Salta al contenuto principale


bad news "AI bubble doomers". I've found the LLMs to be incredibly useful and reduce the workload (and/or make people much, MUCH more effective at their jobs with the "centaur" model).

Is it overhyped? FUCK Yes. Salespeople Gotta Always Be Closing. But this is NOTHING like the moronic Segway (I am still bitter about that crap), Cryptocurrency, which is all grifters and gamblers and criminals end-to-end, and the first dot-com bubble where not NEARLY enough people had broadband or even internet access, plus the logistics systems to support shipping products was nowhere REMOTELY where it is today.

If you are expecting this "AI bubble" to pop anytime soon, uh.. you might be waiting a bit longer than you think? Overhyped, yes, overbuilding, sure, but not remotely a true bubble any any of the same senses of the three examples I listed above 👆. There's something very real, very practical, very useful here, and it is getting better every day.

If you find this uncomfortable, I'm sorry, but I know what I know, and I can cite several dozen very specific examples in the last 2-3 weeks where it saved me, or my team, quite a bit of time.

Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

ok, you caught me… Centaur model?? That's a new one for me. I'm not at the vanguard of LLM stuff, but I'm also not totally out of the loop… or at least I thought I wasn't.
in reply to CM Harrington

@octothorpe nature.com/articles/s41586-025…
@codinghorror
in reply to Joachim Wiberg

@troglobit @octothorpe "centaur" is also used as a term to describe human in control, LLM doing the leg work. Opposed to vibe coding, where LLM is driving. From context it sounds like Jeff could be talking about this -- augmented human with LLM help, rather than autonomous LLM.
in reply to Jeff Atwood

@ctietze @troglobit @octothorpe Can someone give a concrete example of using LLMs in centaur mode? The theory makes total sense, but I can’t really picture what it looks like in practice.
in reply to balu

@Balu @ctietze @troglobit @octothorpe basically, double check its work, and build upon it. Never present a LLM work as "final", but treat it as a rough first draft which you manually improve and build upon.
in reply to Jeff Atwood

value? Sure. No doubt. Can that value make enough to cover the ridiculous spend?
in reply to Patrick Berry

@pberry in my experience, yes. Stop generating synethic porn videos and furry porn images, first of all, if you want to talk about where the REAL energy use is going. Let's drop the artifice and get real here.
in reply to Jeff Atwood

the structural deals between OpenAI, Oracle, NVIDIA, and all the companies they rely on for data centers just feels…frothy. Yes, I read a lot of Ed Zitron.
in reply to Patrick Berry

@pberry sure, froth, but there is substantial basis, unlike a LOT of other historical fads/trends in tech I can point to.
in reply to Jeff Atwood

it's real in much the same way the railroad boom was real (rather than tulip mania, say).

But LLMs are also not remotely worth the level of valuations and investment we're seeing, and that is a bubble that will pop. "Useful" and "a bubble" can both be true.

Many other types of AI systems have been in production use for years but command nothing like this kind of manic investment.

in reply to Jeff Atwood

A lot of companies went bust and a lot of people lost their savings in the railroad boom. Worth remembering that.
in reply to Janne Moren

@jannem for sure. but railroads MATTERED and solved a lot of real problems. There were also, sadly, the railroad robber barons.. sooo...
in reply to Jeff Atwood

@jannem but the "bubble" warnings of financial experts from Deutsche bank aren't about usefulness. It's about assets, revenue streams and the fact this frantic building of generic data centers is hiding the recession in the US.

Housing is useful too. Didn't stop the 2008 crash...

in reply to Joris Meys

@JorisMeys @jannem housing is RADICALLY different and this is totally comparing apples to oranges.
in reply to Jeff Atwood

@jannem you missed the point. Usefullness is no predictor for whether a market is inflated to the point any deflation will ripple through to other markets. As the housing/banking crisis did. They're not saying AI is useless. They're saying too many investors are exposed to too high risks in a volatile market and it will hurt people.
in reply to Joris Meys

the comparison between housing and datacenters is fundamentally broken and not relevant. Literally apples and oranges. But if you want to carry on with this broken comparison, go ahead. Just leave me out of it. Or, don't, and you will regret it.
Questa voce è stata modificata (2 giorni fa)
in reply to Joris Meys

@JorisMeys @jannem Oh, I definitely will have a great day, because I'm putting $69m into action to help desperately poor people and orgs doing amazing work.

blog.codinghorror.com/stay-gol…

blog.codinghorror.com/the-road…

in reply to Joris Meys

and again, the data centers will be useful for a lot of other compute tasks, driving cost down for everyone on a lot of problems. Why people can't see this, I don't know, but at this point? Fuck 'em. I am tired of explaining what I know to be true, and my track record on predictions is pretty goddamn good. Believe me or don't. I don't care. I don't care about anything except doing my GMI studies, and I will absolutely guarantee not just a minimum income in these rural counties in poverty, but that I will never, ever stop doing science on this. You will have to kill me to stop me. And even then, it will go on without me, because the team we have built is extraordinary blog.codinghorror.com/the-road…
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

@jannem and again, nobody said they were useless. But somebody has to pay to keep the lights on. I hope you're right and it won't be tax payers picking up the bill again.
in reply to Jeff Atwood

“I can cite several dozen very specific examples in the last 2-3 weeks where it saved me, or my team, quite a bit of time.”

Please do, if you can. Because most time I’ve tried to use LLMs for work the error rate ends up costing me MORE time than I would have spent without, and most AI boosters are short on specifics. We just had a presentation at my job on how we all need to be using AI with no case studies of how it’s actually been useful so far.

in reply to Seth Richards

here's one: a friend confided he is unhoused, and it is difficult for him. I asked ChatGPT to summarize local resources to deal with this (how do you get ANY id without a valid address, etc, chicken/egg problem) and it did an outstanding, amazing job. I printed it out, marked it up, and gave it to him.

Here's two: GiveDirectly did two GMI studies, Chicago and Cook County and we were very unclear what the relationship was, or why they did it that way. ChatGPT also knocked this out park and saved Tia a lot of time finding that information out, so she was freed up to focus on other work.

I could go on and on and on. Email me if you want ~12 more specific examples. With citations.

But also realize this: I am elite at asking very good, well specified, very clear, well researched questions, because we built Stack Overflow.

