LLMs are a 400-year-long confidence trick
tomrenner.com
tomrenner.com
I'm seeing legitimate 10x gains because I'm not writing code anymore – I'm thinking about code and reading code. The AI facilitates both. For context: I'm maintaining a well-structured enterprise codebase (100k+ lines Django). The reality is my input is still critically valuable. My insights guide the LLM, my code review is the guardrail. The AI doesn't replace the engineer, it amplifies the intent.
Using Claude Code Opus 4.5 right now and it's insane. I love it. It's like being a writer after Gutenberg invented the printing press rather than the monk copying books by hand before it.
The LLM marketing exploits fear and sympathy. It pressures people into urgency. Those things can be shown and have been shown. Whether or not the actual LLM based tools genuinely help you has nothing to do with that.
Of course it is a little more nuanced than this and I would agree that some of the marketing hype around AI is overblown, but I think it is inarguable that AI can provide concrete benefits for many people.
Simplifying the hype into 2 threads, the first is that AI is an existential risk and the second is the promise of “reliable intelligence”.
The second is the bugbear, and the analogy I use is factories and assembly lines vs power tools.
LLMs are power tools. They are being hyped as factories of thoughts.
String the right tool calls, agents, and code together and you have an assembly line that manufactures research reports, gives advice, or whatever white collar work you need. No Holidays, HR, work hours, overhead etc.
I personally want everyone who can see why this second analogy does not work, to do their part in disabusing people of this notion.
LLMs are power tools, and impressive ones at that. In the right hands, they can do much. Power tools are wildly useful. But Power tools do not make automatically make someone a carpenter. They don’t ensure you’ve built a house to spec. Nor is a planar saw going to evolve into a robot.
The hype needs to be taken to task, preferably clinically, so that we know what we are working with, and can use them effectively.
Yes, yes you can. As I’ve mentioned elsewhere on this thread:
> When a con man sells you a cheap watch for an high price, what you get is still useful—a watch that tells the time—but you were also still conned, because what you paid for is not what was advertised. You overpaid because you were tricked about what you were buying.
LLMs are being sold as miracle technology that does way more than it actually can.
It's not LLMs they care about, it's datacenter ownership. US political norms empower owners. If you think of a DC as a mega church and remote users the disciple, it makes the desired network effect obvious. That is leveraged to sway Congress and states.
These tech projects are not intended for users. They're designed to gain confidence of politicians, preferential political support.
Gen pop is not the market. DC is.
Most peoples individual data crunching problems can be resolved with a TI graphing calculator.
Big Tech convinced Congress that culture of helpless consumers of their data center outputs is simpler and will lead humanity to a forever growth future!... nevermind they will all be dead, unable to verify.
A con trick that worked great on older, more religious leaning Americans. One that's not working so well on the younger generation who know how these systems work.
Exactly, and that's the whole point. If you lack genuine mathematical reasoning skill, a calculator won't help you at the IMO. You might as well bring a house plant or a teddy bear.
But if you bring a GPT5-class LLM, you can walk away with a gold medal without having any idea what you're doing.
Consequently, analogies involving calculators are not valid. The burden of proof rests firmly on the shoulders of those who claim that an LLM couldn't invent new mathematical techniques in response to a problem that requires it.
In fact, that appears to have just happened (https://news.ycombinator.com/item?id=46664631), where an out-of-distribution proof for an older problem was found. (Meta: also note the vehement arguments in that thread regarding whether or not someone is using an LLM to post comments. That doesn't happen without intelligence, either.)
I would guess it's only a matter of days before that proof, or one very similar, is found in the training data, if that hasn't happened already, just as has been the case every time.
No fundamental change in how the LLM functions has been made that would lead us to expect otherwise.
Similar "discoveries" occurred all the time with the dawn of the internet connecting the dots on a lot of existing knowledge. Many people found that someone had already solved many problems they were working on. We used to be able to search the web, if you can believe that.
The LLMs are bringing that back in a different way. It's functional internet search with an uncanny language model, that sadly obfuscates the underlying data while making guesswork to summarize it (which makes it harder to tell which of its findings are valuable, and which are not).
It's useful for some things, but that's not remotely what intelligence is. It doesn't literally understand.
>* if you bring a GPT5-class LLM, you can walk away with a gold medal without having any idea what you're doing.*
My money won't be betting on your GPT5-class business advice unless you have a really good idea what you're doing.
It requires some (a lot of) intelligence and experience to usefully operate an LLM in virtually every real world scenario. Think about what that implies. (It implies that it's not by itself intelligent.)
You cannot bluff, trick, or "market" your way through a test like that.
If scoring highly on an exam implies intelligence then certainly I'm not intelligent and the Super Nintendo from the 90s is more sentient than myself, given I'm terrible at chess.
I personally don't agree with that definition, nor does any dictionary I'm familiar with, nor do any software engineers with whom I'm familiar, nor any LLM specialists, including the forefront developers at OpenAI, xAI, Google, etc. as far as I'm aware.
But for some reason (it's a very obvious reason $$$), marketers, against the engineers' protest, appear to be claiming otherwise.
This is what you're up against and what you'll find the courts, and lawyers, will go by when this comparison comes to a head.
In my opinion, I can't wait for this to happen.
