Ars Live recap: Is the AI bubble about to pop? Ed Zitron weighs in.

Despite connection hiccups, we covered OpenAI's finances, nuclear power, and Sam Altman.

See full article...

 

[odikweos]

I have a friend, fellow entrepreneur, whatever, who has been desperately trying to build a product with this crap. He just keeps thinking something vaguely working is so close. It's been like 8 months. I won't even meet with him anymore - I just say "ok so where's the 5 minute video showing it working I asked for back in January?" -- and I get nothing at all.

Because the shit doesn't work. It won't work. They just don't want to believe Santa Claus isn't real.

 

[MHStrawn]

Maybe I'm not giving Team Silicon enough credit; but I suspect that comparing the(admittedly super heroic) essentially-obsolete-when-finished last gasp of large scale tube computing to contemporary very large scale integration will...not be representative...of the efficiency gains that are actually within near to middle term reach when you are already fighting with the wavelengths of your increasingly recalcitrant light sources to get sub-5nm photolithography.

This sounds exceptionally compelling.

I wish I understood one word of it.

 

[MHStrawn]

I like your accessibility argument here as well. It might not be a trillion-dollar market, but it's still a valuable one. LLMs do a pretty good job translating from one context to another, and if the cost of inference goes does down (like through distilling models, as mentioned in another recent article), accessibility could be a good use case.

That's the thing...a $50B industry is a HUGE industry! If people were treating it like that I doubt Zitron would be so apoplectic. But CEOs have been convinced AI can solve anything and rid them of those pesky human employees and much of the public feels like AI is inevitably going to take every job and rid the world of human artists.

AI has been around for a while now. Revolutionary digital technologies of the past had already demonstrated sustained impact on multiple industries....AI hasn't. A NYT writer wrote more than a year ago "AI is mid". Not that it's useless but it's just another tool for humans to do things faster and more efficiently.

"AI is just another tool" doesn't inspire $1Trilliong in investments, though.

 

[HardCover]

I think he's right about the bubble bursting, but an economist I follow has pointed out that the late 1900s and early 2000s internet bubble was recognized, and warned against, three years before it burst. He pointed out that we're only about thirty percent of the way past the recognition of the bubble, if the pattern holds. Which it may not, because things are not the same now. And while the internet bubble burst, it later led to some major economic efficiencies and made billions of dollars for some people, and millions for many more.

Bitcoin and other digital currencies are also probably in a bubble, and have been for a decade. That may also burst, but nobody knows when.

The big question for a lot of people is, "How do I protect myself financially?" They want to be in Nvidia, and other AI companies, but an increasing number of them recognize that they may make some money riding the upward tide, but they could lose it all if they go to the bathroom at the wrong time. Or if the "buy on the dip," which turns out to be a canyon.

 

[MHStrawn]

A lot of valid points, though I’d point out that Tesla has been sailing on sand castles for over a decade and smart people know it’s a bubble, built on empty promises, doesn’t stop it. It seems pretty hard to spook the market these days, even tariffs have had enough of the cry wolf.

All of that is to say don’t underestimate the insanity of the markets.

On the one hand I agree with you...the financial markets largely stopped operating like financial markets when all the 401K money started pouring in in the 90's...and now operates more similarly to a casino. Which means you can have poorly run relatively unprofitably companies like Tesla enjoy grossly inflated values for years and years and years. Like the famous Dutch tulip market, as long as people think it's valuable, it's valuable.

But Tesla is an actual company producing actual products. And at one time they were a cutting edge company seemingly producing the cars of the future. AI is largely vaporware lacking a single killer product that can justify the vast amounts being invested. I think Ed is right that at some point the investors are going to turn the spigot off or slow it to a trickle and when that happens these companies are going to be hard pressed to continue to exist in any form like they are now.

 

[Jim Salter]

This sounds exceptionally compelling.

I wish I understood one word of it.

Translation: don't expect another improvement in computing power on the order of the transition we made from vacuum tubes through discrete transistors, then integrated circuits, then Very Large Scale Integrated Circuits.

Still too technical?

"Yes, you're eating more cookies at once than you ever have before, congratulations--but you're now eating cookies at the same rate that we can grow wheat and produce flour, and we're growing as much wheat as the planet can produce, and milling it as fast as it can be harvested."