You want to get good at LLMS? learn how to ask better questions of evil genies. I was raised on that. 🧞

Questa voce è stata modificata (2 giorni fa)

reshared this

in reply to Jeff Atwood

@sethrichards "learn how to ask better questions of evil genies." Journalism and interviewing gets you halfway there.
in reply to Phil Wolff

or community management, where the best reference books are hostage negotiation books. Also, that's not a joke. They are the best reference books for community management, hands down.
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

@evanwolf @sethrichards wait, what? I'm intrigued! Do you have any specific recommendations for hostage negotiation books that were useful in your community management experience?
in reply to Jeff Atwood

Does the benefit you found here go beyond e.g. Covey's "First try to Understand" principle, or reflective listening and the like? I'm intrigued but also very disappointed by most non-fiction books with that one weird trick you really got to know, spread across 200 pages.
in reply to Christian Tietze

hostage negotation is extremely hardcore and very effective. I'll drill in later in a blog post.
Questa voce è stata modificata (1 giorno fa)
in reply to Jeff Atwood

looking forward to that -- hope I'll catch it in my RSS reader
in reply to Jeff Atwood

@sethrichards good point, this week I learned about prompt engineering to optimize results.
in reply to Jeff Atwood

I see, everyone else is using it wrong... Tools this bad/difficult should not be being presented how thier are. People are not going to invest days in prompt engineering before asking a generic LLM for medical advice.
in reply to ⊥ᵒᵚ⁄Cᵸᵎᶺᵋᶫ∸ᵒᵘ ☑️

@falken @sethrichards the tool over-hype masks the huge underlying utility. If you think of it as a tool that you need to learn to use and not your “AI co-worker” or other hype-fueled nonsense, it is quite productive. Example: I am not an iOS engineer but I’m a principal engineer. With me directing Claude Code, reviewing its work, setting code standards, and asking for refactors, I produced thousands of lines of well-factored, tested code iOS code that passed review and shipped.
in reply to Jeff Atwood

@sethrichards Jeff - Can you package your knowledge on how to "ask better questions"?
in reply to J. Peterson

@isonno @sethrichards there are many pages on Stack about this, ref stackoverflow.com/help/how-to-… and stackoverflow.blog/2010/10/04/… and meta.stackoverflow.com/questio…
in reply to Jeff Atwood

@sethrichards "A friend is unhoused, and although I can afford to bet a million dollars, I didn't house him, I gave him some AI slop". What a saint you are!
in reply to Major Denis Bloodnok

well, we allocated $69m to this, and I'd have to get divorced to do more, because half is what we agreed on (though that is, in fact, more than half). If you'd like me to get divorced, let me know. Refer to: blog.codinghorror.com/stay-gol… and blog.codinghorror.com/the-road…
Questa voce è stata modificata (1 giorno fa)
in reply to Jeff Atwood

@sethrichards Except we know perfectly well you can afford to bet a million dollars. You've got the money, you'd just rather serve slop.
in reply to Jeff Atwood

@sethrichards

Evil genies with a severe form of ADD of some sort.

You hit it on the head - the prompt is the key.

With an experienced human - vagueness is often acceptable, and they will usually ask for clarification. The AI doesn't ask - it guesses, often incorrectly. So you need to over-specify in the prompt, including things it might be insulting to mention when talking to an experienced human. Then iterate, and aggressively steer that conversation.

This is why I don't see the AI as replacing a human except for trivial situations. It's a force multiplier, but not a replacement, and the skills necessary to use them effectively are non-obvious.

in reply to Jeff Atwood

@sethrichards

So your argument is simultaneously:

> LLMs are useful RIGHT FUCKING NOW for SO MANY scenarios

But also, they're only useful because:

> I am elite at asking very good, well specified, very clear, well researched questions, because we built Stack Overflow.

Is it then fair to say that LLMs are likely to be very misleading for people who do not have your "elite" experience?

If not, why not?

in reply to Nik

@nikclayton @sethrichards people also suck at writing titles to emails. So yeah, we need to teach people how to write.
in reply to Jeff Atwood

That's helpful, thanks. I guess the follow-up question is, how do you know you can trust the output? That's the real crux of the issue for me. My experience has been that these systems will confidently generate a mix of right and wrong info, while stripping all the context clues I've traditionally used to gauge validity. If I have to resort to traditional web searches etc. to check their work then I might as well skip the LLM step.
in reply to Jeff Atwood

@sethrichards@mas. Those examples do not make it clear to skeptical drive-by readers like me how you established the extent to which the output you received was actually correct

Is part of the magic value add to embrace the idea that for many activities, being "actually correct" isn't the most important criteria? Compared to, eg, just having a direction to get started in.

If someone could reference or breakdown examples that did unpack actual correctness, that would be persuasive.

Questa voce è stata modificata (1 giorno fa)
in reply to Jeff Atwood

The problem is not "LLMs are useless and when the bubble bursts they go away," they aren't going away any more than websites went away when the .com bubble burst.

The problems are
1. They are a 6/10 tool being advertised as an 11/10 tool with the folks selling this stuff consistently overstating what they're capable of doing.
2. The few hundred billion spent building them needs the 11/10 promises to come true in order to be justified.
3. They're really good at making up answers that appear *plausible* but are also completely wrong, and verifying the answers is becoming increasingly difficult as the top search results are increasingly flooded with output from the same LLMs.
4. 'AI' is being used to try to sell a bunch of completely unrelated stuff like 'copilot+ pcs' even though everything meaningful in the LLM space only runs in datacenters due to GPU memory limitations.

Maho Pacheco 🦝🍻 reshared this.

in reply to Billy O'Neal

@malwareminigun I would add one more, at least 50% of use cases are already solved in a more deterministic/efficient automation way.
in reply to Maho Pacheco 🦝🍻

@mapache Making a claim like that requires data I do not have, and that claim does not match my personal experience. In particular, because the first few things I tried to do with LLMs were tasks like that (e.g. "for every line that says =x64-windows in this file, duplicate that line as =x64-windows-static") and they were *awful* at it. But cases like "I am not a macOS sysadmin expert, help me figure out why our build machines have half their disks filled with garbage after a few weeks" leading to "the problem is Spotlight being awful" have worked out great. @codinghorror
in reply to Billy O'Neal

@malwareminigun
LLMs won't go away but a lot of the companies selling them will. This will be quite disruptive.

Prices have to rationalize, I am not quite sure how that will work if $200/mth is not profitable. I don't see any path to the models becoming 100 times cheaper to generate. This implies that some category of folks will have to accept much worse models for a reasonable price point.

Alternatively, all the start-ups disappear into existing big tech companies and they dramatically reduce spending and folks get model updates at a much much slower frequency.

Either way this will be a very different landscape then we have today.

in reply to Shafik Yaghmour

@shafik @malwareminigun "This implies that some category of folks will have to accept much worse models for a reasonable price point. " you are conveniently ignoring that fact that it IS possible to come up with far more efficient models in compute time, that are almost as good -- or perhaps even better!
in reply to Jeff Atwood

@sethrichards when I was in uni we learned about specifying pres and posts to out functions as contracts and a way to derivate the algorithm from the post to the pre. Not only the derivation was already hard, sometimes defining the pre and post was as hard to define as solving the problem in the first place.