Thrilled to know if I shouldn't wait for that. If you're directly involved with some credible research to the contrary, I would love to hear more.
But IMO, in this case at least, has nothing to do with intelligence. It's performing a search against its own training data, and piecing together a response in line with that data, while including the context of the search term (aka the question). This is run through a series of linear regressions, and a response is produced. There is nothing really groundbreaking here, as best I can tell.
It is unsupportable to claim that ML researchers at leading labs share your opinion. Since roughly 2022, they understand that they are working with systems capable of reasoning: https://arxiv.org/abs/2205.11916
The article you linked discussed reasoning. That's really cool. But, consider that we can say that a chess game computer opponent is reasoning. It's using a preprogrammed set of instructions to predict out to some number of possible moves ahead, and choosing the most reasonable. A calculator, essentially, it is in fact reasoning. But that doesn't have much to do with intelligence. As we read in the dictionary, intelligence implies understanding, and we certainly can't say that the Chess Masters opponent from the Super Nintendo literally understands me, right?
More to the point, I don't see that any LLM has thus far exhibited remotely any inkling of understanding, nor can it. It's a linear regression calculator. Much like a lot of TI84 graphing calculators running linear algebraic functions on a grand scale. It's impressive that basic math can achieve results across word archives that sound like a person, but it's still not understanding what it outputs, and really, not what it inputs beyond graphing it algebraically either.
It doesn't literally understand. So, it is not literally intelligent, and it will require some huge breakthroughs to change that. I very much doubt that such a discovery will happen in our lifetime.
It might be more likely that the marketers will succeed in revising the dictionary. We've seen often times that if you use words wrong enough, it becomes right. But so far at least, that hasn't happened with this word.
Let's say a kid who's not unusually gifted/talented at math somehow ends up at the International Math Olympiad. Smart-enough kid, regularly gets 4.0+ grades in normal high school classes, but today Timmy got on the wrong bus. He does have a great calculator in his backpack -- heck, we'll give him a laptop with Mathematica installed -- so he figures, why not, I'll take the test and see how it goes. Spoiler: he doesn't do so well. He has the tools, but he lacks understanding of how and when to apply them.
At the same time, the kid at the next desk also doesn't understand what's going on. She's a bright kid from a talented family -- in fact Alice's old man works for OpenAI -- but she's a bit absent-minded. Alice not only took the wrong bus this morning, but she grabbed the wrong laptop on the way out the door. She shrugs, types in the problems, and copies down what she sees on the screen. She finishes up, turns in the paper, and they give her a gold medal.
My point: any definition of "understanding" you can provide is worthless unless it can somehow account for the two kids' different experiences. One of them has a calculator that does math, the other has a calculator that understands math.
I very much doubt that such a discovery will happen in our lifetime.
So did I, and then AlphaGo happened, and IMO a few years later. At that point I realized I wasn't very good at predicting what was and was not going to be possible, so I stopped trying.
Don't stop trying. Predictability is an indicator of how well a theory describes the universe. That's what science is.
The engineers have long predicted this stuff. LLM tech isn't really new. The size and speed of the machines is new. The more you understand about a topic, the better your predictions.
Indeed.
I feel that my expectations surrounding "AI" are much more realistic than they were before building the tools.
If you haven't already, it's very much worth giving them a run through.
It may be extremely dangerous to release. True. Even search engines had the potential to be deemed too dangerous in the nuclear pandoras box arguments of modern times. Then there are high-speed phishing opportunities, etc.
It may be an essential failure to miss the boat. True. If calculators were upgraded/produced and disseminated at modern Internet speeds someone who did accounting by hand would have been fired if they refused to learn for a few years.
Its communication builds an unhealthy relationship that is parasitic. True. But the Internet and the way content is critiqued is a source of this even if it is not intentionally added.
I don't like many people involved and I don't think they will be financially successful on merit alone given that anyone can create a LLM. But LLM technology is being sold by organic "con" that is how all technology such as calculators end up spreading for individuals to evaluate and adopt. A technology everyone is primarily brutally honest about is a technology that has died because no one bothers to check if the brutal honesty has anything to do with their own possible uses.
They literally are. Sam Altman has literally said multiple times this tech will cure cancer.
My company just released a year-long productivity chart covering our shift to Claude Code, and overall, developer productivity has plummeted despite the self-reported productivity survey conveying developers felt it had shot through the roof.
Couple that with engineers across the board mentioning that they feel like they're losing proficiency in an understanding of the codebase and where things are.
The trouble with highly productive specialists is that they produce a ton of high quality results where the demand is not really there and has to be artificially made. Even if you find enough work for them it often means the incremental cases are things you wouldn't have bothered with. A specialist branching to work slowly in related yet further related areas is a lot more value and can work with an oracle so flawed that it barely beats chance..
With juniors it is much more complex, but they have always been a useless consideration in productivity. Not having them has always been highly productive in the short term but has long term consequences.
The line becomes a lot blurrier when you work on non trivial issues.
A Django app is not particularly hard software, it's hardly software but a conduit from database to screens and vice-versa; which is basic software since the days of terminals. I'm not judging your job, if you get paid well for doing that, all power to you. I had a well paying Laravel job at some point.