 

[Jim Salter]

That's the thing...a $50B industry is a HUGE industry! If people were treating it like that I doubt Zitron would be so apoplectic.

This is how I felt about Tesla, back before Musk bought Twitter, and my biggest concern was how he kept saying the car literally didn't even matter, only full self driving did.

 

[adrianovaroli]

“Zitron is unfairly dismissing counterexamples”

The author is using $50 billion examples, and Ed is worried about a trillion dollar bubble. I would be dismissive, too. Use a trillion dollar counter example, if you can think of one.

 

[MHStrawn]

Which it may not, because things are not the same now. And while the internet bubble burst, it later led to some major economic efficiencies and made billions of dollars for some people, and millions for many more.

Bitcoin and other digital currencies are also probably in a bubble, and have been for a decade. That may also burst, but nobody knows when.

The big question for a lot of people is, "How do I protect myself financially?" They want to be in Nvidia, and other AI companies, but an increasing number of them recognize that they may make some money riding the upward tide, but they could lose it all if they go to the bathroom at the wrong time. Or if the "buy on the dip," which turns out to be a canyon.

Yeah, I doubt this bubble will be like a previous bubble. I mean, the stock market lost 40% of it's value in a flash during dotcom crash. I could see it being worse this time bc so much of the biggest tech companies are vulnerable here and they're tendrils reach into so much of the rest of the economy.

As a 60 YO with substantial assets I'm really struggling to find a sound strategy. In the past I would have simply put my money in bonds and t-bills but even those formerly safe havens look perilous due to Trump tariffs deprecating the power of the dollar.

 

[terrydactyl]

One of Zitron’s most pointed criticisms during the discussion centered on OpenAI’s infrastructure promises. The company has pledged to build data centers requiring 10 gigawatts of power capacity ... for its Stargate project in Abilene, Texas.

For reference, the state of California's peak demand run about 25 gigawatts, so they are talking about 40% of a state with 40M people!

 

[dobrien75]

but I think revenue will grow dramatically faster than he thinks

based on what though?

His depreciation figures are crazy, but only because I'd never thought about the lifecycle issue with data centres.

 

[josephhansen]

what is wrong with you.

Yes, this was a deeply unprofessional question for Mr. Edwards to ask. I'm surprised Mr. Zitron even graced this off-topic sycophancy with a reply

 

[C.M. Allen]

We don't have the evidence that AI is like the internet, at least not in its current form. Not everyone is going to make significant use of AI, because it isn't right for them.

A lot of research from non-AI companies are showing either no real world gains or even worse poorer performance caused by the use of AI.

In fact, the opposite. The decentralized nature of the internet, combined with a low barrier to entry, meant that even though it was in a bubble, there was plenty of ground beneath it once it popped. The same is simply not true of AI. They require massive data farms and unfathomably vast sums of data to train and operate. None of that is remotely achievable by any entity outside of massive tech firms...who will be the first to jump ship once they realize there's no money to be made with AI. And without those giant tech firms spending billions on AI hardware, software, etc, there's nothing there to sustain any of the market. Small hobbyists don't even represent a rounding error in that equation. They won't be able to sustain the hardware side of the market. They won't be able to sustain the software side. And they won't be able to manage the data sets. Without those big tech firms throwing money at the 'market,' this generation and form of AI is more or less dead in the water. It's why those tech firms are trying so desperately to throw AI at anything and everything they see in the hopes that they can stave off a total collapse long enough for a major breakthrough or 'must have, killer app' to come along and save them.

 

[aaronspink]

Excellent article, thank you. I agree that “A 50 billion-dollar industry pretending to be a trillion-dollar one” is probably the best financial take.

I like your accessibility argument here as well. It might not be a trillion-dollar market, but it's still a valuable one. LLMs do a pretty good job translating from one context to another, and if the cost of inference goes does down (like through distilling models, as mentioned in another recent article), accessibility could be a good use case.

LLMs are meh at best for translating most languages... It can handle kissing cousins English<>french ok as long as it isn't anything technical or literary, but with most other languages it is just horrid, sure better than no translation, but basically the same if not worse than machine translation... Between the lack of temporal knowledge, subject/object/observer inversions, and straight up hallucinations its pretty bad.

FDPR systems are good for one thing: recognizing patterns within chaff. Which shockingly is why they were developed originally. but they are way overused and over-applied to things that there are much better machine learning models. general AI is a complete dead end because the fundamental models they are using (FDPR) to achieve it is not intelligent.

