMs do not remove the onus of responsibility on the human being, if anything they amplify it because LLMs can generate lots more output more quickly that needs to be verified than humans can.
People that previously have barely any experience in tech, now being hired in AI startups because they are good bullshitters.
Obviously.
Can they do the job? Because right now, government decisions are based on AI generated code, which was verified by nobody who can do that. So the cost of an unsatisfactory answer is quite high.
I like the symmetry of people being unable to detect "skill" in people who are great at manipulating language and in LLMs who are also great at manipulating language.
The sort of person that's going to offload their thinking to AI is the exact sort of person that is not going to verify anything because they've already offloaded their thinking to AI.
B...b...but the Anthropic trainer said we'd get the best results if we don't think of it as a tool, but instead give it a name and think of it as our brilliant coworker!
Why should I trust you, internet rando over a stormtrooper-level salesman? /s
Brandolini's principle in action. It takes 10 times more energy to refute BS than to generate it. A related analogy to computing: it is easy to generate propositions, but hard to test if a given proposition is satisfiable or not, which curiously ties to P vs NP.
Watching for unexpected failure modes is surely worth it.
Intelligent life-forms can generate probabilistic outputs based on inputs, but being able to generate probabilistic outputs based on inputs is not what makes us intelligent.
Likewise, there is no reason to think the brain employs super-Turing or quantum computations that cannot be approximated by LLMs.
> exclude LLMs with CoT from the category of intelligent systems with certainty
At least, don't you think that the recent mathematical results of LLMs are a bit like a glimpse of something teapot-shaped in the orbit? (which makes it not a Russell's teapot, which. by definition, can't be observed).
To me, it's an expected progression of ANNs' approximation of human cognitive processes. The universal approximation theorem guaranties the existence of such ANNs barring the super-Turing or quantum superiority of the brain.
I don't? You are presenting opinions as if they are mine, but they are not.
They could have been great, if trained on datasets from a more sensible species.
???
Of course it is. The brain is mechanically not capable of doing anything other than that.
Do you believe the brain is something other than a bundle of probabilistic physical interactions? Or are brains not the source of what we call intelligence?
This is going to elevate your thinking on this no end, if you're interested.
We know that the brain is a probabilistic input → output machine because the universe is a probabilistic input → output machine. The brain is made of universe. There are deterministic relationships (which at high sensitivity or complexity become easier to describe as probabilistic), and quantum relationships. That's it. The brain, like every other thing comprised of "universe" is comprised of those two types of relationships.
If Romain's book provides evidence of relationships in the brain that are neither quantum (therefore random) NOR classical (therefore deterministic), then 1) he would have already won at least one Nobel prize, and 2) anyone in this thread would be able to at least gesture toward what relationship that is.
The brain is an object in the universe.
The universe has quantum behaviors (fully random, not a source of intelligence) and it has deterministic behaviors (fully non-random). Many of those deterministic behaviors are so complex that they're easier to analyze and describe as probabilistic, which is where most brain input → output relationships land.
Please point to any evidence whatsoever that the brain has some third type of interaction going on that has never been observed anywhere in the entire universe, then we can have a discussion about it.
The brain is a (very complex, incredible) input → output machine. That's it! It's incredible!
I don't understand why people are so afraid of this that they will believe otherwise with literally zero evidence whatsoever.
The brain is deterministic at the level of specific interactions, which process inputs in a highly chaotic (but still deterministic) manner into probabilistic outcomes.
The opposite of deterministic is random, i.e. in the quantum sense of truly no relationship between input and output.
There are probably some quantum effects in the brain here and there, but the vast majority of it is just traditional deterministic interactions networked together in such a complex system that the resulting behavior is much, much easier to predict in probabilistic terms than otherwise.
You cannot say an AI model cannot be intelligent because it's a probability machine, when all available evidence points toward natural intelligence also being generated by probability machines (much more complex ones, called brains).
Are newly born babies reacting due to statistical probabilities that they have derived, or are they using something other than their brains?
The answer is obviously yes lol.
The creature is an assemblage of electrical, chemical, and kinetic relationships.
Watching a baby develop is exactly what you'd expect from a system that's predominantly electrical noise triggering behaviors and then gradually refining denoising the relationship between inputs and outputs, with the goal function of achieving more desirable inputs.