What I'm raising though is the fact that AI is not that useful for applications that aren't solving what has been solved 100 times before. Maybe it will be, some day, reasoning that well that it will anticipate and solve problems that don't exist yet. But it will always be an inference on current problems solved.
Glad to hear you're enjoying it, personally, I enjoy solving problems, not the end result as much.
Also, almost all problems are composite problems where each part is either prior art or in itself somewhat trivial. If you can onboard the LLM onto the problem domain and help it decompose then it can tackle a whole lot more than what it has seen during pre- and post-training.
I have two stories, which I will attempt to tie together coherently in response.
I'm making a compiler right now. ChatGPT 4 was helpful in the early phases. Even back then, its capabilities with reading and modifying the grammar and writing boilerplate for a parser was a real surprise. Today 5.2-Codex is iterating on the implementation and specification as I extend the language and fill in gaps in the compiler.
Don't get me wrong, it isn't a "10x" productivity gain. Not even close. And the model makes decisions that I would not. I spent the last few days completely rewriting the module system that it spit out in an hour. Yeah, it worked, but it's not what I wanted. The downsides are circumstantial.
25 years ago, I was involved in a group whose shared hobby was "ROM hacking". In other words, unofficial modification of classic NES and SNES games. There was a running joke in our group that went something like this: Someone would join IRC and ask for an outlandish feature in some level editor that seemed hopelessly impossible at the time. Like generating a new level with new graphics.
We would extrapolate the request to adding a button labeled "Do My Hack For Me". Good times! Now this feature request seems within reach. It may forever be a pipe dream, who knows. But it went from "unequivocally impossible" to "ya know, with the right foundation and guidance, that might just be crazy enough to work!" Almost entirely all within the last 10 years.
I think the novelty or creativity criticism of AI is missing the point. Using these tools in novel or creative ways is where I would put my money in the coming decade. It is mind boggling that today's models can even appear to make sense of my completely made up language and compiler. But the job postings for adding those "Do My Hack For Me" buttons are the ones to watch for.
That's not how book printing works and I'd argue the monk can far more easy create new text and devise new interpretations. And they did in the sidelines of books. It takes a long time to prepare one print but nearly just as long as to print 100 which is where the good of the printing press comes from. It's not the ease of changing or making large sums of text, it's the ease of reproducing and since copy/paste exist it is a very poor analogue in my opinion.
I'd also argue the 10x is subject/observer bias since they are the same person. My experience at this point is that boilerplate is fine with LLMs, and if that's only what you do good for you, otherwise it will hardly speed up anything as the code is the easy part.
It’s like arguing that the piano in the room is out of tune and not bothering to walk over to the piano and hit its keys.
They don't have time to check more stuff as they are busy with their life.
People who did check the stuff don't have time in life to prove to the ones that argue "in exactly whatever the person arguing would find useful way".
Personally like a year ago I was the person who tried out some ChatGPT and didn't have time to dabble, because all the hype was off putting and of course I was finding more important and interesting things to do in my life besides chatting with some silly bot that I can trick easily with trick questions or consider it not useful because it hallucinated something I wanted in a script.
I did take a plunge for really a deep dive into AI around April last year and I saw for my own eyes ... and only that convinced me. Using API where I built my own agent loop, getting details from images, pdf files, iterating on the code, getting unstructured "human" input into structured output I can handle in my programs.
*Data classification is easy for LLM. Data transformation is a bit harder but still great. Creating new data is hard so like answering questions where it has to generate stuff from thin air it will hallucinate like a mad man.*
Data classification like "is it a cat, answer with yes or no" it will be hard for latest models to start hallucinating.
Do I now get the right to talk badly about all LLM coding, or is there another exercise I need to take?
Yes, the technology is interesting and useful. No, it is not a “10x” miracle.
In truth neither claims are reasonable, but because of the door in the face, the victim is more susceptible the the latter claim. Without the more outrageous claim it is unlikely the victim would have believed the latter claim.
In reality, both "AGI" and "100x miracle" AND the "10x miracle" are all outrageous claims, and I call bullshit on all of them.
How long have you been in the industry?
This does not seem a revolution compared with database standardization, abandonment of assembly for most coding, introduction of game engines, etc.
I see a lot of hype for LLMs from people that do not have the experience to compare them to anything else.
Then why is half of the big tech companies using Microsoft Teams and sending mails with .docx embedded in ?
Of course marketing matters.
And of course the hard facts also matters, and I don't think anybody is saying that AI agents are purely marketing hype. But regardless, it is still interesting to take a step back and observe what marketing pressures we are subject to.
Self-reports on this have been remarkably unreliable.
How do I know? Because I am testing it, and I see a lot of problems that you are not mentioning.
I don’t know if you’ve been conned or you are doing the conning. It’s at least one of those.
It was something like this:
"We think we are building Ultron but really we are building the Iron Man suit. It will be a technology to amplify humans, not replace them"
How do you avoid this turning into spaghetti? Do you understand/read all the output?
This may be true. The commenter may "believe in this tech" based on his experimentation with it
But the majority of sentences following this statement ironically appear to be "marketing hype" or "someone telling [us] it's good":
1. "The productivity gains I'm seeing right now are unprecedented."