1/

in reply to Marcos Dione

@sethrichards

And now you're saying that you not only have to double check every answer, you also have to be very good at asking the question to begin with? And these tools get the answers wrong if you don't? And they're released to the unsuspecting, untrained, I-have-no-free-spoons general public?

in reply to Marcos Dione

@mdione @sethrichards not exactly. They don't get it wrong, most people's questions are fairly rudimentary. Just ask anyone who works the help desk at {any company I can think of}
in reply to Jeff Atwood

@sethrichards and still, you don't know if your question is rudimentary or not, so you can't trust the answer until you double check.
in reply to Jeff Atwood

the internet is also incredibly useful and yet the dotcom bubble was a real bubble that destroyed hundreds of companies. The AI race is a circular economy that is 100% dependent on delivering “superintelegence”, not just “very very useful” tech. Will it? Maybe… but it’s not crazy to see risk there.
in reply to poswald

@poswald dotcom bubble was mostly the logistics issue and lack of broadband / internet access. Those problems are LONG gone.
in reply to Jeff Atwood

not sure I follow… The “internet is the future” people were not wrong… and yet they all got wiped out when the bubble popped. It took years to recover and many great companies formed in the aftermath (YouTube put the excess capacity to good use for example). My point is that AI can be very useful, and the future of tech, and also ready to pop and wipe out billions in corporate value at the same time.
in reply to poswald

again, I've explained this like 15 damn times now.. a) the first bubble was because ALMOST NOBODY was ON the internet, and those that were had crap connections. We needed broadband and much wider adoption of broadband everywhere. b) the systems of logistics and shipping were terrible back then, and evolved exponentially since.
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

yeah I was there working at a startup in NYC too… I get your point. You think any AI bubble will be mitigated because the tech can be delivered to the consumer easily this time.

I was making a different point that I think explains why you still hear AI doomers despite it being useful tech. It’s still a very dangerous bubble that will likely misallocate vast funds and careers IMO. Anyway it’s fine. sorry, I didn’t mean to frustrate you with my comment.

in reply to Jeff Atwood

I agree there are great uses for llms. But I'm not sure the dot com boom happened because of broadband and logistics. It happened because revenues and profits took more than half a decade longer to catch expectations. So there are certain similarities to today.
in reply to Von Javi

@javiervg yes, and revenues and profits were not there for the exact reasons I said. a) not enough users on the internet period, and far from enough broadband and b) wildly underdeveloped logistics / shipping systems.
in reply to Jeff Atwood

revenues were no there because just broadband. Broadband did make new applications possible , Netflix streaming, etc. All new technologies take time to adopt. ecommerce took almost a decade to 2.5% of total sales. It is still not even close to what people thought in 2000.

Profits were more because many companies focused on attracting customers, investing no matter the cost. Once money to support losses stopped flowing in. The dance stopped, market crashed, so yes similarities

in reply to Von Javi

there really are not that many similarities to today. the LLMs are useful RIGHT FUCKING NOW for SO MANY scenarios. And everyone is on the internet, because they have to be -- otherwise how do they ever get laid. Or apply for a job. Or look up information. Or.. really ANYTHING AT ALL. Good luck living off the 'net these days. You're gonna need it.
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

@poswald Pithy take: LLM bubble is mostly the lack of cheap GPU bandwidth and the energy/water logistics issue. In 20 years those problems will be LONG gone.
in reply to Eric Gerlach

@egerlach @poswald yeah but I worry about nvidia's extreme dominance here. We need a LOT more competition in GPUs. AMD is trying, but we need like 4-5 more companies building and innovating hardware. This is why I'm kinda sad Apple isn't willing, because their hardware is INCREDIBLE.
in reply to Jeff Atwood

@poswald In terms of monopolists, I'm slightly more concerned about TSMC and ASML. If Nvidia implodes, others will rise from the corpse. Too much talent there not to. But the fab design/implementation cycle is much longer than the chip design cycle. Would take much longer to recover. But TSMC is also more resilient than Nvidia.

I'm also concerned about the new Intel CEO. Doesn't seem to value long-term investment. I think Intel was 1-2 generations from being a competitive player, but I don't have faith they keep it up now.

in reply to Jeff Atwood

one can make a fairly successful argument that the “sons of segway” the lithium-ion powered cheap Chinese made scooters and e-bikes are reshaping cities and last mile mobility.
in reply to Jeff Atwood

I'll tag you in a few days with this project I'm working on. VERY much not a big deal. But way beyond my capabilities. I've been using Cgpt to help build my new portfolio site. During this, I have found it is grossly object blind to its own errors. First drafts are always cool, way beyond anything I could do or even afford to pay someone for. But I'll find a glitch and then spend 10hrs trying to get it to track it down. It just pushes the error further down the line, but still there. The only fix was to dump that chat window and start fresh, completely rephrasing the issue and the desired resolution.

Ironically, this is more like a human than anything else. Humans are invariably unable to see their inherent personality and thinking flaws. No matter how well pointed out, how hard they are worked on, invariably they spend more time pushing the problem around and not actually solving it. We have entire industries built on this very issue, therapists, pop culture self-improvement, religions... For the last 19 days I've run into this same issue with it every single day. And spent way more time not fixing minor issues it generated than actually moving forward.

5 times it even gave me code to drop in that had spelling errors. We track the bug down and it blamed me. I copied and pasted that very code and fed it back to it to find the issue and then deny it wrote that error. Talk about a freakishly human thing to do.

I've used it now for 2yrs to help with art projects. It's far better for that than almost every human I know. With the correct personality framework, it ends up being incredibly useful as a sort of partner in the project.

I do think there's a lot it cannot do, yet. For specific tasks it is better than many humans can be. And I think, given the resources being tossed at it, this is going to rework most all of human culture/industry/interaction. But if it already has human flaws built into it, I suspect that those will grow in a similar way.

in reply to retech

@retech I never, ever use LLMs for code and never will. I barely trust other HUMAN coders, much less these statistical models.
in reply to Jeff Atwood

I don't have a lot of options, given my budget constraints.
in reply to retech

@retech if you're writing code, you've already failed. The best code is no code at all. blog.codinghorror.com/the-best…
in reply to Jeff Atwood

@retech yes yes, and then we do write code—not that useful to me to think of it as failure even though not writing code is always desireable. I‘m not as dismissive towards LLM writing code, but it surely would be criminal not to understand the output.
in reply to Jeff Atwood

I'm right there with you. The increased productivity is staggering when you know how to write the prompts.
1. Pretend you're writing a legal document or contract - say the things that seem obvious and be painfully precise.
2. use the LLM to eliminate tedious tasks entirely.
3. treat it like a smart junior team member you're collaborating with - give it the shape of what you expect the result to be.

Using these rules, what used to take 3 days can be accomplished in 3 hours.

Jeff Atwood reshared this.

in reply to Elijah

@elijah studies indicate that we overestimate how much it actually speeds us up, but treating it as a junior dev or an intern is the way to work with it.

I complain constantly about the mistakes it makes, but I often use it to scaffold boilerplate and make quick small adjustments successfully. I just have to be super vigilant about what parts I commit. When possible providing an example of what you're trying to accomplish helps.

In practice it seems to only speed me up ~20%

in reply to Curtis Carter

@codingcoyote @elijah what a convenient number pivot-to-ai.com/2025/07/11/ai-…
in reply to noodlejetski

@noodlejetski @codingcoyote I'm well aware of that study. I only know what the actual impacts have been in my experience.