 

[benjaminoakes]

Yes, but that was a one-off. That is not going to happen again because we are at ~18 Angstrom process and the average aluminum and silicon atom is closer to 3 Angstrom. We just can’t keep making the features smaller. Yes the light can go smaller. Yes some structures are still big, but to continue to expect more than 1-2 order of magnitude improvement is not realistic.

I agree, and I hope we can figure out novel ways of approaching LLMs which vastly decrease the computing requirements. I saw something about Anthropic and a distillation process today, for example. There's bound to be many discoveries.

 

[Dputiger]

In fact, the opposite. The decentralized nature of the internet, combined with a low barrier to entry, meant that even though it was in a bubble, there was plenty of ground beneath it once it popped. The same is simply not true of AI. They require massive data farms and unfathomably vast sums of data to train and operate. None of that is remotely achievable by any entity outside of massive tech firms...who will be the first to jump ship once they realize there's no money to be made with AI. And without those giant tech firms spending billions on AI hardware, software, etc, there's nothing there to sustain any of the market. Small hobbyists don't even represent a rounding error in that equation. They won't be able to sustain the hardware side of the market. They won't be able to sustain the software side. And they won't be able to manage the data sets. Without those big tech firms throwing money at the 'market,' this generation and form of AI is more or less dead in the water. It's why those tech firms are trying so desperately to throw AI at anything and everything they see in the hopes that they can stave off a total collapse long enough for a major breakthrough or 'must have, killer app' to come along and save them.

The idea seems to be something like this:

Improvements in GPU performance (or the emergence of non-GPU processors who can handle AI more effectively) + improvements in model quantization + improved accuracy from models that don't require such aggressive retraining or can be updated piecemeal + further algorithmic improvements = much cheaper AI.

And I think there is some reason to expect improvements where such things are concerned. I've read some articles suggesting that the "Let's build smarter" approach was ignored for much of the last few years in favor of "Let's throw more data at it," because smarter is always harder and takes longer. Now that high quality data is increasingly hard to come by, companies are taking another look at the problem. Nvidia's dominance of AI is related to its dominance of CUDA, so there's the possibility that another company could develop a more efficient architecture, or that NV itself might create such a thing. Algorithmic improvements are likely given how new AI is, and if companies weren't constantly retraining entire models from the ground up, the GPU requirements and costs would not be so stratospheric.

To me, the question is: Is there enough runway left in the bubble for those improvements to be developed before the bottom drops out of everything? If inference is 10x cheaper in 2030 than in 2025, is that enough? What about 50x? What if inference costs drop but training costs don't?

Right now, everyone is hurling money at AI because each of these companies is terrified of what happens if AI works for a competitor and they aren't onboard. There's a perception that AI could revolutionize computing, but I don't think the primary driving force is optimism -- it's fear. Whoever figures it out first wins the ultimate first-mover advantage.

 

[runnerd00d]

To me the "money" from AI will come more from manufacturing a bit down the road. At the expense of human labor most of the time, also. Me asking Copilot for some random bullshit isn't making anyone money 99.99% of the time, but with manufacturing there are many ways AI can be used create more efficient and far more effective robotics that learn and can manage many different tasks at once, depending on the need. That's my take on AI for now. It will displace a lot of manufacturing jobs including in lower cost countries. Just how long that will take...I have no idea. I'll defer to other smarter and far more knowledgeable people in this chat.

 

[ebolasoup_zaire]

Benj the AI isn’t gonna love you back. Unless you pay for Grok Pro Premium+

 

[doubleyewdee]

Zitron pushed back on this optimism, saying that AI costs are currently moving in the wrong direction. “The costs are going up, unilaterally across the board,” he said.

I feel that the question he answered (around hardware and ecosystem efficiency improvements) with this response was pretty disingenuous. For example, if you wanted to take the inference cost/token, you really need to factor in model complexity as part of this. You could use, say, model parameter count as a basic baseline, maybe mix in reasoning capability, etc. In those cases, comparing GPT3 to GPT-5 on a cost/token basis looks bad if you do not factor in the increased sophistication of the two models. Now, if you want to say "the juice isn't worth the squeeze" (you do not get +$x value going from GPT-3 -> GPT-5 commensurate with the cost), that's an argument to make. However, it is not the argument that was made.