2. "Even a year ago this wouldn't have been possible, it really feels like an inflection point."
3. "I'm seeing legitimate 10x gains because I'm not writing code anymore - I'm thinking about code and reading code."
4. "Using Claude Code Opus 4.5 right now and it's insane."
5. "It's like being a writer after Gutenberg invented the printing press rather than the monk copying books by hand before it."
The "framing" in this blog post is not focused on whether "this tech" actually saves anyone any time or money
It is focused on _hype_, namely how "this tech" is promoted. That promotion could be intentional or unintentional
N.B. I am not "agreeing" with the blog post author or "disagreeing" with the HN commenter, or vice versa. The point I'm making is that one is focused on whether "this tech" works for them and the other is focused on how "this tech" is being promoted. Those are two different things, as other replies have also noted. Additionally, the comment appears to be an example of the promotion (hype) that its author claims is not the basis for his "belief in this tech"
I think the use of the term "belief" is interesting
That term normally implies a lack of personal knowledge:
151 "Belief" gcide "The Collaborative International Dictionary of English v.0.48"
Belief \Be*lief"\, n. [OE. bileafe, bileve; cf. AS. gele['a]fa. See {Believe}.]
1. Assent to a proposition or affirmation, or the acceptance of a fact, opinion, or assertion as real or true, without immediate personal knowledge; reliance upon word or testimony; partial or full assurance without positive knowledge or absolute certainty; persuasion; conviction; confidence; as, belief of a witness; the belief of our senses. [1913 Webster]
Belief admits of all degrees, from the slightest suspicion to the fullest assurance. --Reid. [1913 Webster]
2. (Theol.) A persuasion of the truths of religion; faith. [1913 Webster]
No man can attain [to] belief by the bare contemplation of heaven and earth. --Hooker. [1913 Webster]
4. A tenet, or the body of tenets, held by the advocates of any class of views; doctrine; creed. [1913 Webster]
In the heat of persecution to which Christian belief was subject upon its first promulgation. --Hooker. [1913 Webster]
{Ultimate belief}, a first principle incapable of proof; an intuitive truth; an intuition. --Sir W. Hamilton. [1913 Webster]
Syn: Credence; trust; reliance; assurance; opinion. [1913 Webster]
151 "belief" wn "WordNet (r) 3.0 (2006)"
belief
n 1: any cognitive content held as true [ant: {disbelief}, {unbelief}]
2: a vague idea in which some confidence is placed; "his impression of her was favorable"; "what are your feelings about the crisis?"; "it strengthened my belief in his sincerity"; "I had a feeling that she was lying" [syn: {impression}, {feeling}, {belief}, {notion}, {opinion}]
151 "BELIEF" bouvier "Bouvier's Law Dictionary, Revised 6th Ed (1856)"
BELIEF. The conviction of the mind, arising from evidence received, or from information derived, not from actual perception by our senses, but from. the relation or information of others who have had the means of acquiring actual knowledge of the facts and in whose qualifications for acquiring that knowledge, and retaining it, and afterwards in communicating it, we can place confidence. " Without recurring to the books of metaphysicians' "says Chief Justice Tilghman, 4 Serg. & Rawle, 137, "let any man of plain common sense, examine the operations of, his own mind, he will assuredly find that on different subjects his belief is different. I have a firm belief that, the moon revolves round the earth. I may believe, too, that there are mountains and valleys in the moon; but this belief is not so strong, because the evidence is weaker." Vide 1 Stark. Ev. 41; 2 Pow. Mortg. 555; 1 Ves. 95; 12 Ves. 80; 1 P. A. Browne's R 258; 1 Stark. Ev. 127; Dyer, 53; 2 Hawk. c. 46, s. 167; 3 Wil. 1, s. 427; 2 Bl. R. 881; Leach, 270; 8 Watts, R. 406; 1 Greenl. Ev. Sec. 7-13, a.
TLDR: everyone thought AI made people faster, including those who did the task, both before and after doing it. However, AI made people slower at doing the task.
But it's clear the LLM's have some real value, even if we always need a human-in-the-loop to prevent hallucinations it can still massively reduce the amount of human labour required for many tasks.
NFT's felt like a con, and in retrospect were a con. The LLM's are clearly useful for many things.
When a con man sells you a cheap watch for an high price, what you get is still useful—a watch that tells the time—but you were also still conned, because what you paid for is not what was advertised. You overpaid because you were tricked about what you were buying.
LLMs are useful for many things, but they’re also not nearly as beneficial and powerful as they’re being sold as. Sam Altman, while entirely ignoring the societal issues raised by the technology (such as the spread of misinformation and unhealthy dependencies), repeatedly claims it will cure all cancers and other kinds of diseases, eradicate poverty, solve the housing crisis, democracy… Those are bullshit, thus the con description applies.
* LLMs are a useful tool in a variety of circumstances.
* Sam Altman is personally incentivised to spout a great deal of hyped-up rubbish about both what LLMs are capable of, and can be capable of.
The dependency here is that if Sam Altman is indeed a con man, it is reasonable to assume that he has in fact conned many people who then report an over inflated metric on the usefulness of the stuff they just bought (people don’t like to believe they were conned; cognitive dissonance).