All I can share is the real-world experience I have.
ex: Creating production-quality services implementing business logic between two APIs used to take a few days with learning the APIs, writing the tests, setting up dev and e2e testing environments, integrating with CI, etc. Now, AI does 90% of the drudgery while we're doing meaningful things elsewhere.

in reply to Elijah

@elijah @noodlejetski @codingcoyote that seems very carefully phrased. You didn't say "now it takes less time". I don't mean to be antagonistic, so please forgive me for if I'm being silly. I'm just trying to understand, so being explicit is helpful.
in reply to Jonathan Hartley

@tartley @elijah @noodlejetski @codingcoyote again, I think programming and coding is one of the WORST uses cases for LLMs. It requires extreme precision and total accuracy and .. that's not really what they're good at.
in reply to Jeff Atwood

Nobody is questioning the practical utility, the problems are all fundamentally about economics. Unless someone makes breakthroughs that can at scale generate ROI, you're going to reach a threshold where there's not enough capital in the market to sustain the ongoing investment while also simultaneously starving investment in other industries.

Obviously the investors know what they're doing right, that's probably what everyone assumes at this stage 😀

in reply to deepfryed

@deepfryed having a lot of competition in compute time is a very good thing. Compute time is useful for a LOT of stuff, and competition pushes prices down for all.
in reply to Jeff Atwood

@deepfryed

You seem to be assuming compute time is fungible. The GPUs being built for ML are heavily optimised for multiplying sparse matrixes of very low-precision floating-point values. They are not even very good at graphics, let alone other workloads.

in reply to David Chisnall (*Now with 50% more sarcasm!*)

@david_chisnall @deepfryed I only recently found out that most of the GPUs that NVidia makes nowadays ... don't do graphics. At least not at a quality much better than a C64 would. Investments into electricity infrastructure are likely going to be useful past the bubble but most of these GPUs are kinda pointless.
in reply to Peter Bindels

@dascandy @david_chisnall they're good on numerical computational workloads, perhaps you can put them into use in weather modeling etc but it's not going to recoup initial investment that assumes returns in the AI / ML space.

It remains to be seen, but unless there is a good plan to use these massive data centers to somehow to monetise the numerical compute resources in a different / profitable problem space they'll go broke very soon.

Questa voce è stata modificata (1 giorno fa)
in reply to deepfryed

@deepfryed @dascandy
As @dascandy says, they're not generic floating point, they're very tailored to specific workloads. The last AI accelerator I worked on had a neat vector floating point representation with a single exponent and a vector of mantissas. This was great for AI workloads because it saved you a load of work in normalisation, but was useless for general matrix problems. Similarly, modern AI accelerators are tuned for a 4-bit or 8-bit floating-point operations.

For a lot of GenAI workloads, you really care about negative infinity, zero, and positive infinity. There are some interesting papers on using tri-state logic to represent exactly these, but you can also use 2-bit or 3-bit (and get 1 and -1 as well) representations to cover that range. Things that can very rapidly multiply matrixes of 2-bit floating-point values are great for LLM inference, they're completely useless for scientific computing.

Add to that, it turns out that AI training is actually better if your floating point units introduce random errors. You normally add some entropy intentionally, but if your hardware is slightly non-deterministic you need to do this less. So it's possible to push the compute units in an AI-tuned floating-point pipeline much closer to their physical limits because some errors are beneficial. This was also a problem when people started repurposing GPUs for GPGPU workloads: a bit of error in a single pixel or vertex in one frame will not be noticeable, whereas the same error in a finite element simulation may lead to the wrong result. It's much worse for accelerators aimed at AI workloads.

They're as much general-purpose compute resources as a size-6 Torx screwdriver is a general-purpose fixing tool.

in reply to Jeff Atwood

@deepfryed @dascandy

How? Hardware tends towards consolidation because designing a new chip is really expensive and having a larger market lets you amortise that. Manufacturing has already been consolidated because it costs around $100bn to build a fab for a new generation, but design is increasingly consolidated as well. We’re producing microcontrollers, some of the simplest chips you can make (they are tiny, a couple of square millimetres) and the cost of getting the first one to market, using 22fdx (which is several generations old technology at this point, so starting to be affordable for low-cost devices) is several million. Getting something competitive with Intel or NVIDIA to market is at least a hundred million, assuming you get it right first time. More practically, it’s a multiple of that.

A few years ago, there were a couple of hundred startups designing AI chips. Every single one of the, was targeting an exit where one of the big tech companies bought them. They all wanted to work with us at Microsoft because being bought and folded into Azure was one of their few paths to success.

in reply to Jeff Atwood

interesting anedoctal evidence!

Now, how about we get serious and publish/wait for some (at least potentially) unbiased study/research on that?

Because I haven't seen any. All I've seen are the likes of this one, negative about Centaur:

circumstances.run/@davidgerard…

in reply to tinsukE

@tinsuke here's the specific examples. Feel free to explain why I'm wrong. I'll be waiting. Good luck, pal. infosec.exchange/@codinghorror…


here's one: a friend confided he is unhoused, and it is difficult for him. I asked ChatGPT to summarize local resources to deal with this (how do you get ANY id without a valid address, etc, chicken/egg problem) and it did an outstanding, amazing job. I printed it out, marked it up, and gave it to him.

Here's two: GiveDirectly did two GMI studies, Chicago and Cook County and we were very unclear what the relationship was, or why they did it that way. ChatGPT also knocked this out park and saved Tia a lot of time finding that information out, so she was freed up to focus on other work.

I could go on and on and on. Email me if you want ~12 more specific examples. With citations.

But also realize this: I am elite at asking very good, well specified, very clear, well researched questions, because we built Stack Overflow.

You want to get good at LLMS? learn how to ask better questions of evil genies. I was raised on that. 🧞


in reply to Jeff Atwood

those examples sound OK, but I'm not particularly interested on their specifics or picking them apart.

I'd be interested on how representative of the overall experience they are. Because they're still anedoctal evidence and I don't think one could generalize LLMs' usefulness from them.

That's what I meant by expecting some unbiased research or study with thorough analysis, specially given how LLM users seem to be bad at estimating the benefits: metr.org/blog/2025-07-10-early…

in reply to tinsukE

@tinsuke "those examples sound OK, but I'm not particularly interested on their specifics or picking them apart."

Which means they are valid examples. I can provide a few dozen more via email if you want. Highly specific, too.

To your point, I think LLMs will continue suck a lot at code and coding, because it requires very precise, very accurate language. But a LOT of human tasks do not require this level of precision. Compare with answering the phone at the local public zoo. How many times do you think you say the same 20 answers over and over?

in reply to Jeff Atwood

I didn't bother asking about the examples (and still ain't!) because anedoctal evidence little matters in supporting statements like that LLMs would be useful tools for the general public (or a decent, monetizable, chunk of it), and not just a product of hype.

If LLMs are useful, why wouldn't trustworthy studies/research be able to show that? That would make convincing people about them so much easier.

As easy as addressing/waving off the ethical concerns around them, that is.

in reply to Jeff Atwood

largely agreed, but given the literal trillions we’re spending I feel the bar for this not being a financial bubble is much higher than mere existence of utility.