The raw cost-to-inference of the same model is absolutely going down over time. The delta is that models are changing, reasoning models will make multiple inferencing passes, etc. So, yes, "the costs are going up" for the absolute latest foundation model invocations, but there are plenty of scenarios that are served suitably by older models, non-reasoning models, purpose-built distilled models (e.g. Claude Haiku, the OpenAI -minis, etc), and so on.

I really appreciate this article and the attempt to provide a nuanced discussion about the state of AI, the genuinely real and dangerous speculative investment bubble happening, and so on. I just wish it had been done with someone who was ready to speak in anything but absolutes fueled, seemingly, by at least some amount of emotional attachment to a specific and entirely antagosnitic view with no room for flexibility of viewpoint.

In fairness to EZ, I imagine his counterpoint would be that this is akin to pushing back on, say, climate science denialism bullshit. At some point you see no value in entertaining falsifiable nonsense or discussing "nuance" with bad faith actors. I just don't believe the "is AI good/bad/neither?" discourse featured here is in the same realm, nor that Benj or others who do find periodic value from using AI tools are somehow acting in bad faith.

 

[doubleyewdee]

The big question for a lot of people is, "How do I protect myself financially?" They want to be in Nvidia, and other AI companies, but an increasing number of them recognize that they may make some money riding the upward tide, but they could lose it all if they go to the bathroom at the wrong time. Or if the "buy on the dip," which turns out to be a canyon.

Hedge your bets (investing is just gambling with a gentel veneer) and diversify. Not super complicated. You lose out on stratospheric highs, but you're less exposed to collapses.

 

[aaronspink]

Yeah, I doubt this bubble will be like a previous bubble. I mean, the stock market lost 40% of it's value in a flash during dotcom crash. I could see it being worse this time bc so much of the biggest tech companies are vulnerable here and they're tendrils reach into so much of the rest of the economy.

As a 60 YO with substantial assets I'm really struggling to find a sound strategy. In the past I would have simply put my money in bonds and t-bills but even those formerly safe havens look perilous due to Trump tariffs deprecating the power of the dollar.

Its going to be bad because everything AI related is a huge Enron level cluster that will multiplicatively decay. if any part of it starts to get a cold, the death spiral will be nearly instant because everything is circular.

 

[j00ce]

The thing is, he's also dismissive of the $50b example portion. He's basically saying "this stuff has effectively no value." He even says at an adjacent point "I’m ready to accept issues, but AI is all issues, it’s all filler, no killer," implying that AI does nothing of value at all, ever. This is just absurd. People are finding value in these systems, which is a big part of why the speculative bubble has grown to ridiculous proportions.

There's a lot of AI which has some degree of value. The question is: is the amount of value worth the true cost of said AI?

The Register published an article a couple of days back, about an AI startup called Augment which changed from a set of "all you can eat" options, to charging "per message".
https://www.theregister.com/2025/10/15/augment_pricing_model/

However, per-message charging didn't take into account how many tokens were needed to process said message(s):

A new post from CEO Matt McClernan said [...] one user on the $250 max plan is costing the company "approaching $15,000 per month

As such, they've switched to a credit-based system, which more closely maps to token usage. But even then: a user who was paying $250 per month had this to say:

"Over the last 7 days, your usage totaled 31 user messages, corresponding to 40,982 credits under the new pricing model. This equals an average of 1,322 credits per message."
...
So, my 4500 monthly messages EQUAL NOWHERE NEAR the 520k credits they are giving us for the same cost.

Or to put it another way: where $250 previously paid for 4500 messages, at an average of ~1300 credits per message, their new 520k limit only lets them generate around 400 messages.

To be fair, at 30 messages a week, they're comfortably under this limit. But equally, it's still effectively over a tenfold drop in what they're getting for their money.

So, what happens if/when the rest of the industry starts to put its prices up?

To take a quick example: Cursor's Pro+ pricing scheme costs $60 for ~1,500 GPT 5 requests, which is roughly in the same ballpark as Augment's "messaging" pricing scheme, which charged $100 for 1500 messages.

Which means it's not unreasonable to assume that the same ten-fold increase in cost will apply.

How many people will be willing to pay $600 for that same service, if/when Cursor decides it needs to start making a profit?

LLMs may well be a $50 billion industry. But the people investing in it are currently expecting a trillion dollar return. And that's not going to happen.

