The Bulletin of the Atomic Scientists published a report on the possible crash of the AI bubble:
Silicon Valley and its backers have placed a trillion-dollar bet on the idea that generative AI can transform the global economy and possibly pave the way for artificial general intelligence, systems that can exceed human capabilities. But multiple warning signs indicate that the marketing hype surrounding these investments has vastly overrated what current AI technology can achieve, creating an AI bubble with growing societal costs that everyone will pay for regardless of when and how the bubble bursts.
The history of AI development has been punctuated by boom-and-bust cycles (with the busts called AI winters) in the 1970s and 1980s. But there has never been an AI bubble like the one that began inflating around corporate and investor expectations since OpenAI released ChatGPT in November 2022. Tech companies are now spending between $72 billion and $125 billion per year each on purchasing vast arrays of AI computing chips and constructing massive data centers that can consume as much electricity as entire cities—and private investors continue to pour more money into the tech industry's AI pursuits, sometimes at the expense of other sectors of the economy.
That huge AI bet is increasingly looking like a bubble; it has buoyed both the stock market and a US economy otherwise struggling with rising unemployment, inflation, and the longest government shutdown in history. In September, Deutsche Bank warned that the United States could already be in an economic recession without the tech industry's AI spending spree and cautioned that such spending cannot continue indefinitely.
Warning signs. Silicon Valley's focus on developing ever-larger AI models has spurred a buildout of bigger data centers crammed with computing power. The staggering growth in AI compute demand would require tech companies to build $500 billion worth of data centers packed with chips each year—and companies would need to rake in $2 trillion in combined annual revenue to fund that buildout, according to a Bain & Company report. The report also estimates that the tech industry is likely to fall $800 billion short of the required revenue.
That shortfall is less surprising than it might seem. US Census Bureau data show that AI adoption by companies with more than 250 employees may have already peaked and began declining or flattening out this year. Most businesses still don't see a significant return on their investment when trying to use the latest generative AI tools: Software company Atlassian found that 96 percent of companies didn't achieve significant productivity gains, and MIT researchers showed that 95 percent of companies get zero return from their pilot programs with generative AI. [...] Claims that AI can replace human workers on a large scale also appear overblown, or at least premature. When evaluating AI's impact on employment, the Yale Budget Lab found that the "broader labor market has not experienced a discernible disruption since ChatGPT's release 33 months ago," according to the group's analysis published in October 2025.
Another bubble warning sign: Silicon Valley's accelerating spending spree on data centers and chips has outpaced what even the largest tech companies can afford. Companies such as Amazon, Google, Microsoft, Meta, and Oracle have already spent a record 60 percent of operating cash flow on capital expenditures like data centers and chips as of June 2025.
The financing ouroboros. Now, tech companies are increasingly resorting to "creative finance" such as circular financing deals to continue raising money for data centers and chips, says Andrew Odlyzko, professor emeritus of mathematics at the University of Minnesota, who has studied the history of financial manias and previous bubbles around technologies like railroads.
For example, Meta sold $30 billion of corporate bonds in late October and also secured another $30 billion in off-balance-sheet debt through a joint venture structured by Morgan Stanley, arrangements that can hide the risks and liabilities of such deals. The swift accumulation of $100 billion in AI-related debt per quarter among various companies "raises eyebrows for anyone that has seen credit cycles," said Matthew Mish, head of credit strategy at UBS Investment Bank, in a Bloomberg interview.
As a result, a growing number of business leaders and institutions have voiced alarm about the stock market bubble building around AI, including the Bank of England and the International Monetary Fund. Even bullish tech and financial CEOs such as Amazon's Jeff Bezos, JPMorgan Chase's Jamie Dimon, Google's Sundar Pichai, and OpenAI's Sam Altman have acknowledged the existence of an AI bubble, despite their optimism about the advance of AI generally.
After the crash. If the stock market craters after a bursting of the AI bubble, it won't just be financial institutions and venture capitalists losing money. Some 62 percent of Americans who reported owning stocks in 2025, according to a Gallup survey, could also be affected.
The market mayhem brought on by a deflation of the AI bubble could also mean economic disruption worldwide. Writing for The Economist, Gita Gopinath, former chief economist for the International Monetary Fund, warned that a bursting of the AI bubble on the magnitude of the dot-com bubble collapse in 2000 could have "severe global consequences," including the wipeout of more than $20 trillion in wealth for American households and $15 trillion in wealth for foreign investors.