After the dust settles, will we have useful LLMs? Yes. Will most AI investors have lost their shirts? Also yes.

in reply to Marko Karppinen

@karppinen totally agree, once the bubble bursts and all the techbros move onto their next Ponzi scheme then maybe the real people can make something actually useful out of ai (and sensibly rename it)
in reply to Johnnyvibrant

@johnnyvibrant @karppinen look, people, if you can't see what garbage cryptocurrency is compared to this, then again -- you're not worth my time to even talk to.
in reply to Jeff Atwood

@johnnyvibrant sure, crypto is useless if you are not a criminal. I don’t follow how that means there’s no AI bubble
in reply to Marko Karppinen

@karppinen @johnnyvibrant because a bubble based on something that actually works, and has value, and isn't completely impractical is very different from a bubble based on bullshit.
in reply to Jeff Atwood

Bad news "subprime housing bubble doomers". I've found homes incredibly useful and reduce life on street (and/or make people much happier of their conditions).

This is NOTHING like previous overleveraged financing and not REMOTELY like a true bubble because people live in houses and banks won't yank them.

If you find this to be uncomfortable, sorry, but lessons have to be learned.
@codinghorror

in reply to Osma A 🇫🇮🇺🇦

how is housing even remotely like this? What are you even talking about?
Questa voce è stata modificata (1 giorno fa)
in reply to Jeff Atwood

it looks like the dotcom bubble : overhype, overinvestment, the bubble burst but the web is still here and thriving.
in reply to Jeff Atwood

How do you square the excessive environmental costs with your “stay gold” ethos?
in reply to John Parker

@Middaparka provide detailed specifics, please, with citations, or else walk away now
in reply to Jeff Atwood

Ahh… so the onus is on me to somehow uncover the true costs of the model training, etc., despite the fact that all of the players in the industry go to great lengths to obfuscate them?

Guess I’ll be walking away then. 🫡

in reply to John Parker

no, you can do nothing and provide no evidence for your arguments. It doesn't matter to me; nothing matters to me. I am doing the county level GMI studies no matter what. No one can stop me from doing science and putting my money in the hands of poor people in rural counties in the USA. You do whatever you need to do, John, and I will do what I need to do to stay alive. How about that. blog.codinghorror.com/the-road…
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

To be clear, Jeff, I firmly believe what you’re doing in terms of wealth distribution, both in terms of your personal wealth, and the “stay gold” initiative is incredibly admirable. Whilst it’s an option only available to a few, taking a top-down approach such as the one you’re taking is one of the few ways meaningful change can be enacted.

It’s a pity there’s not more people out there with the same attitude, and the courage to put their money where their mouth is.

in reply to John Parker

Concepts like Universal Basic Income/Guaranteed Minimum Income and an acceptance of the environmental trade-offs of AI feel like uncomfortable bedfellows to me though, although perhaps I’m just ill-informed. Either way, I’ve the luxury of not being in a position where my opinion or influence matter in the slightest.

Irrespective, I very much wish you well with “stay gold”. If there were more people with your convictions, the world would be a measurably better place.

in reply to John Parker

@Middaparka I'm not super worried since all the solar graphs are far, far exceeding the most positive predictions and estimates. And we can use compute time for a lot of very useful things. A decline in cost of compute time benefits the whole world. But yes, they should be built sustainably and in harmony with the earth, JUST LIKE EVERYTHING ELSE WE HAVE EVER BUILT.
in reply to Jeff Atwood

@Middaparka

Here's a couple to start with:

1. mit-genai.pubpub.org/pub/8ulgr…

> The unfettered growth in Gen-AI has notably outpaced global regulatory efforts, leading to varied and insufficient oversight of its socioeconomic and environmental impact [...]

2. Google's 2025 Environment Impact report, sustainability.google/reports/…

> Compared to 2023, our total [CO2] emissions increased by 22%, primarily due to increases in data centre capacity [...] for AI.

in reply to Nik

@Middaparka

Google used to say "Net zero by 2030". Now that report I linked to has changed that to "50% of 2019 emissions by 2030".

in reply to Nik

@nikclayton @Middaparka I'll have to research this because Google alone seems unlikely to move the needle on emissions based on how awful internal combustion engines are, airflights, etc. I'd shut down all the airlines first if I was really serious about emissions. Let me do some deeper research and get back to you in 15 minutes.
in reply to Jeff Atwood

I heard a podcast recently (ProfG I think) predicting the $$$ bubble will pop, but the utility will remain.

A good analogy is PCs. Originally the 1980s PC built many fortunes (remember Gateway 2000?) but it eventually became a low-margin commodity.

in reply to J. Peterson

@isonno ProfG sucks, and doesn't know what he's talking about, because he's all about his own brand.
in reply to Jeff Atwood

Whatever the source, I think commoditization of AI tools is inevitable.

AI being a useful tool and AI being an investment bubble can both be true.

See also railroads and PCs.

in reply to Jeff Atwood

I'll stop calling it a bubble when core functionality stops being neglected for shoehorning into everything no matter how hard one tries to actively avoid it as justification for circular investments.

As much as I'm sceptical of anyone saying it actually improves their ability to do X task with Y amount of people, websites didn't die when the dotcom bubble burst and neither did cryptocurrencies. They just got relegated to the tasks they were actually useful for after enough blood was spilled to write the regulations with it. All of my complaints with using LLMs for things can be resolved without killing it off entirely.

Lately my issue is more with the zero sum game nature of it. it'd be difficult now, but I could've easily got along without internet when that bubble burst. I got along just fine each time cryotocurrencies went bust. With what people are reducing a the two-letter marketing phrase of half a century ago is something I'm constantly having to actively avoid at basically every step, and even then there's very likely personal data of mine that I cannot prevent from being fed into training data no mater how loudly and how explicitly I state that I DO NOT consent.

If it's so useful, I don't need to be cautious about my operating system, the tools I use, where I host my projects, what configuration I have set for everything down that pipeline, and risk remaining in a perpetual state of unemployment if I don't change the workflow I've had for over a decade so that at every one of those steps my hand is forced further and further away from the vision I as the creator of said projects had in mind and more towards tweaking a system I never asked to become the entirety of my work. If it's so useful, its own merits will have me curious and I'll actually take the plunge of my own volition.

But how it's done now, how pervasive and inescapable it's becoming, how stigmatised wanting to perfect a craft with your own two hands at every step of the way is becoming, it's less reminiscent of a revolutionary paradigm shift and more reminiscent of the cult I left when I entered adulthood.

in reply to Jeff Atwood

there is a bubble because there is no way these AI companies will be profitable. The dotcom bubble burst bot because the internet wasn't useful (it was) but because all the dotcom companies were unprofitable.

Investors expect exponential growth, but there is no way for openAI to grow any further, and it's difficult for them to charge any more money from customers. AI models are too easy to replicate by competitors, so there is no lockin, costomers can go to competitors any time.

in reply to Marius Gundersen - mdg 🌻

then there are the investments being made now, which will be impossible to downpay ($100 billion???). Nvidia is singlehandedly holding up the US economy, but that share price is based on expected future earnings. But how can Nvidia expect even bigger investments in the future? What happens to all the funds invested in Nvidia when Nvidia suddenly won't expect AI companies to make huge investments the next quarter?
in reply to Marius Gundersen - mdg 🌻

and we've seen the diminishing returns on new LLM models. There is exponential growth in costs to develop a new marginally better model. There just isn't demand or willingness to pay for that model.

