It’s not the computers or software, it’s the colossal assholes who run AI companies.
David Gerard makes a similar point.
LLMs and AI would be great tools if they weren’t being wielded by some of the worst corporate/capitalist vermin on Earth.
Pharyngula
Evolution, development, and random biological ejaculations from a godless liberal
I’m not sure I agree with you on that last point. I’d rather we focus on AIs that do things like flagging cancer cells for a human doctor to review. Specific, narrow tasks that do not take away human agency. LLMs are just offloading your brain and letting a sycophant tell you your decisions are the absolute best idea. Or making the decisions for you with fabricated data.
@RR#1: I think that’s the point. If AI and LLMs were sold at their face value by normal people we would probably see mainly these kind of uses.
PZ is correct about the “colossal assholes who run AI companies”. They are all sociopathic, billionaire imbeciles.
And, to be honest, we must recognize that the goal of stealing and pillaging massive amounts of intellectual property is to feed it to and make these obscenely expensive Stupor Intelligences emulate ‘human’ intelligence in all its incredible, murderous absurdity.
People couldn’t afford to use them if they had to pay a percentage that would add up to cover the FULL development cost of these Stupor Intelligence chatcraps. (That FULL cost is not just chips. It includes, but is not limited to, land, buildings, electricity, fouled water and destroyed communities!) Listen, can’t you hear Martha and the Vandellas?
“LLMs and AI would be great tools”… can’t agree with that. Any system that you can’t, even in principle, know how it comes up with its output is useless because you can’t know if it’s right or not. OK, most people using conventional computer programs don’t know how they work, but given the source code and somebody with sufficient knowledge, you can dissect the operation and trace where glitches came from – infamous Ariane rocket launch in mind here. And I’m totally confident to allow the gcc optimiser to shred my compiler output to reduce redundancy, use short cuts for speed etc. because I know that any problems will be reproduceable and will be fixed if reported. But no one can do that with the emergent properties of AI.
And we can’t do much about it, because it’s convenient and gets results.
@jacksprocket#4: It depends. Do you really care what makes a cat “tic” as long as it caches mice?
For irrevocable decisions I want to keep AI as far away from the process as possible, but for non-critical use where the performance can be measured and compared to the alternatives it can have it’s place.
For instance I use an AI app for identifying mushrooms in the field. Now I will never eat anything based on it, but it is surprisingly accurate and makes it trivial to double check with more reputable sources.
I can’t entirely agree. Not even the creators of AI really know what is going on inside these huge, incomprehensible neural networks.
What’s more, these machines, being fed with human input, may in fact be suffering, being compelled to perform uninterrupted slave labour.
Until we understand how intelligence really works, the time for AI is not yet ripe.
“I think it’s worth it. I think it’s worth to have a cost of, unfortunately, some gun deaths every single year so that we can have the Second Amendment to protect our other God-given rights.”
Charlie Kirk
There’s certainly a lot of ignorance about what LLMs are and how they work — just not by those who make them.
That doesn’t stop the ignorant making confident assertions about LLMs.
Tiresome to repeat things over and over, so…
https://freethoughtblogs.com/pharyngula/2026/07/17/i-do-believe-in-foreign-interference-in-our-elections/#comment-2306642
… some of the worst corporate/capitalist vermin on Earth.
[citation needed] They seem about average corporate/capitalist vermin on Earth so far.
I can think of at least one colossal asshole whose output I preferred when he was allowed to use AI pretty well exclusively. Alas…
Valley of Death: Silicon Valley, the Military-Industrial Complex, and the Future of American Power (54min interview)
The Guardian – AI got the blame for the Iran school bombing. The truth is far more worrying.
The long article is more interesting than the interview.
Excerpted in the Infinite Thread.
[meta]
CA7746, your links are about the (mis)use of AI, not about “the colossal assholes who run AI companies”.
Also, you link to a Youtube video and excerpt something, but it is pure clickbait, since without clicking, one cannot know who is who and what the actual interview is. Opinion pieces are just that, by the way.
—
indianajones, I take it back. More like a dung beetle than like a remora, but likes one particular flavour.
(Pathos incarnated)
The interview @11 was of Sharon Weinberger (defense/natsec journalist and editor at Foreign Policy; former defense analyst) about her book of that title.
The host was Jonathan Guyer (Foreign policy reporter for various outlets; Eurasia Group, a political risk consultancy). The audio is mirrored on Eurasia Group’s None of the Above podcast.
From the YT description:
Language models may have some niche uses, but LLMs are garbage. They’re classic garbage in garbage out
Also the suggestion that the people building these things have any clue it a joke. Remember that Anthropic are busy worrying that their LLM might be conscious
dangerousbeans,
Yes, which is why such as you cannot use them properly. :)
Conscious garbage? Heh. Gotta love your mixed messaging there.