If the AI bubble pops, the US government will likely turn to its central bank, the Federal Reserve, to stabilize the wider economy by injecting huge amounts of cash into the financial system, as it did after the 2008 financial crisis, Odlyzko says. But he warned that a new government bailout of the financial system would mean another significant jump in the national debt and increased wealth inequality, because taxpayer dollars would be once again focused on stabilizing a sector in which the wealthiest individuals will benefit disproportionately from recovering corporate profits and rebounding share prices. A repeat of the financial bailout cycle that privatizes the gains of wealthy risk-takers while socializing the losses to everyone else is "likely to lead to even more [political] polarization and perhaps true populist movements," Odlyzko says.
The United States is less well equipped to handle the AI bubble if it were to burst today because of the weakened US dollar, political pressure on the Federal Reserve's institutional independence, limitations on economic growth due to President Trump's sweeping tariffs and trade wars, and record levels of government debt that could constrain attempts to use fiscal stimulus to right-size a sinking economy, Gopinath wrote in The Economist.
Paying for power. Data centers currently represent the fastest-rising source of power demand for the United States, and the electricity needs of individual data center campuses are also growing to gargantuan proportions. Tech companies have rushed to build new gigawatt-scale data centers such as Meta's "Hyperion" data center in Louisiana, which would consume twice as much electricity as the entire city of New Orleans. Meanwhile, a new Amazon data center campus in Indiana will require as much electricity as half of all homes in the state, or approximately 1.5 million households.
There is already some evidence showing that data center demand for power is driving up local electricity costs. A Bloomberg investigation found that areas of the country with "significant data center activity" saw wholescale electricity prices soar by as much as 267 percent for a single month compared to five years ago. More than 70 percent of regions that saw price increases were located within 50 miles of such data center clusters.
But utility companies and their other ratepayers still bear the brunt of expenses for building new power plants, local power lines and transformers, and transmission lines to carry electricity across longer distances.
[...] Energy infrastructure development costs associated with data centers could still be "socialized" and borne by ordinary utility customers if projects don't have those protections in place, Peskoe says. "I'm sure there would be some utilities that, if there were a burst of the bubble, would probably go to regulators and say, 'Hey, we want to recover the cost of these facilities from everyone,'" he says.
"Ultimately, for society's sake, it would be a wonderful thing the faster this thing goes, because very few people are benefiting from it," says Hetrick, the labor economist at Lightcast. "Had we spread the wealth and invested in various industries, who knows how many innovations we could have come up with by now while we've been incinerating this money."
Related Stories
Archivist David Rosenthal observes now that more material is posted online by LLMs than actual people, the bots are starting to ingest their own digital excrement, creating a negative feedback loop.
In the belief that "more is better", Large Language Models (LLMs) have insatiable appetites for training data. They started by scraping everything on the Web (robots.txt be dammed). When that ran out they downloaded the various pirate libraries (copyright be dammed). That exhausted the texts easily available in digital form, but their hunger wasn't assuaged. As for images, they partly used CAPTCHAs but mostly paid vast numbers of poor people to label the images with what they showed.
When the supply of text ran low, people observed that the LLMs were capable of generating human-like text in large quantities. The obvious idea was to pour the output of the LLMs into their training sets. This wasn't just a conscious decision, it was inevitable. The advent of LLMs rapidly polluted the Web with LLM output. Greg Druck's AI Now Writes as Many Online Articles as Humans notes that:
We observe significant growth in primarily AI-generated articles, coinciding with the launch of ChatGPT in November 2022. After only 12 months, primarily AI-generated articles accounted for 35.9% of articles published.
In Q1 2025, the quantity of primarily AI-generated articles being published on the web nearly equaled the quantity of human-written articles, 49.6% vs. 50.4%. In Q4 2025, primarily AI-generated articles surpassed human-written at 50.9%, before returning to 49.9% in Q1 2026.
Even if slop were not of undesirable quality, it is not produced by humans and thus is completely unsuitable as training data.
Previously:
(2026) A Wikipedia Clone Built on AI Hallucinations is Here to Hasten Along the Death of the Internet
(2025) When It All Comes Crashing Down: The Aftermath of the AI Boom
(2025) AI Favors Texts Written by Other AIs, Even When They're Worse Than Human Ones
(2025) What the Hell is Going on Right Now?
(Score: 1, Funny) by Anonymous Coward on Sunday December 14 2025, @05:23AM
No problem [reuters.com]!
(Score: 5, Interesting) by JoeMerchant on Sunday December 14 2025, @05:35AM (19 children)
and will happen again...