 

[crawler_carl]

I expressed a belief that criticism over the cost and power requirements of operating AI models will eventually not become an issue.

The argument that compute costs (or model compute requirements, for that matter) may decrease significantly making the economics of the industry work is not taking into consideration something important: if anyone can run a model on lower level hardware, Nvidia and OpenAI are toasted!
That's the main reason the market wobbled earlier this year when DeepSeek released.

 

[crawler_carl]

Adding to the last message, the stocks valuations therefore assume the cost of the AI industry will not change significantly. Which makes little sense to me considering it's balance sheets are terrible.

 

[BoCaro]

OpenAI is on track for profitability. I heard they will do porn soon.

 

[Snark218]

The thing is, he's also dismissive of the $50b example portion. He's basically saying "this stuff has effectively no value." He even says at an adjacent point "I’m ready to accept issues, but AI is all issues, it’s all filler, no killer," implying that AI does nothing of value at all, ever.

If you define value as "delivers a tangible benefit in excess of whatever it costs to provide," yeah. Yeah, man. That's the point. Of course AI does stuff. Does it do stuff of net value? Does it produce more value than what it costs to train and run an LLM? No.

This is just absurd. People are finding value in these systems,

Not hard when there is no cost to the user.

which is a big part of why the speculative bubble has grown to ridiculous proportions. I

The fact that individual users are finding uses for LLMs has nothing to do with why the speculative bubble is so inflated. Nothing at all.

t's not a reasonable position to take. You can say "80% of AI trials in [xyz scenario]" are failures," it's easy to find case studies showing that "slap an LLM on a workflow" isn't a panacea or an instant improvement. This isn't how he frames it, though. He's calling out an expected 100% failure rate, with no upside in the future.

No. He's saying that whatever trivial bullshit things people are putting AI to work on, they would not actually pay what it costs to train and run that LLM if they were not being subsidized by venture capital.

It's, frankly, a bit of an unserious position given the rate of adoption and relative stickiness of customer workloads. Yes, there's a lot of failures, but there are also many successful cases, and there's a reason people come back to the tools and OpenAI, Google, AWS, Microsft, etc have already accumulated those billions in revenues in a short time-frame. Growth of those numbers will likely stagnate as speculation and specious projects dry up, at least for a time, but the "less insane" (imo) view of things is that the core technology isn't going away wholesale, and will very likely see increased use over time if/until supplanted by [next technology thing that gets its own stupid bubble].

Billions in revenues aren't impressive if they're offset by tens of billions in costs.

At least, this is how EZ comes across to me on this topic. Just utterly dismissive. Not cautionary, not "this hype has gotten out of hand," but entirely and completely in the "this is all garbage and should go away entirely" camp. I use AI for work and for assistive / accessibility things every day of my life now, and so I struggle to take that framing seriously.

Would you pay $2000 a month for "work and assistive/accessibility things?" Call me doubtful.

 

[isparavanje]

And what they are selling as AI is quite literally 1960s technology. There is nothing new in "AI", there wasn't 20 years ago and there wasn't 40 years ago.... Yes the models are bigger, but the models aren't actually better because the technology of the model is the same. Its just a bigger pile of poop. So sometimes it will have something in it that the previous smaller piles of poop didn't. But its still just poop.

Transformers are pretty new (2010s), and earlier RNNs (eg. LSTM) did not take off both because they were terrible at long context and because they were sequential and hence terribly slow to train. Even RNNs are a modern architecture. I mean sure, they use backprop under the hood, but if you're just going to take a random piece and say it's that old you might as well say it's 150BCE because it's linear algebra.

 

[isparavanje]

There's a lot of AI which has some degree of value. The question is: is the amount of value worth the true cost of said AI?

The Register published an article a couple of days back, about an AI startup called Augment which changed from a set of "all you can eat" options, to charging "per message".
https://www.theregister.com/2025/10/15/augment_pricing_model/

However, per-message charging didn't take into account how many tokens were needed to process said message(s):

As such, they've switched to a credit-based system, which more closely maps to token usage. But even then: a user who was paying $250 per month had this to say:

Or to put it another way: where $250 previously paid for 4500 messages, at an average of ~1300 credits per message, their new 520k limit only lets them generate around 400 messages.

To be fair, at 30 messages a week, they're comfortably under this limit. But equally, it's still effectively over a tenfold drop in what they're getting for their money.