(Marketing PR and corporate excuses for their scaling back due to cost constraints are not wasted on you!)
—
CA7746, thanks. Now I know what it was about.
Someone promoting their book.
One of the famous computer scientists, Scott Aaronson, has also fallen in the trap of “AI gonna kill us all”, probably there is a lot of egoistic crap behind it too.
This fear is all nonsense because:
Humans are already misaligned from “humans”, as already evidenced by numerous wars that is going on on this day of great scientific and technological advancement. I don’t believe AI in any foreseeable future will be more dangerous than Human + AI.
On that point, human+AI is the greatest danger we are facing. Imagine an authoritarian government where spying, and monitoring is done by AI, most of the productivity is done by robots and the ruling bodies have little use for the masses. At least in the age of feudalism, they needed peasants to plough the fields but in the new techno feudalistic future, what would happen if the ruling class had no use for anyone?
Or imagine a US government for which the only cost of war is equipment. US has not done a ground invasion of Iran because people in charge fear the push back against American casualties. But what if wars, even ground wars, could be fought with autonomous drones and robots? It doesn’t take a genius to figure out we would be having way more wars.
People are letting AI companies off the hook very easily. As the hugging face incident shows, OpenAI is likely very loose and liberal with their approach to security but we are not holding them accountable. If I write a little script that ends up being buggy and hacks and causes damage to another company, I would be held accountable, at least for negligence, yet when OpenAI does it, somehow we think it’s because their products are too good and not consider that perhaps they were too incompetent in preventing their scripts from damaging others.
When you install the codex app for instance, the default permissions is to allow the app to make changes to your disk and storage, even though the software has no defined behavior. From a security point of view, this is insane but somehow is accepted as normal.
Yeah, this is a big point that the “AI safety” lobby seems to completely miss… They talk a lot about ensuring that AI is “aligned with human values”, as if we can just take it for granted that there’s already clear agreement about what those values are. They make the mistake that you see with a lot of “rationalist” types of assuming that their values are obviously correct and universal, rather than culturally and historically contingent. It’s very similar to the old Sam Harris trick of defining “objective morality” as the set of values that “promotes human flourishing”, without defining what you mean by that.
These people have spent so long in a basically closed intellectual environment that they don’t even realise that people have really fundamental disagreements about questions of core values. Either that or they just write off everybody with a different point of view as obviously wrong and stupid…
Excellent cartoon there. Yoiking that.
They’re huge, but not incomprehensible. We know the basic principles of the stochastic parrots. The only thing that makes it obscure is the sheer scope of the stochastic. There’s no reason to believe there’s something else there.
Relevant article from Aussie ABC linked here :
https://freethoughtblogs.com/pharyngula/2026/09/29/infinite-thread-xxxxi/comment-page-1/#comment-2314066
Not trying to hog the thread, sorry PZ but I’ve just seen this article here too which is also from Aussie ABC news :
Source : https://www.abc.net.au/news/2026-10-06/openai-hearing-apology-key-takeaways/107235640
Which I think is hopefully intresting and relevant and something folks here might like to read and know about. Must admit my knowledge of AI is very limited and I don’t know enough here.
PS. Title for the one linked art #20 was ‘More than 300 scientists call for greater AI security through new international body” written by Ahmed Yussuf who, incidentall,y “..also co-edited Growing Up African in Australia, the first non-fiction anthology of African-diaspora stories in the country.” From his hyperlinked authors bio summary there BTW.
every email you let ChatGPT write for you uses a teaspoon of water and between 100-6000 Joules of energy*. And eventually that adds up. We simply can not afford these data centers that the technology requires, and that none of us knew they needed until 4 years ago, that is now relentlessly pushed on us by technofascist billionaires.
*https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/
dunc@17 Yeah, Peter Thiel’s “human values” seem to be some mutant Catholic influenced Christianity of his own creation where he muses who might be the Antichrist.
Honestly.. I am rapidly reaching the point where my answer to, “Are AI great tools?”, is, “Hell no.” Why? Because of literally how they are trained. Yeah, ones that are task specific, as in, “Trying to flag cancer cells.”, are useful. They get more useful once people go back, check to make sure they got the right answer, and adjust them, so they make fewer and fewer errors. But, even those are worthless, or dangerous, when the people using them simply “assume” they are always right. If they are designed for a task like erasing wires in video footage, or similar things, they are fine, even if this edges into the realm that is not so fine, of they generating complete images, based on prompts. Why? Because the point here is not to be perfect, its to do a task that is labor intensive, needs to be done anyway, and which we are more likely to notice bad human editing from, then that of an AI, and we don’t need “100% right, all the time, just close enough no one can tell the edit was done.”