Surely you can at least gesture toward one thing in the brain that appears not to be a probabilistic relationship between input and output?
“LLM has made legitimate mathematical discoveries” —> Wow the rate of progress is amazing. Highly upvoted.
“LLM does something not good” -> Does everyone else not realize LLMs are just dumb next token predictors? Highly upvoted.
So tired of this discourse and this site.
These models can do a lot of things but they also can't do a lot of things. In order to use these models effectively you have to understand that they are next token predictors and how that allows it to do what they do.
The issue is of course with using the word "dumb": they are next token predictors, no doubt about it, but whether LLms as a class of system are smart or dumb is entirely unknown and entirely variable in time.
To interact with them effectively you must know how they behave, just like you have to know how humans behave to interact with them effectively. If you disagree, find someone with autism and have a conversation with them.
Or should the discourse in a diverse community like HN only reflect the positions you personally hold?
I mean even perennially contentious topics will get this behavior.... some thing about emacs makes the front page, within a day or two there will be a vim post up there. Same with Rust is (good|bad), or if systemd creates an even more awesome tool, the haters will come along and recycle stories about bugs from over a decade ago.
There's a lot of people here. Not all of them read it every hour, and discussions like this among large groups often take a very long time with lots of repetition. Human group dynamics (aka politics) is slow.
> So tired of this discourse and this site.
You're welcome to leave if you don't like it. The site was like this long before you joined, and will like it long after you leave I'm sure.
It's also worth noting, that an awful lot of math discoveries are perfectly in line with dumb next token generators - they are finding a way to formally construct an argument and being surprised when it doesn't work, or surprised at the outcome of the grind. Not all of them are made by brilliant leaps of intuition.
I agree that humans must verify LLM-produced facts, but strongly disagree with these kinds of "stochastic parrot therefore dumb" arguments.
Yes, an LLM is a "stochastic parrot". No, that doesn't imply that it is dumb. Enough to look at how Terence Tao asks ChatGPT to help him understand a solution that nobody had ever discussed before [1], or how a random guy asks ChatGPT in a handful of words to disprove a 30-year-old conjecture, with zero technical input [2].
If your parrot in a birdcage with internet access can finish the sentence, "The counterexample to the Dinitz–Garg–Goemans conjecture is...", then it's a pretty smart parrot, by all reasonable definitions of "smart". Just because someone bottled up the formula into matrix multiplications and added some random sampling to the outcome, that doesn't take away from the fact that the parrot said provably correct statements that the biggest experts in the field couldn't imagine.
And no, I'm not implying that the LLMs are correct all the time, or that their intelligence and reasoning works in any way like ours.
[1]: https://chatgpt.com/share/6a5fdc7a-d6f8-83e8-bbea-8deb42cfed... [2]: https://chatgpt.com/share/6a60b2eb-0b64-83ee-9c76-7931ca1de0...
This line of critique is pernicious because it is both technically correct, as description, and profoundly misleading.
Saying that outputs are a product of inputs is not interesting and to the point it is not explanatory.
What is interesting, is how they do what they do. What is the "statistically likely* next token? To answer that you can do exactly one thing, run the LLM. That's because what they are doing is interesting and not reducible.
What is more interesting is that in order to do what they do, given the architectures we apply and the training strategies we use and the harnesses we situate them in, LLM are recapitulating in their deep layers strategies observed in the animal brain. This is still suggestive, interpretibility is nascent: but it is also more than a little interesting. In some respects, for cognitive scientists interested in the manner in which mind merges from computational substrates, it is profoundly interesting.
One can incorporate this, and, still be viciously critical of bother the success and failure of LLM in the applications we have put them to, and of how we (as individuals and as institutions such as corporations) are integrating them into our work.
There is a lot to criticize!
But criticism can be taken more seriously when it is not obscured by misunderstanding or misrepresentation (intentional, or not) of what LLM are and why they are not remotely "parrots" in the pejorative sense.
The technology, as technology, at the scale we are architecting it, is doing things we did not imagine would be witnessed in our lifetime, if ever. Dismissing that and denying it because of the career, industry, society, and civilization challenges that technology brings are existential, is bad argumentation or bad faith.
Both can be true at once.
I see no credible corroboration. More likely its folks having no more care for what they are doing than the bots themselves.