So, what happens if/when the rest of the industry starts to put its prices up?

To take a quick example: Cursor's Pro+ pricing scheme costs $60 for ~1,500 GPT 5 requests, which is roughly in the same ballpark as Augment's "messaging" pricing scheme, which charged $100 for 1500 messages.

Which means it's not unreasonable to assume that the same ten-fold increase in cost will apply.

How many people will be willing to pay $600 for that same service, if/when Cursor decides it needs to start making a profit?

LLMs may well be a $50 billion industry. But the people investing in it are currently expecting a trillion dollar return. And that's not going to happen.

For coding at least, at least if we just consider the cost of inference (eg. assuming a model is trained and then the training cost is amortised over a long time because AI has stagnated and we don't train new models every few months anymore), it's easily worth it for things like code. You can see how much model inference costs on third party providers with big open models, such as if you hosted your own open models on AWS, and it's not far of from what the big companies like OAI or Anthropic charge. (API costs, not subscriptions, which they definitely lose money on)

Even with API costs, though, it costs maybe tens of cents to complete a refactoring task, and it's pretty good quality if you're doing basic refactoring tasks. Sure, it's very easy copy-paste type work, but it saves money and time, and even with API costs, at least for me, it's not going to be anywhere near your hypothetical costs. Probably high tens-low hundreds per user per month, and I, and probably many software companies, will happily pay that because it's negligible compared to the cost of an employee, so even a percent-level efficiency boost is worth it.

A lot of the maths you're doing for AI subscriptions will also give insane numbers for Google drive or Dropbox subscriptions, or many other digital subscriptions where the average user isn't expected to use the subscription to the max. (eg. How much do you think is the value of the movies watched by the most prolific Netflix users? Netflix almost certainly loses money on them)

 

[jdale]

There's a lot of AI which has some degree of value. The question is: is the amount of value worth the true cost of said AI?

The Register published an article a couple of days back, about an AI startup called Augment which changed from a set of "all you can eat" options, to charging "per message".
https://www.theregister.com/2025/10/15/augment_pricing_model/

However, per-message charging didn't take into account how many tokens were needed to process said message(s):

As such, they've switched to a credit-based system, which more closely maps to token usage. But even then: a user who was paying $250 per month had this to say:

Or to put it another way: where $250 previously paid for 4500 messages, at an average of ~1300 credits per message, their new 520k limit only lets them generate around 400 messages.

Paying per token seems sensible, certainly it works for the company. But for the user? If I ask an AI to produce something, and it makes significant mistakes which I have to tell it to correct, that's a little bit annoying when there's no additional charge, but a big deal if every iteration is costing me more money. And at an accelerating rate since it needs to re-process all the previous tokens each time. Each query is more expensive than the last.

To be fair, at 30 messages a week, they're comfortably under this limit. But equally, it's still effectively over a tenfold drop in what they're getting for their money.

So, what happens if/when the rest of the industry starts to put its prices up?

To take a quick example: Cursor's Pro+ pricing scheme costs $60 for ~1,500 GPT 5 requests, which is roughly in the same ballpark as Augment's "messaging" pricing scheme, which charged $100 for 1500 messages.

Which means it's not unreasonable to assume that the same ten-fold increase in cost will apply.

How many people will be willing to pay $600 for that same service, if/when Cursor decides it needs to start making a profit?

LLMs may well be a $50 billion industry. But the people investing in it are currently expecting a trillion dollar return. And that's not going to happen.

I'm not even willing to pay for it at all at this point, so....

But, hey, Microsoft is building it into Windows so it will be free, right?

 

[Bilateralrope]

If you like ai, consider that enshitification will ruin it like everything else.

Yes. The shift from a growth focus to a profit focus will pop the bubble if nothing does before then.

"Pretty much ever corporate shop in town is paying for it for their devs" - right now they are, because it's priced (almost certainly) unprofitably low, and every C-suite and investor in the world is singing from the "you cannot possibly spend too much money on AI" songbook. Like, it's hard to overstate this effect. My company never wants to spend any money on anything, but right now if you say you want to spend some money on AI the only pushback you get is "are you sure you wouldn't like to spend more?"

That's not going to last, though. There's no way it can last. At some point, they have to increase the price to something higher than the service costs to provide. At some point, C-suites and investors will get tired of setting fire to money and start asking questions like "is this worth the much higher price we are now paying for it"?