Once a technology becomes good enough then it's more about convenience than quality. MP3 isn't the best audio, but it's dominant, same as streaming movies, even though Blu-ray is technical much better.

Even small LLMs have been shown to be good enough for most use...

in reply to Marius Gundersen - mdg 🌻

@gundersen yep. Already said that here. Feel free to read it. Or don't. I really don't care. You do you. infosec.exchange/@codinghorror…


The LLM / GAI people are hitting exponential difficulty walls with massively diminishing returns. I don’t care how many GPUs and “training data” you throw at the problem, you can’t brute force your way out of this… but you can certainly waste billions trying. My self driving car bet with Carmack is the canary in the coal mine. When, and only when, we have fully autonomous SAE level 5 self driving cars, we thus by definition have true, full general purpose artificial Intelligence: blog.codinghorror.com/the-2030…

in reply to Jeff Atwood

my experiences differ from yours and it has actually been more of a problem than a help in my team. Even if this wasn't the case, the fact many LLMs are based on dubious content sources and consume massive amounts of energy make them deeply unappealing to me from an ethical standpoint
in reply to Sephster

@sephster depends on the task. For code and coding I expect it to be horrible, and wouldn't even try.
in reply to Jeff Atwood

deep AI/ML bubble or GenAI bubble? I think there is a difference and unless deep AI/ML can take up the momentum, I think GenAI AI will pop. There was a huge web bubble and yet here I am 25 years later replying directly to a legend via the web.

I hold with those who feel we're overestimating in the short term and underestimating in the long term.

Don't have a million though.

in reply to Sasha

@sashabilton waymo execs straight up say they think we'll NEVER get to SAE level 5, so if "deep AI" means real artificial intelligence, certainly not in my lifetime, and probably not in my children's lifetime, would be my best estimate. Beyond that, who knows. We need deep breakthroughs in Quantum computing to make any progress IMO. Computers are currently too shackled in their binary world of 1 and 0, all or nothing.
in reply to Jeff Atwood

not that deep 😀 more multi-agentic, successful processes that can do more than automate straightforward tasks. I think physical world AI is really tough, but the digital equivalent of drive to the store, collect the shopping, check the barcodes for accuracy, ask to have failures fix or offer alternatives, drive home and suggest recipes from the order, knowing who is home and what they like. Tell you how much exercise you need to burn off the calories.
in reply to Jeff Atwood

Considering how these models are trained and how the fair use principle is abused, praising current crop of big AI models is a bit in contradiction with your values, no?

Or do I have a wrong impression of you about ethics, privacy, and doing the right thing?

in reply to Hakan Bayındır

@bayindirh as I said here infosec.exchange/@codinghorror…


if only everything ever published on the web was Creative Commons!

Atwood's Third Law: Content licensing is now the hardest problem in computer science.


in reply to Jeff Atwood

Creative commons license also has a non-commercial, no-derivs, attribution, share-alike license (CC-NC-BY-SA), which I license my blog with. This normally blocks AI training (no transformation, no-sell, must-cite), so, CC doesn't allow free-reign over training, and I don't want to feed models with my output.

So your stance is, "tech is more important, we can figure ethics later" AFAICS.

Thanks.

in reply to Hakan Bayındır

not at all what I said. I'm saying licensing is a FUCKING NIGHTMARE problem. Have you ever even once looked at how difficult music licensing is, alone? Protip: read this: infosec.exchange/@codinghorror… and then this infosec.exchange/@codinghorror…

Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

but feel free to imagine things I didn't actually say and then complain about them, I guess? Is that the kind of person you want to be?
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

I might have misunderstood you, sorry if I did.

My rule-0 is don't do anything that you don't want to experience. So, again, if I misunderstood you, sorry about that (English is not my native language to begin with). It's not my intention to stuff words into anyone's mouth.

Yes, I know how music licensing is a hell of an onion. I played in an orchestra and have enough musician friends to experience in close proximity.

in reply to Hakan Bayındır

yeah and music licensing -- which is a living nightmare -- is trivial compared to how complex "INDIVIDUALLY NEGOTATE AND LICENSE EVERY SINGLE BIT OF CONTENT EVER CREATED FOREVER" is gonna be. I'm not saying steal everything. I'm saying this is an INSANELY difficult problem to take on and solve.
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

I understand what you are saying, but actually it's not "that" hard if you want to solve it, but the AI industry doesn't want to solve it, because current form of "steal first, apologize later, but never remove that data" is much more profitable for them, for now.
in reply to Hakan Bayındır

@bayindirh I don't condone stealing at all but "let's create another fifty thousand different nightmare mode bureaucracy licensing systems like music" is really not appealing to me either. Creators should get paid for their work, for sure.

You, of all people, know that musicians get screwed more than anyone else with that "perfectly legal and OK" licensing system. So is it a good system, then?

Hard things are hard.

in reply to Jeff Atwood

No, music licensing and academic publishing is two mediums that rip the creators the most. I support neither model in their current forms.

My proposition is a narrower interpretation of fair use and license detection on the page.

If you gonna sell this model or to access to it, assume everything is "All rights reserved". Any license preventing transformation stops scraping. Viral licenses assumed affect all output, exclude cite requiring content if your model can't cite. Simple.

in reply to Hakan Bayındır

@bayindirh the disconnect is that web crawling was always allowed, so this effectively destroys all web crawling. It's not nearly as simple as you're painting here. This is "shit's easy syndrome"
in reply to Hakan Bayındır

Guys building "The Stack" use a license filtering system to select what to include. LLMs are "smart" enough to understand licensing lines on the things they ingest.

If industry wants, we can add relevant HTTP headers to our pages to signal our stance.

They are simple, open ways to communicate what creators want. The only obstacle is the AI companies. Will they cooperate?

in reply to Hakan Bayındır

@bayindirh I mean, everything's simple when you don't have to actually do it or implement it yourself and are imagineering stuff. All I'm saying is, you are RADICALLY underestimating how hard this problem is. I agree with you that it matters and we need to figure it out.
in reply to Jeff Atwood

@bayindirh (this reminds me of people who think web crawling is an easy problem and OH MY GOD it is SO difficult. Bing sucked SO bad at this for a decade+)
in reply to Hakan Bayındır

Creators want to monetize their content can "just" return HTTP402 to relevant bots, again if the AI companies stop masquerading their bots as home users via distributed residential proxies and fake user agent strings.
in reply to Jeff Atwood

I will try. Try, because I'm an unpleasantly busy period of my life. On the other hand, 500 character limit makes us look like more conflicted than we are.

I'm aware of the pitfalls and shortcomings of my proposal, because it's purely technical, but the problem is mostly social.

Again, technical problems are easy, humans are hard.