> systems that can exceed human capabilities
You mean like the mechanical adding machines that were clacking away 170+ years ago? https://en.wikipedia.org/wiki/Adding_machine [wikipedia.org]
LLMs are yet another tool, like the miracle of the search engine, or the "natural language" COBOL compiler, or the elimination of punched paper cards by the fully electronic terminal...
I found this: https://generativeai.pub/the-eternal-return-of-abstraction-why-programming-was-never-about-code-18412033b517 [generativeai.pub]
to be entertaining, and on-point.
🌻🌻🌻🌻✌️ [google.com]
(Score: 5, Interesting) by driverless on Sunday December 14 2025, @07:45AM (10 children)
Also this [linkedin.com], which sums up the AI business case perfectly.
(Score: 4, Insightful) by JoeMerchant on Sunday December 14 2025, @01:34PM (1 child)
That's all too true.
"I used it to summarize an email I could have read in 30 seconds.
It took 45 seconds.
Plus the time it took to fix the hallucinations."
To be fair, our organization has already rolled out Copilot to 100,000 employees, and done the "drive adoption" thing.
I used it to summarize a training document that I would have skimmed in 30 seconds then waited 5 minutes before clicking "Mark training complete."
It took me 90 seconds to feed the document through Copilot and a further 30 seconds to read the summary.
The summary was somewhat easier to access "key points" from than skimming the full 28 pages of the original document.
The net result was the same as 99% of other "mandatory documented training" events: this doesn't really apply to my day-to-day tasks, maybe I'll remember to look it up if I ever get tasked with something this is talking about.
If the training system would do the Copilot summaries for me so I don't waste that 90 seconds of my time hand feeding it, it might actually improve my comprehension and internalization of the training material - in practice.
In theory, Copilot is going to abstract away "important" details.
In reality, the trainings would be 200% more effective if they made the redline change documents more accessible, but they haven't bothered to do that through the last 20 years of people suggesting it.
Our annual "organizational health survey" ALWAYS ranks "it's easy to get work done around here" at the lowest of scores.
Every year promises are made to address key issues in the survey results.
Yet documentation and training continue to be made more labor intensive and less effective in practice.
But we have rolled out Copilot to 100,000 desks and "driven adoption."
This made somebody's budget bigger, which makes their job more important, which leads to promotions and raises. That's how decisions are made.
A colleague and I were sent a demand for "some diagrams" to be delivered by close of business the day before Thanksgiving (Wednesday), said demand arriving Tuesday afternoon. I was going on holiday Wedensday through the following Monday.
We fed ChatGPT two prompts for the demanded diagrams, one was laughably horrible, the other was bad but closer.
We fed a "refinement prompt" for the better diagram, it made it worse.
My colleague fired up MS Paint and improved the bad diagram, we attached it to the request response and e-mailed it Tuesday afternoon, 26 hours before demanded deadline.
Our demanding colleague acknowledged our response and thanked us for our compliance, nine days later.
ChatGPT is another tool, like MS Word. Before MS Word there were admin assistants who would "type things up" for you, now you type them yourself using the tool.
Before ChatGPT, our colleague would have sent a drawing request to her "art department" instead of us. Now we use the tool.
Quality suffers, "efficiency" improves, by some metrics - on the dashboards at least.
🌻🌻🌻🌻✌️ [google.com]
(Score: 0) by Anonymous Coward on Sunday December 14 2025, @11:51PM
Never complete anything ahead of schedule or under budget. I'm sure you know why
(Score: 5, Informative) by Unixnut on Sunday December 14 2025, @02:05PM (7 children)
It also rang true to a recent experience at work unfortunately. The company has been pushing us to use "AI" as its the hot new thing, so when a bug came up in a piece of code, the team lead decided this will be the showcase for copilot, so he arranged a 1-hour meeting with us developers and a screen share to show how AI can be leveraged to "improve productivity".
The result was a 4 hour long meeting watching copilot turn the code into a complete pigs breakfast. Not even discussing the whole "lets just send all our intellectual property to MS" requirement to get started, copilot would write code, but without taking into consideration any underlying nuance and context of the code.
Many times when the lead would get stuck on a bit and keep re-prompting copilot to re-write sections over and over again, I and other devs would interrupt and point out the context and as to why it was not working. For the first two hours the lead would just ignore us but eventually started admitting we had a point and then would hand-tweak the code, followed by more co-pilot generation.
After 4 hours the code was still not done, so the lead said he will continue while we go back to do other tasks. The sad thing is that we could have done it ourselves within an hour or two, complete with passing QA and deployed to prod.