> Now, we all pay the consequence, to the tune of hundreds of thousands if not millions of dollars of wasted productivity from teams that have to deal with the resulting fall-out of this usage of “AI”.
People said the same about email spam ... until they engaged spam filters. CVE report slop is simply spam. Complaints are better directed at the filters, not the filtered.
I didn't get this at all from the parent. They're simply stating that LLMs aren't entirely trustworthy, and that the responsibility is ultimately ours, not the LLM's.
Each layer of attention can more through feature space “lit up weight clusters” in a way no other previous AI can. It can from that decode some rudimentary logic and world modeling and make deductions. Certainly better than any previous AI. Only a goof here would believe this wasn’t a serious advancement.
So don’t over sell it. But don’t sell it short with this “grrr in an engineer don’t threaten me with new tech” attitude.
This take is akin to teenage angsty takes and doesn’t really belong here.
I hate AI slop as much as the next guy but the amount of tribalism over AI is taking near-religious forms.
Nobody knows what intelligence is, therefore we don't know what does or does not possess it, therefore we don't know whether LLMs currently, or in the future, possess it.
Yes, LLMs can be stupid, guess what: so can I. That doesn't really change the argument at all.
I feel like I'm on a deja-vu from when DALL-E was released and everybody was fighting over whether AI can be creative yes or no. Same story, different words.
Intelligence, creativity: we have no idea what these words mean, and AI is helping us understand them better. That alone is an achievement of epic proportions. I am not joking here. Any computer scientist before 2015 would be absolutely blown away by what you can now do for 10 cents and an API call, yet somehow because of the tech-bro-iness of it all we get a tribal war over what is plainly visible in front of us:
LLMs are uncomfortably close to what we thought intelligent machines would look like
I've thought about this for quite some time now.
No. A human doesn't need to verify everything. And the argument is really simple: stochastic.
Think of self-driving cars: We can show today - based on evidence and real data - that self-driving cars are safer than human drivers. That's a fact and the consequences are clear, more self-driving cars, less human-driven cars, less accidents, less hurt people, less dead people.
Are the cars 100% safe and NEVER make a mistake? No. But they don't need to. Nothing is ever 100% (in the real world).
Now back to AI for software creation. "Review is the bottleneck because EVERYTHING must be judged by a human." No. It doesn't. We just need to build AI review systems, that will do reviews better than (or at least as good as) humans. The human review quality bar is far below 100%. Far far far. If we can show (likely in the next 12-24 months I think) that AI review quality is consistently above the human review quality - again, based on evidence, based on real data - then that's it, then there's no good reason to have humans review the code.
Yes, there will be another layer in the system, another level of abstraction that will/must end at the human boundary.
Flesh-based “brain” is able to use its vast corpus of inputs and calculate the most statistically likely output in a given situation. It is probabilistic, and when you are dealing with probabilities in a situation where certainties, not probabilities, matter, you’re going to get dinged on credibility massively when your flesh-based brain gets the probabilities wrong at best, or in this case, claims a line of code generates a vulnerability when it is, in fact, a code comment.
Humans are prediction engines. They are not Pure Intelligence, and shouldn’t not be treated in any form or fashion as if they possess pure intelligence. What bothers me about this entire situation is that presumably the folks that have relied on the flesh-based “brains” to generate these vulnerabilities knew (or should have known) enough about their "tool" to know this would happen, but did not: To err is to be human.
Now, we all pay the consequence, to the tune of hundreds of thousands if not millions of dollars of wasted productivity from teams that have to deal with the resulting fall-out of this over reliance on fallible “brains".
A human must verify everything another human presents as fact. Everything. If you don’t, we all pay the price. Using a human does not remove the onus of responsibility on the human being in charge, if anything they amplify it because humans work for peanuts in some countries, and can generate lots more output more quickly that needs to be verified by the humans in charge.
However, error margins are in the center of any engineering discipline. We cannot produce things measured with 100% accuracy. This is accepted fact. The focus is always not on eliminating errors, but on reducing them to acceptable minimum. With LLMs we should not expect an ideal logical thinker, but a process that may error sometimes, and we must design quality controls instead that push LLM outputs within acceptable margins. And it can work.
In the current AI mania, there's a lot of due diligence simply being ignored. Plenty of "Well humans make mistakes too!" going on here on HN too.