I mean, Github is owned by Microsoft and it makes money (I think), it will be just fine. But I don't think you can state, based on current information, whether or not Copilot will turn out to be a long-term viable business. We just don't have the information to know.

From memory, there have been plenty of developers complaining that upper management requiring them to use AI led to more work fixing whatever the LLM excreted than it would have to have the devs code it themselves in the first place.

LLMs are meh at best for translating most languages... It can handle kissing cousins English<>french ok as long as it isn't anything technical or literary, but with most other languages it is just horrid, sure better than no translation, but basically the same if not worse than machine translation... Between the lack of temporal knowledge, subject/object/observer inversions, and straight up hallucinations its pretty bad.

FDPR systems are good for one thing: recognizing patterns within chaff. Which shockingly is why they were developed originally. but they are way overused and over-applied to things that there are much better machine learning models. general AI is a complete dead end because the fundamental models they are using (FDPR) to achieve it is not intelligent.

We had an article recently about Reddit's LLM screwing up English <> English translation.

 

[aaronspink]

Transformers are pretty new (2010s), and earlier RNNs (eg. LSTM) did not take off both because they were terrible at long context and because they were sequential and hence terribly slow to train. Even RNNs are a modern architecture. I mean sure, they use backprop under the hood, but if you're just going to take a random piece and say it's that old you might as well say it's 150BCE because it's linear algebra.

doing addition via abaqus or via pen and paper is still doing addition. Fundamentally what these models are doing is no different from the original FDPR systems of the 60s, they are faster and bigger but still exhibit all the limitations and issues inherent in the baseline technology. Its why they fail in the same ways that they've failed for decades.

 

[Sideros]

“Zitron is unfairly dismissing counterexamples”

The author is using $50 billion examples, and Ed is worried about a trillion dollar bubble. I would be dismissive, too. Use a trillion dollar counter example, if you can think of one.

The trillion dollar example is that the big AI players get their slop embedded in so many systems (especially government) that we have no choice but to deal with it or risk massive infrastructure and economic collapse all at once.

 

[Bilateralrope]

The trillion dollar example is that the big AI players get their slop embedded in so many systems (especially government) that we have no choice but to deal with it or risk massive infrastructure and economic collapse all at once.

But will that prevent the bubble bursting ?

Or just delaying it while making the eventual burst worse ?

 

[Dusfud]

Hmm.. I lost a little respect when he whiffed the easiest softball question of all:

Q: "Say one nice thing about Sam Altman"

A: "Despite lying about everything, he's still a better human being, and better for humanity, than Elon Musk".[1]



[1] https://www.science.org/content/art...epal-s-malnourished-children-are-paying-price

 

[Sideros]

But will that prevent the bubble bursting ?

Or just delaying it while making the eventual burst worse ?

Maybe a smaller burst? Maybe no burst?

If the government thought NVidia or any of the big players collapsing would cause catastrophic collapse of government you bet the government would pay whatever it takes to keep that from happening.

Much like all the bailouts that have happened in the past 16+ yrs. We might avoid the massive acute economic hit, but suffer from the economic attrition of inflation.

And much like the current investing environment we'd accept the short-term gain at the cost of long-term insolvency.

So, I think it's a good trillion dollar counter argument. Doesn't mean it's correct, just a counter point.

 

[zoltan_merc]

It really can help speed up development time significantly, especially with boilerplate stuff.

Automating the generation of boilerplate is the wrong fix for the existence of boilerplate code, because it does nothing to help readers. The right way to fix the problem of boilerplate code is to redesign whatever system requires its presence.

Granted, when the system is a widely-used programming language such as Java, the right fix (choosing a better language) is typically available only at the very start of a project.

It's also good for an automated code review.

AI is totally useless at one of the main jobs of code review, which is increasing the number of people on the team who are familiar enough with a piece of code to take over its maintenance if its original author gets hit by a bus.

 

[RoninX]

AI is like the Internet in the year 2000.

• Yes, it's a bubble. Replace "Do X, but on the Internet" with "Do X, but with AI". A lot of useless companies will go out of business and a lot of stock value will evaporate.
• No, it's not a fad. It's going to change every aspect of modern life, for better or worse, like the Internet changed the way we work, play, shop, date, learn, and conduct politics.

 

[dsound]

Ah please fix your internet connection