In other words, if Sam Altman is indeed a con man, it is very likely that most metrics of the usefulness of his product is heavily biased.
Theranos on the other hand… That was a con and the founder was prosecuted.
And again, Sam Altman has a history of deceit.
https://www.technologyreview.com/2022/04/06/1048981/worldcoi...
https://www.buzzfeednews.com/article/richardnieva/worldcoin-...
The latest slogan i can remember is “open happiness”.
I like fizzy sugar drinks as much as the next guy but Coca-Cola is not liquid happiness.
Edit: Dropbox marketing slogan is “for all things worth saving”
The promises of Theranos and LLMs are concrete measurable things we can evaluate and report where they succeed, fall short, or are lies.
Theranos was an outright scam that never produced any results, whereas LMMs might not have (yet?) lived up to all the marketing promises (you might call them slogans?) they made, but they definitely provided some real measurable value.
You’re being ridiculous. These are not slogans, and no one arguing in good faith would claim they are.
https://youtu.be/l0K4XPu3Qhg?t=60
> but they definitely provided some real measurable value.
Something which provides value can still be a scam.
https://news.ycombinator.com/item?id=46614364
I addressed all your points some eight levels up the thread.
We had an "essential" reporting function in the business which was done in Excel. All SMEs seem to have little pockets of this. Hours were spent automating the task with VBA to no avail. Then LLMs came in after the CTO became obsessed with it and it got hit with that hammer. This is four iterations of the same job: manual, Excel, Excel+VBA, Excel+CoPilot. 15 years this went on.
No one actually bothered to understand the reason the work was being done and the LLM did not have any context. This was being emailed weekly to a distribution list with no subscribers as the last one had left the company 14 years ago. No one knew, cared or even though about it.
And I see the same in all areas LLMs are used. They are merely pasting over incompetence, bad engineering designs, poor abstractions and low knowledge situations. Literally no one cares about this as long as the work gets done and the world keeps spinning. No one really wants to make anything better, just do the bad stuff faster. If that's where something is useful, then we have fucked up.
Another one. I need to make a form to store some stuff in a database so I can do some analytics on it later. The discussion starts with how we can approach it with ReactJS+microservices+kubernetes. That isn't the problem I need solving. People have been completely blinded on what a problem is and how to get rid of it efficiently.
There is a finite amount of incremental improvements left between the performance of today's LLMs and the limits of human performance.
This alone should give you second thoughts on "AI doomerism".
That could also apply to LLMs, that there would be a hard wall that the current approach can’t breach.
The "walls" that stopped AI decades ago stand no more. NLP and CSR were thought to be the "final bosses" of AI by many - until they fell to LLMs. There's no replacement.
The closest thing to a "hard wall" LLMs have is probably online learning? And even that isn't really a hard wall. Because LLMs are good at in-context learning, which does many of the same things, and can do things like set up fine-tuning runs on themselves using CLI.
OpenAI's o3 was SOTA, and valued by its users for its high performance on hard tasks - while also being an absolute hallucination monster due to one of OpenAI's RLVR oopsies. You'd never know whether it's brilliant or completely full of shit at any given moment in time. People still used o3 because it was well worth it.
So clearly, hallucinations do not stop AI usage - or even necessarily undermine AI performance.
And if the bar you have to clear is "human performance", rather than something like "SQL database", then the bar isn't that high. See: the notorious unreliability of eyewitness testimonies.
Humans avoid hallucinations better than LLMs do - not because they're fundamentally superior, but because they get a lot of meta-knowledge "for free" as a part of their training process.
LLMs get very little meta-knowledge in pre-training, and little skill in using what they have. Doesn't mean you can't train them to be more reliable - there are pipelines for that already. It just makes it hard.
I do think though that lack of online learning is a bigger drawback than a lot of people believe, because it can often be hidden/obfuscated by training for the benchmarks, basically.
This becomes very visible when you compare performance on more specialized tasks that LLMs were not trained for specifically, e.g. playing games like Pokemon or Factorio: General purpose LLMs are lagging behind a lot in those compared to humans.
But it's only a matter of time until we solve this IMO.
Pre-training a base model on text datasets teaches that model a lot, but it doesn't teach it to be good at agentic tasks and long horizon tasks.
Which is why there's a capability gap there - the gap companies have to overcome "in post" with things like RLVR.
I didn’t say that is the case, I said it could be. Do you understand the difference?
And if it is the case, it doesn’t immediately follow that we would know right now what exactly the wall would be. Often you have to hit it first. There are quite a few possible candidates.
So far, there's a distinct lack of "wall" to be seen - and a lot of the proposed "fundamental" limitations of LLMs were discovered to be bogus with interpretability techniques, or surpassed with better scaffolding and better training.
I think you’re confused. You are the one making the extraordinary claim, the burden of proof is on you.
You asserted LLMs have a finite number of steps to go to reach (overcome?) human limits. You don’t know that. It hasn’t happened. You can’t prove it.
I, on the other hand, merely pointed out that is not a certainty.
Your teapot argument works against you.
The case for "hard wall": wishful thinking.
In the case of contractors, the contractors buy the subscription but they need authorization to give access to the code. That's obvious if the property of the code is of the customer but there might be NDAs even if the contractor owns the code.