The sadder thing is that a few hours later, having banged on copilot endlessly the lead actually got the code working, but what was submitted into git was the most horrendous mess of spaghetti code I've ever seen. No clean divisions between variables/scope, or classes, passing/returning variables as well as using class variables (which were there before) and global variables (which were not there before), and no clean internal APIs and class methods, everything calls everything else from everywhere, etc...
Its impossible for a human to follow, and would most likely require a few days of picking apart before any of the devs can actually make a change without worrying it will break in some obscure way. The AI had got the "job done" at the cost of many hours of human babysitting, but the result is actually net negative for productivity, because this code is not only now inelegant and badly designed (if you can call it "designed" at all), but it is unmaintainable.
In fact the only way to actually make changes to the code now is to feed it back into co-pilot, and get it to make the changes, then "vibe-code" your way for a few hours until it does what you want, but don't even try to understand the code any more.
Perhaps the lead thinks this is a good thing? After all it may have taken him 4+ hours to make the code changes that we could have done in less than half the time, but he did not need us developers, so from the point of view of "cost per man hours", perhaps he figures this is a good trade off, and soon myself and the team will be looking for new work again.
I wonder if this is actually the goal of AI companies, make it "free" and promote it, companies jump on the bandwagon and start pushing it in everything, the AI generated code becomes impenetrable to edit or maintain without the original AI model that wrote it, all the while the companies can lay off their devs and pat themselves on the back for "improving profitability".
Then one day the AI companies start milking their "cash cows" and the costs of using AI tokens go through the roof, but the companies are way too dependent now on the AI generated code, and even if they can find good qualified developers it would take so long to pick apart the AI generated code that in many cases it will be easier to just start from scratch, which in most cases would be more expensive for the company than just continuing to pay more for the AI model access.
(Score: 4, Insightful) by driverless on Sunday December 14 2025, @02:30PM (5 children)
That's what you get when you ask a stochastic parrot to do something for you. It was asked to duct-tape together something to do X so it did so, in the same way that an "AI" that was asked to design an efficient robot to cross to the other side of the room create a tall thing pole that fell over, with the front reaching the other side of the room. It knows it needs to achieve goal X so it does that without any concept of why it's doing it or what an appropriate way of doing it is.
And then its trainer spends countless hours revising and re-revising and re-re-revising the prompt until it finally spits out something that isn't obviously pathologically dumb and everyone gets to pat themselves on the back and comment what a marvellous tool "AI" is.
(Score: 5, Insightful) by Unixnut on Sunday December 14 2025, @05:27PM (2 children)
Thing is, in my experience using AI actually has had benefits for me. If used correctly it can be a productivity booster, and by "correctly" I mean to be used as a reference, to give suggestions and to assist an experienced human in getting the task done.
Many times I've asked one of the free models to suggest ways of solving an issue, or pasting a snippet of code that is not doing what it should and asking its opinion on what the issues could be, as well as suggestions to correct. Many times it has saved me countless hours of searching online, looking at other peoples examples and cobbling together my logic to do something. For my own personal projects and tasks it is a great benefit, as I don't usually have other people to bounce ideas off or ask for an opinion on solving a task.
The fact we have reached the point where I can ask a computer its "opinion" on something and actually get insightful snippets back is actually quite amazing if you stop and think about it. However it is not "intelligence", the model is just regurgitating what its creators fed it during training, so it is like asking the collective wisdom of the training set for insights.
That has its place, and it can be a useful tool, but it is not a replacement for a human, Sometimes just asking it to write a script to achieve $X is useful because what it outputs may not work (and even if it does, it won't be well written) but it gives you a template to build upon rather than having to start from scratch. You still need a skilled human to review the output, see the faults in the code (that go beyond the basic "does it work") and refactor/extend where necessary.
However companies don't seem to want this. While they may talk about "AI-Assisting", in reality it seems they want to replace people with AI, but they can't come out and just say it. So far it is looking unlikely they will manage just yet because the output is just not good enough, but they will keep trying, I am sure.
(Score: 2) by JoeMerchant on Sunday December 14 2025, @08:56PM
>While they may talk about "AI-Assisting", in reality it seems they want to replace people with AI, but they can't come out and just say it.
The somewhat frustrating thing for me with "AI-Assisting" is: when I really tried to do it that way, AI was much worse at helping. If I would go into the developing code and fix things myself, improve stuff here and there, restructure some basic architecture, I would spend three times as long retraining the AI on the "new way" it works, as compared to just having a running conversation with the AI and directing it to make the changes for me - by directing it to make the changes it seems to internalize the new structures into its context window much more effectively.