The proposal I'll write will technically work until it hits real world, because of humans and tragedy of commons.

in reply to Jeff Atwood

@bayindirh You do condone it, obviously, since you're an enthusiastic user of plagiarism machines.
in reply to Major Denis Bloodnok

@denisbloodnok @bayindirh can you provide citations? I mostly use ChatGPT and Gemini. Is there any source on what percent of the training content is stolen? And how are we defining "stolen"? Like normal web crawling isn't allowed?
in reply to Jeff Atwood

@denisbloodnok

The license of the thing you crawled may not allow transformations, selling or sharing w/o citations (CC BY-NC-SA).

-or-

The license of the thing you crawled doesn't allow its inclusion in any other codebase, ever (source available).

-also-

We have torrented books, mislabeled data, etc. etc.

-so-

The case gets complicated as you want to enable exploitation of the thing you crawled, for free.

See: notes.bayindirh.io/notes/Lists…

in reply to Jeff Atwood

@bayindirh Not only do you condone it, then, but you're willing to engage in sophistry to justify it. The argument above that "not ripping off creators would be hard so let's keep doing it" is pitiful.
in reply to Jeff Atwood

Here are four easy ones to start with:

- Training on copyrighted data without compensation and/or respecting the license
- "bias washing"
- "accountability washing"
- Environmental impact

Also interested to hear what other ethical issues you've considered.

in reply to Nik

@nikclayton see my other replies on this; licensing is a nightmare mode problem (just try licensing music for a tiny taste of how complex this is, and now it's 2-3 orders of magnitude more difficult) .. as far as the rest, all things that should be labelled for all users as "this product has the following ingredients" so you know what you are ingesting. Environmental impact, not sure, I think local rules should be followed, but big picture I am very bullish on solar due to the graphs showing solar wildly outpacing even the most optimistic predictions.
@Nik
in reply to Jeff Atwood

I've read your other replies on the thread, where you repeatedly state that licensing is a "nightmare problem".

1. Is your position that this problem is so difficult that AI model builders should just ignore licensing and ingest stolen content?

2. If that's not your position, do you agree that the vast majority of AI models are built on stolen work?

3. If you agree with (2), then do you think it is ethical to use AI models that you know are trained on stolen work?

in reply to Nik

@nikclayton

  1. No. But having a lot of freely available creative commons data and content is good for the world in all possible futures, full stop.
  2. It depends on the model, and how ethical the builders are. If it's Zuck, then yeah, they are completely amoral and do whatever it takes to win. They literally don't care about anything else except, ya know, "winning".
  3. Again, I'd need them to label the contents of the product I'm about to ingest accurately, and .. are they? We should require them to.
  4. Can we provide potential viable solutions to the problems rather than just endlessly complaining about them? Also, honestly acknowledge that licensing is INCREDIBLY complex.
@Nik
in reply to Jeff Atwood

The fact that it's not completely useless aggravates the effects of the bubble, not mildens them. I think bubble will pop because even though it's useful, the profits doesn't make up to the cost of making those models. Current market value of AI companies is at least 150x of the revenue right now, while it was at best 10x pre-AI. There is no indication whatsoever that the companies will make up that much revenue in anytime soon.

I highly recommend reading profgalloway.com/bubble-ai/

in reply to Aiono

oh that fuckin' guy. Missed the point by 10 miles. He thinks he's a whole lot smarter than he is. Here's the thing: it's FAR more than "not completely useless". It's like 20% productivity boost for a lot of stuff (depends what it is). That's MASSIVE.
Questa voce è stata modificata (2 giorni fa)
in reply to Jeff Atwood

20% is still a lot less than the investment increase. Also that's also just your claim, while there are studies for the contrary metr.org/blog/2025-07-10-early… . I know you say it's not about coding but the study shows that self assessments are in general unreliable.

Your language is completely rude. I just gave a different opinion than yours and you are talking disrespectfully. My argument might be bullshit, but then you can call out my argument not the person.

in reply to Aiono

@aiono oh, there's still a gold rush mentality, for sure. Also I think GALLOWAY is bullshit, NOT you! And he is. This dude is one of those "I must build the brand of {my name}" at all costs people, like Ezra Klein. Not a fan. Of Galloway! I like you!
in reply to Jeff Atwood

i don’t understand how it cannot be both useful technology and a bubble that burst though, those are not mutually exclusive.
in reply to Jorge Salvador Caffarena

@jorgesalvador because the previous bubbles were built on sand, stuff that was bullshit, there was no "there" there.
in reply to Jeff Atwood

still doesn’t explain it. You can have something useful and make companies overspend massively, create a bubble, burst it, and still the technology is useful and will still be afterwards no matter.

Housing was, is, will be useful after the 2008 bubble.

Questa voce è stata modificata (1 giorno fa)
in reply to Jorge Salvador Caffarena

@jorgesalvador LLMs are, is, will be useful forever. Compute time (for ANY purpose) are, is, will be useful forever. Carry on.
in reply to Jeff Atwood

yes that is what I am saying, you are repeating what I say, so it’s not the counter argument you expect?
in reply to Jeff Atwood

this would need an in person conversation, this is going nowhere as I’m not explaining myself clearly, or you are not getting my point.
in reply to Jeff Atwood

what's your setup? I'm skeptical but I might have used the wrong tools 🤷
in reply to Jeff Atwood

I've used LLMs to successfully input data into my brain. Analyze, compare and basically make sense of data in various shape from multiple sources. I even used it to generate common patterns to guide my learning of a craft like languages and technology.

I have yet to use it successfully to output anything from my brain though. Be it writing code or an email, the mechanics of transferring my thoughts to a destination format is hardly ever the limiting factor : The bottleneck is my brain. It's the silver bullet problem all over again.

The solution seems easy. Bypass the brain and have the LLM go from input to output on its own. Luckily, we have hundred of vibe coders live streaming their fall from optimism to show that's a bad idea. If I don't understand what the machine is doing, there is no way I'll trust that work, at least until there is a revolutionary leap in the technology.

That leaves one area I can think : Have it challenge my output. I can imagine significant incremental gains in productivity there, but I haven't had the chance to try any offering like that for either for code or prose...

in reply to Jeff Atwood

But If it's useful or nor IS Not the whole point to decide if it's a bubble?

I find it useful, and I would pay about 3-5$/month max for the usefulness it provides for me. So we will see, if they are either able to operate the text generators for such prices or if there are many people who are indeed willing to pay >100$+ per month and seat.

This is IMHO the question the coming months/years have to answer to decide bubble or not or how big. I don't know the answer, we will see.

Questa voce è stata modificata (1 giorno fa)
in reply to Frank

@frank yeah, it gets down to compute time costs, and I am worried because GPUs ain't getting cheaper. If anything the opposite. We need a LOT more competition in GPUs.
in reply to Jeff Atwood

Honest question: how are are you seeing it making people more effective?

I work in tech, and in the last three years I've seen it not only being adopted, but also made mandatory in some cases, in three different companies. At this point, everyone is using LLMs for one or other thing.