I want to see some numbers before I believe this. So far my feelings is that the best case scenario is that it reduces the time it needs to do bureaucratic tasks, tasks that were not needed anyway and could have just been removed for an even grater boost in productivity. Maybe, it seems to be automating tasks from junior engineer, tasks which they need to perform in order to gain experience and develop their expertise. Although I need to see the numbers before I believe even that.
I have a suspicion that AI is not increasing productivity by any meaningful metric which couldn’t be increased by much much much cheaper and easier means.
I don't think that's of any doubt. Even beyond programming, imo especially beyond programming, there are a great many things they're useful for. The question is; is that worth the enormous cost of running them?
NFT's were cheap enough to produce and that didn't really scale depending on the "quality" of the NFT. With an LLM, if you want to produce something at the same scale as OpenAI or Anthropic the amount of money you need just to run it is staggering.
This has always been the problem, LLMs (as we currently know them) they being a "pretty useful tool" is frankly not good enough for the investment put into them
At this point the "trick" is to scare white collar knowledge workers into submission with low pay and high workload with the assumption that AI can do some of the work.
And do you know a better way to increase your output without giving OpenAI/Claude thousands of dollars? Its morale, improving morale would increase the output in a much more holistic way. Scare the workers and you end up with spaghetti of everyone merging their crappy LLM enhanced code.
The main reason being: even SOTA AIs of today are subhuman at highly agentic tasks and long-horizon tasks - which are exactly the kind of tasks the management has to handle. See: "AI plays Pokemon", AccountingBench, Vending-Bench and its "real life" test runs, etc.
The performance at long-horizon tasks keeps going up, mind - "you're just training them wrong" is in full force. But that doesn't change that the systems available today aren't there yet. They don't have the executive function to be execs.
This sounds like a lot of the work engineers do as well, we're not perfect at it (though execs aren't either), but the work you produce is expected to survive long term, thats why we spend time accounting for edge cases and so on.
Case in point; the popularity of docker/containerization. "It works on my machine" is generally fine in the short term, you can replicate the conditions of the local machine relatively easily, but doing that again and again becomes a problem, so we prepare for that (a long-horizon task) by using containers.
The problem is: we don't have an AI exec that would outperform a meatbag exec on average, let alone reliably. Yet.
Opus 4.5 saved me about 10 hours of debugging stupid issues in an old build system recently - by slicing through the files like a grep ninja and eventually narrowing down onto a thing I surely would have missed myself.
If I were to pay for the tokens I used at API pricing, I'd pay about $3 for that feat. Now, come up with your best estimate: what's the hourly wage of a developer capable of debugging an old build system?
For the reference: by now, the lifetime compute use of frontier models is inference-dominated, at a rate of 1:10 or more. And API costs at all major providers represent selling the model with a good profit margin.
> And API costs at all major providers represent selling the model with a good profit margin.
Though we don't know for certain, this is likely false. At best, it's looking like break even, but if you look at Anthropic, they cap their API spend at just $5,000 a month, which sounds like a stop loss. If it were making a good profit, they'd have no reason to have a stop loss (and certainly not that low).
> Yeah. Obviously. Duh. That's why we keep doing it.
I don't think so. I think what is promised is what keeps spend on it so high. I'd imagine if all the major AI companies were to come out and say "this is it, we've gone as far as we can", investment would likely dry up
It's not going to mean they can employ 0 engineers, but maybe they can employ 4 instead of 5 - and a 20% reduction in workforce across the industry is still a massive change.
Assuming all the stars align though and all these things come true, a 20% reduction in workforce costs is significant, but again, you have to compare that to the cost of investment, which is reported to be close to a trillion. They'll want to see returns on that investment, and I'm not sure a 20% cut (which, as above, is looking like a best case scenario) in workforce lives up to that.
Any standard of intelligence devised before LLMs is passed by LLMs relatively easily. They do things that 10 years ago people would have said are impossible for a computer to do.
I can run claude code on my laptop with an instruction like "fix the sound card on this laptop" and it will analyze what my current settings are, determine what might be wrong, devise tests to have me gather information it can't gather itself, run commands to probe hardware for it's capabilities, and finally offer a menu of solutions, give the commands to implement the solution, and finally test that the solution works perfectly. Can you do that?
1. create a skeleton clone of frontend A, named frontend B, which is meant to be the frontend for backend project B, including the oAuth configuration
2. create the kubernetes yaml and deployment.sh, it should be available under b.mydomain.com for frontend B and run it, make sure the deployment worked by checking the page on b.mydomain.com
3. in frontend B, implement the UI for controller B1 from backend B, create the necessary routing to this component and add a link to it to the main menu, there should be a page /b1 that lists the entries, /b1/xxx to display details, /b1/xxx/edit to edit an entry and /b1/new to create one
4. in frontend B, implement the UI for controller B2 from backend B, create the necessary routing to this component and add a link to it to the main menu, etc.
etc.
All of this is done in 10 minutes. Yeah I could do all of this myself, but it would take longer.
With adults usually having no more than 2-3 hours of free time per work day, this allows you to productively program in your free time without fully burning out.
Also, my company pays for claude and does not give a shit what I do with it.
Talk to any model about deep subjects. You ll understand what I am saying. After a while it will start going around in circles.