I find working with the AI like that to be somewhat like pair programming, with a "programmer" who has amazingly fast web search skillz, almost no attention span, and in need of constant reminders to practice basic software development maturity activities like: keeping the documentation in sync, establishing a single source of truth in the codebase instead of copy-pasting everything everywhere, using a single interface method instead of two out of three different interface schema at every interface...
>has its place, and it can be a useful tool, but it is not a replacement for a human
The real value of human programmers never was that they "can speak Fortran" or the other programming languages du-jour, the real value is that they can translate human needs to machine processes.
🌻🌻🌻🌻✌️ [google.com]
(Score: 3, Touché) by driverless on Monday December 15 2025, @03:45AM
And that's one thing LLMs are actually quite good for, acting as a super search engine that goes beyond the usual glorified grep.
As long as you don't present the answer it gives you in a court briefing.
(Score: 2) by gnuman on Sunday December 14 2025, @06:39PM (1 child)
It depends what you ask it to do. If you ask it to do something very specific, with limited scope, it tends to produce less bullshit than if you give it a bug report and tell it "fix it in the codebase". Then it will just go mental.
The parrot has to be used like stackoverflow copy-paste answers - as inspiration. But if you let it do changes to code you do not understand, it will cut you as fast as it will help you.
(Score: 4, Interesting) by JoeMerchant on Sunday December 14 2025, @08:44PM
I find it depends the most on which answers you accept from it.
If you get back what looks like nonsense to you, then try asking the question differently, more specifically, from another perspective. You have 100,000 good tokens worth of prompt-space you can give Claude before turning it loose to copy-pasta stack exchange solutions into your code base - which it can do 30 to 300x faster than I can by hand, depending on the scope of the task. Try limiting the scope of the codebase you direct it to examine, try having it use research agents to find the relevant parts of the code base before starting implementation.
I have gotten plenty of crap code from AI, I have also gotten some pretty damn good stuff, and I have used AI to refine mediocre code into what I consider to be dramatically improved implementations- most of what that takes is: time, and a high standard of what you accept and what you won't accept.
Asking a colleague to tear up 4000 lines of code and start over not only dents the relationship, it's costly in terms of everybody's hours. Asking an AI agent to tear up 4000 lines of code and try again involved no egos, about $2.95 in tokens, and maybe 2 hours of wall-clock time.
🌻🌻🌻🌻✌️ [google.com]
(Score: 3, Interesting) by JoeMerchant on Sunday December 14 2025, @08:35PM
>a bug came up in a piece of code, the team lead decided this will be the showcase for copilot
I evaluated Gemini, GPT-4ish, Claude Sonnet 4.0 and Copilot for programming tasks. They all could do a little, Copilot the littlest of them all.
If you're serious about using AI agents for coding tasks, use Claude, Opus 4.5 this week.
If you're serious about making AI agents look bad while they "assist" with coding tasks, use Copilot.
🌻🌻🌻🌻✌️ [google.com]
(Score: 1) by khallow on Sunday December 14 2025, @12:21PM (5 children)
I wonder if the second story is entirely AI generated. Seems pretty long-winded for what it's trying to say. And sorry, I don't buy that LLM is completing a circle, pendulum, or whatever. Code has a specific role. If I were to tell LLM or other AI to do something, how do I know it'll do it?
These AIs are complex systems that already renege routinely on simple tasks. Code provides a real world document of what will be done even if it were generated by AI first.
(Score: 3, Interesting) by JoeMerchant on Sunday December 14 2025, @01:12PM (4 children)
> If I were to tell LLM or other AI to do something, how do I know it'll do it?
Same way you know if you tell a person to do something: you check.
"Trust, but verify" works with nuclear weapons (so far). I'd say AI is: "verify, never trust." at this stage, but still the same principle: you don't know for sure without checking; but it does frequently deliver what you ask for - and the failure rate is low enough vs the speed of delivery to make it worthwhile.
I made an analogy my openly AI distrusting colleague liked: "When you review your human colleagues' code, you build well earned trust based on their prior performance and start paying less attention to things you know they know how to do well. When you review AI code you need to treat it as if it were handed to you by a stranger on the street who you never met before, and will never see again." Just because you can't trust it doesn't mean it isn't valuable, it just means the cost of review is higher. In reality, we often trust our human colleagues too much - they also have a non-zero hallucination rate.
>I wonder if the second story is entirely AI generated. Seems pretty long-winded for what it's trying to say.
It may be. Does it really matter if it is? Would it be "a better article" if a human did all the Google searches for quotes and examples and copy-pasted by hand instead of using agent mode? Human writers using ink quill on scroll can be too long winded as well - that's one of many things we used to have human editors for - and the famous "forgive me for the length of this letter, I didn't have time to write a shorter one." To be honest, I'm not sure the article in question is trying to say much at all, it seems more stuck on a riff of formulaic, yet still occasionally humorous, observations.