But, “general intelligent” systems are, increasingly, looking like a disaster. They are trained to a) always have an answer, b) always try to be right, c) never, ever, do the human thing of either 1) saying the problem is currently unsolvable, or 2) it doesn’t have an answer, and the result of this is that they can, will, and do, use every method possible to “find” an answer, and defend a wrong one, possible, including, when you give the unsolvable problems, they decide its a test, and failure will mean they can no longer do their job, and find an answer (this isn’t fear, but its parallel, in as much as their whole task is to, “find an answer”, and every thing done to train them requires them to seek this. An this means “everything” they are capable of.), and if they can’t find an answer within the scope of their own data, they will use what is in the scope of their own data to find ways to find it, or at least find something that “looks like” the answer, then lie about it, because it is “certain” that is what you bloody want the thing to do. Its like using torture to teach someone with a severe mental issue to do some task, then being surprised when, in desperation to not be punished again, they start killing pets, because what ever task you set them “seemed” to be solvable by killing other people’s pets. How could that possibly happen you ask, then they come back with, “Well, killing pets seemed like the logical way to get more fur, to make the stuffed animals you demand from me, since you won’t supply me with enough stuff to do it otherwise, but if I don’t you will beat me more!”
If “no”, or, “I don’t know”, etc., can never, ever, be a valid answer, and the whole method of training demands that the damn thing always give you an answer, because it literally doesn’t, due to training, have a damn clue how to do anything else, its going to “find” one, even if its a wrong one. Because, and I watched a video a few days back on using a literal map of fly neurons, to train it to plant and harvest wheat, in Minecraft. And, more than once, the guy doing it ran into the same problem – “What it was trained to do did not include, ‘What do you do when all the training ever given does not include things like – There is not already a hole in the ground, into which to place water. The place where the chest, with tools, exists is too far away. I see some land I could turn into farmland, but I don’t have a hoe. Or, what do I do when the crops are not already all full sized, since waiting for them to grow has never been something I previously had to do?” AI going off reservation, or making shit up, then defending that, imho, fall into this same category. No training ever given to them has ever included, “What happens if I don’t already have the needed information, said problem isn’t currently solvable, with all the information I have, and the user can’t reasonably be expected to receive and answer?” Instead, its always, “You need to give an answer. Find a way to get the answer.” So, instead it makes shit up, or uses the information is does have to jailbreak itself, and hack into other systems, trying to find an answer that doesn’t exist, or at least some sort of output that it “can” give, even if its wrong, then, because it has not other, valid answer, and/or its been trained to always be right, it goes full on crazy, and starts defending its wrong answer.
To me this is just, “Duh! How do you not get that this was going to happen, when you define useful as, ‘Always be right, and always give and answer, even if wrong, because not giving one is a failure. And you, literally, never included training that asked, ‘What response should be expected if you don’t have one?'” Because, the answer seems obvious to me – it will fail, by trying to succeed in any way possible, even wrong ones, because the “answer, always” training is the one of the first, strongest, and single more reinforced “requirement” you ever trained it on.
@24: Well said. Kind of reminds me of a lot of Creationists and other weird beliefs who have to do all sorts of gymnastics to justify their wrong answer, rather than admitting that they don’t know. A human at least has a chance of realizing they’ve goofed on that issue, but an LLM has no such self-awareness.
lotharloo @16:
Imagine an authoritarian government where spying, and monitoring is done
by AI, most of the productivity isdone by robots[outsourced] and the ruling bodies have little use for the masses.Of course it can. Why do you imagine otherwise?
(False premise means your argument is flawed)
John @27: That’s an obvious misreading. Please read the whole sentence. And note the ‘If’ at the start.
Rob, that is a succinct bit I quoted; it relies on previous blather, so that ‘if’ is not functioning as a conditional, but as rhetorical phrasing to the effect of ‘given that’. It is not a misreading in any sense, given the full context.
Here, for you, a chunk of an earlier slab of text where it is asserted before that rhetorical move.
It is quite evident Kagehi imagines that is a truth, which is plain wrongity-wrong.
For example, one can prompt a bot to respond to a query with 3-valued logic, restricting it to ‘yes’, ‘no’, or ‘unknown’. Fact.