FFS ask it to make an original joke, and be amused..
so like your average human
> FFS ask it to make an original joke, and be amused..
let's try this one on you - say an original joke
oh, right, you dont respond to strangers prompts, thus you have agency, unlike an LLM
If an average human has seen and read all that is written till now, I bet that they can hold the conversation going for quite a long time...
>say an original joke
I asked an LLM if it had a good night with sweet dreams, It said, "I don't sleep and I only dream when I work!"
But by some subset of definitions my calculator is intelligent. By some subset of definitions a mouse is intelligent. And, more interestingly, by some subset of definitions a mouse is far more intelligent than an LLM.
It works because people have answered similar questions a million times on the internet and the LLMs are trained on it.
So it will work for a while. When the human generated stuff stops appearing online, then LLMs ll quickly fall in usefulness.
But that is enough time for the people who might think that it going to last for ever to make huge investments into it, and the AI companies to get away with the loot.
Actually it is the best kind of scam...
EDIT: Another thought. Thus it seems that AI companies actually have an incentive to hinder developements, because new things mean that their model is less useful. With the widespread dependence on AI, they might even get away with manipulating the population to stagnate.
My pocket calculator is not intelligent. Nor are LLMs.
I don't think conflating intelligence with "what a computer can do" makes much sense though. I can't calculate the X digit of PI in less than Z, I'm still intelligent (or I pretend to be).
But the question is not about intelligence, it's a red herring, it's just about utility and they (LLM's) are useful.
Yes, I have worked in small enough companies in which the developers just end up becoming the default IT help desk. I never had any formal training in IT, but most of that kind of IT work can be accomplished with decent enough Google skills. In a way, it worked the same as you and the LLM. I would go poking through settings, run tests to gather info, run commands, and overall just keep trying different solutions until either one worked or it became reasonable to give up. I'm sure many people here have had similar experiences doing the same thing in their own families. I'm not too impressed with an LLM doing that. In this example, it's functionally just improving people's Googling skills.
you're lumping together two very different groups of people and pointing out that their beliefs are incompatible. of course they are! the people who think there is a real threat are generally different people from the ones who want to push AI progress as fast as possible! the people who say both do so generally out of a need to compromise rather than there existing many people who simultaneously hold both views.
I feel this framing in general says more about our attitudes to nuclear weapons than it does about chatbots. The 'Peace Dividend' era which is rapidly drawing to a close has made people careless when they talk about the magnitude of effects a nuclear war would have.
AI can be misused, but it can't be misused to the point an enormously depopulated humanity is forced back into subsistence agriculture to survive, spending centuries if not millennia to get back to where we are now.
I think that's good, but the whole "AI is literally not doing anything", that it's just some mass hallucination has to die. Gamers argue it takes jobs from artists away, programmers seem to have to argue it doesn't actually do anything for some reason. Isn't that telling?
It's not really hard to see... spend your whole life defining yourself around what you do that others can't or won't, then an algorithm comes along which can do a lot of the same. Directly threatens the ego, understandings around self-image and self-worth, as well as future financial prospects (perceived). Along with a heavy dose of change scary, change bad.
Personally, I think the solution is to avoid building your self-image around material things, and to welcome and embrace new tools which always bring new opportunities, but I can see why the polar opposite is a natural reaction for many.
And if AI assisted products are cheaper, and are actually good, then people will have to vote with their wallets. I think we’ve learned that people aren’t very good at doing that with causes they claim to care about once they have to actually part with their money.
Or would you prefer these things be outlawed to increase employment?
There should have been one good game by now. But there isn’t.
By an extremely loud group of activists, as always. I'd wager most gamers don't care one way or the other.
Unless AI is used for code (which it is, surely, almost everywhere), then Gamers don't give a damn. Also, Larian didn't use it for concept art, they used it to generate the first mood board to give to the concept artist as a guideline. And then there is Ark Raiders, who uses AI for all their VO, and that game is a massive hit.
This is just a breathless bubble, the wider gaming audience couldn't give two shits if studios use AI or not.
Btw there was an update on this, larian did an AMA on reddit because of the backlash because the "just moodboard" excuse didn't satisfy many, they are now saying they won't use ai for anything at all:
> We already said this doesn’t mean the actual concept art is generated by AI but we understand it created confusion. So, to ensure there is no room for doubt, we’ve decided to refrain from using genAI tools during concept art development. That way there can be no discussion about the origin of the art.
https://www.ign.com/articles/everything-we-learned-about-div...
https://www.reddit.com/r/Games/comments/1q870w5/larian_studi...
I know LLMs won't vanish again magically, but I wish they would every time I have to deal with their output.
But they don't. Instead, "AI safety" organizations all appear to exclusively warn of unstoppable, apocalyptic, and unprovable harms that seem tuned exclusively to instill fear.
The problem is not that no one is trying to solve the issues that you mentioned, but that it is really hard to solve them. You will probably have to bring large class action law suits, which is expensive and risky (if it fails it will be harder to sue again). Anthropic can make their own models safe, and PauseAI can organize some protests, but neither can easily stop grok from producing endless CSAM.
[1] https://www.anthropic.com/news/protecting-well-being-of-user...