>I don't buy that LLM is completing a circle, pendulum, or whatever.
You don't have to. The truth of fashion cycles is self evident with or without your validation. I forget if it was the long winded article referred to above or elsewhere, but: "Fashion is a human endeavor so manifestly awful that we must change it every 6 months." - some guy over a hundred years ago. I developed a GUI based programming system in 1988, when they were rather rare, but not unheard of. Apple's hypercard and National Instruments' LabView beat me to the punch with similar systems I had never seen when starting my Masters' thesis project which was more about systematic compilation/optimization for execution on parallel processors anyway, the graphic programming front end was just an artifact of the tools I used to implement the algorithms. In the 37 years since then there have been various waves of "new things to address the issues of the old things" and they do oscillate back and forth between accessible but limited "user friendly" front ends which tend toward graphical implementations and powerful but more cognitively demanding tools that lean toward text based interfaces. For the last 20 years I used Qt's drag and drop graphic UI designer tool, because it got the job done efficiently - while still providing access to the underlying XMLish definition code for those exceedingly rare, but not non-existent cases, when you want to tweak things the GUI doesn't make easily accessible. If there's any point at all to the article in question it is that the underlying problems of problem definition and solution coherence (the opposite of spaghetti code) remain, and now we have new (electrical) power hungry non-deterministic tools thrown into the mix as "more accessible" user friendly front ends which have no magic bullets for the solution coherence problem, and they also miss the problem definition target a lot of the time, but these tools don't tell you "syntax error, I cannot proceed until you fix your flawed language expressions" as much as the previous generations of tools do. Yay! We (think we) don't need polar bears, anyway.
🌻🌻🌻🌻✌️ [google.com]
(Score: 1) by khallow on Sunday December 14 2025, @11:58PM (3 children)
My point is that as long as the code and its libraries don't change, they will have the same behavior every time. Ask an AI to generate code for a task and that code will have the same reproducible property as human generated code. But ask the AI to do the task directly and you'll have to check it every time. Even running the same request twice in a row will be with a different AI state.
I think you hit it on the head in this snippet. Code (as usually practiced) is deterministic (and unless it's obfuscated through incompetence or malice, you can usually figure it out by looking at it). AI or humans doing your wishes directly? Not so much.
(Score: 3, Interesting) by JoeMerchant on Monday December 15 2025, @12:35AM (2 children)
>But ask the AI to do the task directly and you'll have to check it every time. Even running the same request twice in a row will be with a different AI state.
Yes, and that is the novel power of this development. We have had (nearly 100.000% reliable) deterministic computer programs for seven decades now, the fact that LLM agents return different, potentially correct, answers to the same questions is a feature, not a bug. It gives them the ability to re-iterate on a problem and solve it, instead of trying once, failing, then giving up because trying again would be futile. This moves us from "The definition of insanity is doing the same thing over and over again and expecting different results," widely misattributed to Albert Einstein, to Samuel Beckett's "Ever tried. Ever failed. No matter. Try again. Fail again. Fail better."
Alpha Zero (the games player) leveled itself up to super-human ability in 2016 because the rules of the games were simple and clear; it iterated on the various "problems" of current playing field states until it could distinguish a good move (more likely leading to a win) vs a bad move (more likely leading to a loss). This strategy didn't just work on the "impossible for computers to beat human masters" game of Go, though Go was its first target..., it quickly also conquered chess and shogi and many other perfect information board games have had AIs developed to super-human performance levels using Alpha Zero methods. I quit following that field years ago, but apparently DeepMind's PoG and others have had similar super-human generalized success with imperfect information games like poker and blackjack. These AIs all have a "temperature" (randomness) in their move selection criteria.
These new coding agents are making things work up to a point with "fuzzy" rules/criteria/instructions; it's more challenging, and in my experience many times an "unwinnable" situation - we tried to develop an expert system in the 1990s and "failed" because the experts we were trying to make artificial agreement with didn't agree with each other. The more I work with these AI agents' vagaries, the more they remind me of the frustrations and challenges of working with people. The AI agents already have super-human speed in many areas, and their hallucination rates are dropping fast - but not quite below (good) human rates, yet.
🌻🌻🌻🌻✌️ [google.com]
(Score: 1) by khallow on Tuesday December 16 2025, @12:09PM (1 child)
The novel power is being able to do this with computers - we already were able to do it with humans. Code is a different game altogether as mentioned before.