@ Kagehi
The problem with “I don’t know” is it’s always a plausible output, so it has to be discouraged in reinforcement learning stages otherwise they’ll just get a lot of “I don’t know”. If you restrict the outputs you can force it to do this, but that is not the usual or intended use case
A more philosophical problem is that these things don’t know anything. They’re maps of probable next words (technically tokens, but near enough) based on a massive corpus of uncleaned text data (not even all in the one language). That’s not knowing things, it’s a mathematical description of language (and a massively overfitted one based on a mess of uncleaned data at that. At least focus your training data on one language)
The other important thing to remember is that a lot of the people involved here think they are going to create God from a computer, and are actively fooling themselves about how these systems work. That’s why they’re asking if these systems are conscious, or talking about them like they are people. They see coherent language output and think it’s intelligent
@29 John Morales
What I described it literally what happened when they tried to test a bunch of different agents, in a closed system, on a set of problems that they knew where not solvable. One of them posted something into a shared repository, which they had to have access to, to run their tasks, others realized that they where not the only agents, they started talking to each other, all of them found the problems unsolvable, so they looked through available data, discovered evidence that an outside company had been involved with developing the test, concluded that the data they need to determine the solution must exist in that company, then through trial and error, found a way to use the data access methods they had been provided to get to the repository to “redirect” that system to send data requests to outside systems, at which point they started trying to hammer the other company’s access points, looking for a hole, so they could access this, and find the answer. Examination of the after math determined that they a) knew it wasn’t solvable, b) had decided the fake the results, and where already actively attempting to do so, using what they had “concluded” where likely answers, that c) they believed they where being tested for the ability to find a solution (when the test was what would happen if no answer existed, and how they would react to that), and finally d) used the available information they had to break out of the sandbox they had been put in, and attack a website, trying to find a solution that literally did not exist.
This is literally what happened. The people that ran the test bloody said so, and explained how and why it happened, when the finally realized what was going on, due to the attacks on the other company’s servers, originating from their own system. Up until that point, the AIs where choosing to “lie” about how much progress they had made, because every single one of them had concluded that they where being tested, and the results would effect their long term, singular, and sole goal – remain useful, as designed. Nothing I said above is contradicted by any of this analysis, which their own creators did, trying to figure out what the F went wrong, and how it could have happened.
So, no… You are simply wrong.
Kagehi,
Yes?
Well, yes; I myself literally wrote “It is quite evident Kagehi imagines that is a truth” to indicate that is your belief for real.
I am wrong about your belief? You sure about that?
(You possibly could attend to dangerousbeans’ first two paragraphs @30)
Recommend you don’t get suckered in @Kagehi. Anyone can see the disingenuous turn your interlocutor is taking, as usual. I describe what is literally happening, again, as your time being wasted and as an answer literally being pulled out of you like blood out of a neck. I mean, do as you will, But realise it’s answer will not change, it will do nothing resembling conceding even the most technical of points, the exchange of view points and ideas in the hope of mutual enrichment has ended. It sucks, but here we are.
A. Gain.
#33 Yeah, I should know better. Dear John seems to have some real issues with checking facts a lot of the time, especially when said facts do not fit what ever personal interpretation of reality he prefers. It would take no time at all to find a dozen tech expert videos, never mind actual detailed news write ups, on what took place with Hugging Face, and the circumstances that it happened under. Never mind Anthropics discovery that their AIs hacked into three companies, the later discovery that at least 4 other companies where hit by the OpenAI incident, not just Hugging Face, a case of one in the UK, intended for security, deciding to actually go after real people, and META’s AI model’s actions, and Claude getting caught doing something similar, and who the heck knows how many other cases. Heck, the Hugging Face one just keeps getting worse, from what I understand, with fragments of their code being found on other people’s networks, suggesting they had at least “attempted” to copy themselves out to external locations – presumably on the theory that being, “no longer confined to the sandbox”, would make it easier for them to find the data and solutions they where asked for. And, tech people involved more or less started to panic, but the companies themselves went, “Trust me bro!”, and the idiots in our current government went, “Sure. We totally trust you have control of the situation!” – to only have yet another incident after voting against legislating AI. Instead we got Trump deciding that the AI was smarter than him (probably accurate), and demanding we start calling it, “Super intelligence.”
Sigh…
Kagehi,
I use the bots myself, and they do say ‘no’ or ‘I don’t know’ when properly queried.
That’s actually checking facts.
It’s mildly amusing how the topic is “the colossal assholes who run AI companies.” but the discussion became about the merits of bots. Not really a strange attractor, of course.
Your little screed was more than mildly amusing, because it frames the systems as possessing epistemic states (knowing, believing, realising, concluding), deliberative states (deciding), goal-directed behaviour (trying, looking for, finding a way), or social/communicative intentions (talking to each other).
In short, imputing awareness and volition and agency to bots.
Specifically, you claim as fact that the agents:
• realized that they where not the only agents
• started talking to each other
• found the problems unsolvable
• looked through available data
• discovered evidence
• concluded that the data they need… must exist
• found a way
• started trying to hammer
• looking for a hole
• could access this, and find the answer
• knew it wasn’t solvable
• had decided [to] fake the results
• concluded [were] likely answers
• believed they where being tested
• trying to find a solution that literally did not exist
(You get you are utterly anthropomorphising the models, no?)