[2] https://www.anthropic.com/research/team/societal-impacts
I appreciate you pointing out the Risks page though, as it does disprove my hyperbole about ignoring present-day harms completely, although I was disheartened that the page just appears to list things that they believe actions "could be mitigated by a Pause" (emphasis mine).
The catastrophic AI risk isn't "oh no, people can now generate pictures of women naked".
In a vacuum, I agree with you that there's probably no harm in AI-generated nudes of fictional women per se; it's the rampant use to sexually harass real women and children[0], while "causing poor air quality and decreasing life expectancy" in Tennessee[1], that bothers me.
[0]: https://arstechnica.com/tech-policy/2026/01/x-blames-users-f...
[1]: https://arstechnica.com/tech-policy/2025/04/elon-musks-xai-a...
The whole thing with "AI polluting the neighborhoods" falls apart on a closer examination. Because, as it turns out, xAI put its cluster in an industrial area that already has: a defunct coal power plant, an operational steel plant, and an operational 1 GW grid-scale natural gas power plant that powers the steel plant - that one being across the road from xAI's cluster.
It's quite hard for me to imagine a world where it's the AI cluster that moves the needle on local pollution.
People are (or were, as of a couple of weeks ago) having nude images generated from pictures of them, and were posted publicly, by a company worth billions. You think the outage around that is just puritan hand-wringing? Not that real people are actively being harassed?
So there will be laws because not everyone can be trusted to host and use this "dangerous", new tech.
And then you have a few "trusted" big tech firms forming an oligopoly of ai, with all of the drawbacks.
That’s exactly what a con is: selling you something as being more than what it actually is. If you agree it’s overhyped by its sellers, you agree it’s a con.
> Current agents can do around 70% of coding stuff I do
LLMs are being sold as capable of significantly more than coding. Focusing on that singular aspect misses the point of the article.
That’s quite a funny take, because I bet you someone will have made that same argument to criticise “crypto-deniers”.
> pretending they're not doing anything clever or useful
That isn’t at all the argument of the article. No one is claiming LLMs are completely useless or that they aren’t interesting technology. The critique is they’re being sold as way more than what they are or could be, that that has tangible negative consequences we can already feel, and the benefits don’t offset it.
> This has, of course, not happened.
This is so incredibly shallow. I can't think of even a single doomer, who ever claimed that AI will destroy us by now. P(doom) is about the likelihood of it destroying us "eventually". And I haven't seen anything in this post or in any recent developments to make my reduce my own p(doom), which is not close to zero.
Here are some representative values: https://pauseai.info/pdoom
And that's the anthropic fallacy. In the worlds where it has happened, the author is dead.
Though I personally hope that we'll have enough of a warning to convince people that there is a problem and give us a fighting chance. I grew up on Terminator and would be really disappointed if the AI kills me in an impersonal way.
While there might be open issues with AI, those AI companies are providing *far* more value than null.
What parallel world are they living in? Every single online platform has been flooded with AI generated content and had to enact counter measures, or went the other way, embraced it and replaced humans with AI. AI use in scams has also become common place.
Everything they warned about with the release of GPT‑2 did in fact happen.
Hm... is it wrong to think like this?
Let’s not forget these innovations are on the heels of COVID. Strong, swift action by government, industry, and individuals against a deadly pathogen is “controversial”. Even if killer AI was here, twice shy…
I’m angry about a lot of things right now, but LLM “marketing” (and inadequate reporting which turns to science fiction instead of science) is not one of them. The LLM revolution is getting shoehorned into this Three Card Monte narrative, and I don’t see the utility.
The criticisms of LLM promise and danger is part of the zeitgeist. If firms are playing off of anything I bet it’s that, and not an industry wide conspiracy to trick the public and customers. Advertising and marketing meets people where they’re at, and “imagines” where they want to go, all wrapped up with the product. It doesn’t make the product frightening. It’s the same for all manner of dangerous technologies—guns, nuclear energy, whatever. The product is the solution to the fear.
> “The LLMs we have today are famously obsequious. The phrase “you’re absolutely right!” may never again be used in earnest.”
Hard NO. I get it, the language patterns of LLMs are creepy, but it’s not bad usage. So, no.
I can handle the cognitive dissonance of computer algorithms spewing out anthropomorphic phrasing and not decide that I, as a human being, can no longer in humility and honesty tell someone else they’re right, and i was wrong.
You have not actually made clear how mechanical calculators were a scam.
Ironically, this article feels like it was written by an LLM. Just a baseless opinion.
> Simply put, these companies have fallen for a confidence trick. They have built on centuries of received wisdom about the efficacy and reliability of computers, and have been drawn in by highly effective salespeople selling scarcely-believable technological wonders.
Calculators are ok, but LLMs are not calculators.
However, the title implies that they were a trick - otherwise why is the "confidence trick" 400 years old?
I feel like this kind of imprecise use of language is what makes it difficult to interact with LLMs in a meaningful way - perhaps that is the reason the author seems to dismiss the value of them.
The 'are LLMs intelligent?' discussion should be retired at this point, too. It's academic, the answer doesn't matter for businesses and consumers; it matters for philosophers (which everyone is even a little bit). 'Are LLMs useful for a great variety of tasks?' is a resounding 'yes'.