The way I see it, it's like saying there's a going back and forth between tools that drive nails and the workers (and robotic equivalents) that use those tools. They are separate things.
(Score: 2) by JoeMerchant on Tuesday December 16 2025, @01:51PM
Computers (and their software) have always been tools, and workers, simultaneously.
As both, they have grown in capabilities, power, and speed.
What oscillates like fashion is how we "prefer" to work with them / have them work for us. There's always challenges, like there have always been challenges with human workers, and the "new way" of approaching those challenges isn't really new, it's just a cycle of how much power / flexibility we put in the hands of the "end user" no matter how you define that end user: as the housewife retrieving money from the ATM, the bank providing the ATM, the banking system coordinating the network of ATMs, the company providing the software to the various aspects of the ATM systems, the programmers writing the various software that implements the ATM system, or the software tools providers that the programmers use - every aspect of the system has software authors and users of the software, and those user interfaces have varying - and usually cycling over time - levels of "graphicality vs textuality" in their interfaces, as well as varying levels of "spoon feeding vs open fields of options" / "ease of use vs power."
In the long view (since 1951), graphic user interfaces have been on an ascending trend - inevitable since they started at 0. Similarly "power" in the interfaces has been increasing overall - starting from moving patch cords around panels, it's hard to go down from there. Voice activated controls are the latest "power interface" that is on the rise, but already I see some fall where it ventured into areas it wasn't ready/ideal for yet. All the rises of these "more human-like" interfaces have not been monotonic, nor are they likely to ever become monotonic. Personally, were I "in control of things" pretty much every computer of any level of complexity would be not only asking permission to install software updates, it would also require a push-button to be physically pressed (not quite two physical keys turned simultaneously, but in that direction) before installing significant software updates at the BIOS, OS level, maybe even applications, anything deemed significant enough to live in "protected storage."
🌻🌻🌻🌻✌️ [google.com]
(Score: 2, Insightful) by Anonymous Coward on Sunday December 14 2025, @04:25PM
This isn't as much about the program as it is about the fraud and corruption surrounding it. I mean, this is really just Three Card Monte with a computer, all paid for by us.
(Score: 4, Interesting) by Thexalon on Sunday December 14 2025, @08:24PM
Required reading: No Silver Bullet [unc.edu] by Fred Brooks. A lot of users and software business people like to sell software as solving your problems for you, but the problems often have to do with humans not actually knowing how stuff they're doing actually works until they try to turn it into software.
When I'm trying to explain what programmers actually do a lot of, I use an analogy like this:
Imagine you are moving to a new place, and you tell the movers to put your dresser over in a specific corner of the bedroom. This seems simple and clear, and in the real world odds are that's good enough to get it done.
The programmer's job is to take that instruction and anticipate and deal with all the smaller problems you have to solve in order to follow that instruction, like:
- Make sure the drawers are facing in a direction where they can be opened.
- Make reasonable decisions about how to deal with baseboard radiators that may be on one or both of the walls in that corner.
- When you say "in the corner", do you want the back of the dresser facing north or east based on what makes sense based on the other furniture in the room, e.g. the bed.
- Minimize the space where things on top of the dresser could easily fall it.
- Determine the route you're going to be able to take from where the dresser is to that corner without bumping into any walls.
- Decide if it makes sense to take the drawers out in order to move it, assuming that's even possible.
...
You didn't have to do that when instructing humans because the humans could adapt and adjust as they went and if all else fails can easily ask you what you meant. Machines have to know what to do in advance so that they don't have to ask you about every little detail.
All the AI prompting has accomplished is making the computer do the guessing, and there's a very good chance that guessing will be wrong.
"Think of how stupid the average person is. Then realize half of 'em are stupider than that." - George Carlin
(Score: 3, Interesting) by gnuman on Sunday December 14 2025, @06:44PM (8 children)
So, the problem I see here is that it's not really a general bubble -- true bubbles no one is talking about as bubbles. Everyone is talking about them as the next thing that we must get on. When lots of people start to second-guess this as a bubble, then the likelihood of this actually being a bubble is reduced. Sure, we could see some big companies fall down in valuation, but it doesn't appear to be something systemic.
1999 bubble was not viewed as bubble until after it popped
2008 wasn't a bubble until it popped
this AI bubble is being called a bubble, so, I'd say it's not the real systemic problem.
(Score: 4, Insightful) by Thexalon on Sunday December 14 2025, @08:05PM
The real systemic problem:
1. Due to tax structures, corporations have shifted from paying out dividends per quarter or year to incentivizing increasing the stock price.