What I have not seen is any significant gain in productivity. If any. We don't ship faster, we don't produce less bugs, we don't communicate better, our documentation is not better than it used to be. (Some) People enjoy using it, for sure, it makes their job funnier... But I haven't seen it making them more effective.

As you say, it can save a bit of time in some tasks. I've also seen it create messes that sucked up an entire team days to solve. So I'm not sure of the overall result being too spectacular.

in reply to javi

@javi examples were provided here, have a look infosec.exchange/@codinghorror…


here's one: a friend confided he is unhoused, and it is difficult for him. I asked ChatGPT to summarize local resources to deal with this (how do you get ANY id without a valid address, etc, chicken/egg problem) and it did an outstanding, amazing job. I printed it out, marked it up, and gave it to him.

Here's two: GiveDirectly did two GMI studies, Chicago and Cook County and we were very unclear what the relationship was, or why they did it that way. ChatGPT also knocked this out park and saved Tia a lot of time finding that information out, so she was freed up to focus on other work.

I could go on and on and on. Email me if you want ~12 more specific examples. With citations.

But also realize this: I am elite at asking very good, well specified, very clear, well researched questions, because we built Stack Overflow.

You want to get good at LLMS? learn how to ask better questions of evil genies. I was raised on that. 🧞


@javi
in reply to Jeff Atwood

That's fair, yeah, that's the kind of things I've seen people using it for successfully LLMs for. Nothing you probably wouldn't have gotten to on your own after 15/30 minutes of online searches.

I'm not saying they are not useful, but I see two major problems that make me be skeptical of how important this technology is:

First, both of your examples show cases where an incomplete/slightly wrong answer is not really a deal breaker: if some of the resources in the first example were outdated, or if the second answer would had missed some extra connection, the answer would still be valuable.

The problem that is usually not "good enough" for a lot of "work" problems, and I often see people spending almost as much time validating and correcting LLM outputs as if they have not used an LLM to begin with. Not in every case, of course, but it makes their "productivity boost" effect diminished by how important is the task they are being used for.

On the other hand, in a more broad terms, it's the economics issue. Right now, to get the current level of functionality, The different providers are burning cash at a literally never-seen-before rate. Unless there is a real and unexpected breakthrough in the technology, there is no way to keep it at the current cost for the end user in the long term. I know plenty of people who pay $200/month, or get a license from their employer that pays something like that for them. Are they going to be willing to pay $1k or $2k for the current (or even slightly improved) capabilities ? I'm highly skeptical.

So I can't help but not seeing any of this as ground breaking as you folks seem to see it

in reply to javi

@javi "Nothing you probably wouldn't have gotten to on your own after 15/30 minutes of online searches. " correct, so I saved 30 (more like 60, really, being unhoused is a VERY complex topic) minutes.. twice. That's worth a LOT of money, not an infinite or ridiculous amount, but it clearly has a lot of value. Compare with crypto. Zero value to anyone in the world, other than inveterate gamblers.
@javi
in reply to Jeff Atwood

Jeff, you're a nice, intelligent, and terribly geek person I would love to discuss with in details (about plenty subjects).
But, on this particular subject, I think you may have a too much narrow view on the subject.
LLM are pushed by top-level CEO because it allow their wet dream of an intelligent, but non-reflecting, task force to exist.
in reply to Nicolas Delsaux

I have many other reasons to think that, on multiple levels, LLMs are a bubble, and maybe I'm wrong, maybe there are legitimate, good use cases, of these alorithms. Nevertheless, I think they're currently horsemen of the apocalypse
in reply to Nicolas Delsaux

@Riduidel "LLM are pushed by top-level CEO because it allow their wet dream of an intelligent, but non-reflecting, task force to exist." they will learn the hard way that this doesn't work. Will some extremely repetitive jobs that no human should have anyway be eliminated? hopefully, so they can be replaced with much more meaningful and interesting jobs for humans to do.
in reply to Jeff Atwood

it can be useful and still a bubble! I mean tulip bulbs could be still grown even if their valuation changed dramatically.

AI companies are doing seriously some questionable financing, laundering valuations to pull more funding... It seems like some of the big providers are on increasingly shaky financials. We'll still have AI, of course, but their valuations might change dramatically.

in reply to Mike Johnson

@mrj tulips are NOWHERE near as useful as a 20% boost in productivity.
in reply to Jeff Atwood

But did you really shoulder the cost of the use of the LLMs? I wager you didn't, I wager it was VC money. Delivery pet food was futilely propped up by VC money in the 90s, too. It's an every day reality today.

The dot-com bubble bursting didn't kill the WWW. This bubble bursting won't kill LLMs or AI in general. But it will slow it down a lot and wipe out a bunch of investor-class money, which will kill VC confidence for a while, I predict.

in reply to Jeff Atwood

I look forward to the consoling productivity boost for the legions of workers fired because of managements' dedication to realizing the promises of salesfolks' closings.
As those workers are rehired to the jobs the LLM could never do, I'm certain they've invested their generous severance in becoming domain experts so that they are now the premier LLM output-validators, deftly capable of approving generated summaries without the need to read primary sources at all
/s
Questa voce è stata modificata (1 giorno fa)
in reply to Jeff Atwood

Whether or not it's useful is not entirely related to whether it's a bubble. OpenAI is predicting it will need to charge $5,000/month for a BASIC professional user. For more advanced models, OpenAI is predicting $10,000 (coding) - $20,000 (PhD) per month. How many people currently paying ~100s/month are going to switch to thousands? They've also been up front that they're losing money on individual pro accounts that are $200/month, and are DEFINITELY getting hosed by the 90% of users on the free tier.

Furthermore, while AI models have been getting better, they've ALSO been getting more expensive. Sustainable technologies tend to get cheaper over time as they get better, and AI is just not. OpenAI just signed a deal with Nvidia for building out massive new datacenters that will require as much electricity as 10 nuclear reactors can produce and cost more than half a TRILLION dollars just to build.

AI is a bubble because the finances do not work, and cannot work without a massive reorganization of the industry (i.e., a crash). The companies are vastly overleveraged with debt and floating along on investor money that's going to dry up. Their physical assets are largely computing datacenters that are obsolete almost as soon as they're built and full of computing hardware that burns out faster than any other hardware in existence. Their finances only look as "good" as they do--and they don't look good at all--because they're performing weird accounting tricks like having Microsoft "invest" in the company, then using that money to pay MS for Azure; same with Nvidia and graphics cards. The money is counted by each company simultaneously as payment and investment and profit and cost, but it's all just the same money getting passed around the same companies.

Sooner or later--but even The Wall Street Journal is starting to predict sooner--the 1/3 of the stock market that's wrapped up in AI is going to have their shareholders demand a return on investment, and the money is not there to be returned. They almost literally are lighting it on fire in the form of burned out graphics cards and coal power.

Some of AI is going to be salvaged, but that doesn't mean this isn't a bubble (just like dot com).

in reply to Azuaron

@Azuaron agreed, there are ABSOLUTELY fundamental compute cost issues, since GPUs are about the only computer part I can think of that keep getting more and more expensive over time. This is an anomaly.