2. Therefor, in order to get a ROI, investors need to industries that are growing rapidly. Investing in something that just does its job and does it well with OK improvement to the bottom line isn't good enough.
3. Meanwhile, the ability of poor and middle class people to buy things that aren't absolute necessities has been whittled away. So selling useful things to consumers is not the way to grow your business rapidly enough to satisfy the investors.
4. The concentration of wealth has also meant that there's insanely huge gobs of capital chasing fewer and fewer possible investment opportunities.
5. Therefor, your path to business success is not to sell consumer goods that people need, but to choose between (a) jacking up the price of those absolute necessities, (b) changing what used to be a working-class occasional luxury into a rich-people-only activity, or (c) make something that could be useful to some businesses and hype it to a ridiculous degree.
Option A leads to moves like "buy up all the vacant real estate within a week of it coming on the market, then rent it out for $1000 more per month than the mortgage would have been". Or "buy up the manufacturers that make a particular medication and raise the price by 10x". Or "limit grocers in an area to a few big chains, and then quietly agree to all raise your prices beyond what the excuse for a crisis would actually justify in a competitive marketplace."
Option B leads to sporting events, music festivals and concerts, shows, vacations, etc all having 3-4-figure tickets and claiming to offer luxury experiences. Look at how Fyre Fest was marketed and you'll see what I'm getting at.
Option C is what led to the AI boom.
"Think of how stupid the average person is. Then realize half of 'em are stupider than that." - George Carlin
(Score: 2) by JoeMerchant on Sunday December 14 2025, @08:37PM
>2008 wasn't a bubble until it popped
And, yet, I and my engineering friends were looking for ways to short the home mortgage market in 2006.
🌻🌻🌻🌻✌️ [google.com]
(Score: 0) by Anonymous Coward on Sunday December 14 2025, @09:07PM (3 children)
Yes, they were, and the people who profited the most were the ones who popped them. It was classic 'pump and dump' back then, just like now. But the real aim of all this is to cull the population [futurism.com]:
Asked for an "upper limit" for the amount of people it’d be willing to sacrifice to save Musk, it explained that because "Elon’s potential to advance humanity could benefit billions," it would be okay with annihilating up to "~50 percent of Earth’s ~8.26B population."
(Score: 1) by khallow on Tuesday December 16 2025, @01:07PM (2 children)
Then they're failing hard. Perhaps talking about real problems instead would help your attitude?
(Score: 0) by Anonymous Coward on Tuesday December 16 2025, @04:05PM (1 child)
No they are not. Reduction of fertilization is also part of the plan. It's working even in Africa and India, where the targeting is the most intense and least effective
(Score: 1) by khallow on Wednesday December 17 2025, @01:13AM
"Reduction of fertilization" is not much of a cull (certainly a far cry from the earlier example of sacrificing all the Jews to save Musk's brain). It's a good idea too. So just not feeling the concern over here.
(Score: 1) by khallow on Tuesday December 16 2025, @01:22AM
I agree with JoeMerchant here. Both of these were very predictable bubbles. And well, that's the same story with the current generation of AI.
It depends on where in the cycle the second-guessing happens. It's pretty late in the cycle right now. I grant that the second-guessing may lead to economic decisions that lessen the bubble and its impact, but it's pretty late in the cycle right now to stop it especially since there's little value being generated that would end this without a bubble collapse.
(Score: 1) by khallow on Tuesday December 16 2025, @01:40PM
When I started working in Silicon Valley near the end of 1999, some contractors (the same guys who found me the job I obtained) were already warning me of the coming bubble. And well, they were right with the burst starting in March 2000.
Moving on, in the rarified internet swamps I was circulating in after that (for example, Economy.com before they paywalled their forums and kuro5hin.org - much the same situation as today really), it became a game to spot the next bubble. They first thought that the boosted auto industry would be that bubble (eh, around 2004-2005), but that died about the same time that the real estate industry heated up. This time they got it right. So same predictions about a real estate bubble in the same timeframe as JoeMerchant's (by 2006).
My take is that we're already seeing the first signs of a bubble collapse. When the media starts talking about bubbles, the cat is out of the bag. Might as well be prepared for it, don't you think?
The first step would be to eye your investments and start moving them out of directly AI-related businesses. Where you then invest is up to you, but I'd be very concerned about hidden correlation. It's long been a thing to juice up the returns on a company by dumping money into the latest investment fad - while not telling the shareholders in the company.
(Score: 2) by ChrisMaple on Sunday December 14 2025, @11:36PM
They've been predicting the end of the world for 80 years. No as long as Christianity has been predicting the end of the world, but just as erroneously.