> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential. Everything else is transient
As long as the term “AI” means by-and-large LLMs with additional features sprinkled on top, the answer is no. More likely (without careful vetting by the folks aggregating these models) is that the quality will go down as more and more AI-generated output gets subsumed into these models.
Even without that particular problem, LLMs-as-AI can only give us probabilistic outputs based on inputs; and by definition they’re reliant on humans to provide the training data for their model. Without specialized knowledge or training on that knowledge (And even with it, viz. Meta’s engineering), we don’t have to worry about AI itself. We do have to worry what investors who are looking for outsized returns will do to get those returns, job market be damned.
The problem for us isn’t that AI will take our jobs; it’s that snake-oil salesmen can sell the idea that AI will take our jobs, investors buy into it, companies try it, fire their folks, the snake-oil salesmen IPOs, the companies that bought into this idea implode in some form or fashion, and the salesmen have already taken the money and ran. Of course, we still lose our jobs, but maybe (!) we get them back when this all fails?
It’s no longer true that AI tools primarily get knowledge from their pre-training input data. That gives them a baseline, but nowadays AI chatbots and coding agents routinely assume they need to get up-to-date information in other ways, via web searches and other tool calling.
So I don’t see accuracy declining at least for programming.
> nowadays AI chatbots and coding agents routinely assume they need to get up-to-date information in other ways, via web searches and other tool calling.
So I don’t see accuracy declining at least for programming.
How do those chat bots discern that the ‘web searches’ they’re using are returning human generated information only that’s been vetted instead of LLM output?
I know within a second whether the search result I am looking at is obvious AI slop (search for almost any health condition, recipe, esoteric questions etc almost always have slop at the top of the result), but LLMs regularly source those sites as the basis for conclusions.
Sure and thats fair, there's probably many things that are good enough that I miss, but I'm talking about the most obvious possible websites being used regularly as sources in web search. Identifying poor sourcing or misinformation on the internet and what is a credible source has been a lifelong skill I've had to build from the early days of the web, and in the AI boom its only gotten more necessary to be able to not get taken by hucksters with billion dollar budgets.
When you do your own web searches, you learn to trust certain signals over others. You learn which sources are trustworthy and which are suspicious. LLMs don’t do that, and they present every information to you as being equally valid.
Upvoted you, of course; but it’s worse than that. It’s vibes being marketed as correctness. To the lay person (and unfortunately, to more than a few folks who should know better), computers don’t “make up” information. Maybe some good (in some weird way) that comes from all of this is that we stop using LLMs for recitation of facts.
If someone has a vibe coded website with bad statistics but shows up on a web search, what tools will it use to check those statistics? How will it know what data it needs to validate? What tools will it use?
Humans face the same problem. So this at least shouldn't make AI perform worse relative to humans, even if AI slop degrades the performance of both over time.
I find it to not be acceptable if AI's trend is to degrade performance of both AI and humans at the same time, that's kinda not the goal, right? Why are we spending money to make us dumber?
AI isn't going to go away, and AI-generated content isn't going to go away. So while it's an open question the extent to which AI training will be hampered by the proliferation of AI-generated content, I think that the existence of such content is a reality that we'll have to accept, whether you like it or not.
Right, but we’re not talking about a worst case scenario where models are trained primarily on their own output. We’re talking about what will happen in the long run, as future models are trained on a realistic mix of content. It’s surely an open question what will happen.
> More likely (without careful vetting by the folks aggregating these models) is that the quality will go down as more and more AI-generated output gets subsumed into these models.
This assumes that there aren't algorithmic breakthroughs which reduce training/inference costs by several OOMs.
How much do these models need to do before people throw their hands in the air and say, ok this is happening. The Erdos unit distance problem, which as far as I understand was approached by multiple competent mathematicians was solved by a frontier model. Sure people argue there was no novelty there (I cannot comment as a non-mathematician) but it feels like they can draw lines laterally from deep knowledge in different fields (in this case combinatorics and algebraic number theory I believe) and solve problems.
Now if you have millions of instances running in parallel, all "probabilistic", working on frontier AI research I really don't see the blocker (and believe me I wish I did).
> How much do these models need to do before people throw their hands in the air and say, ok this is happening
What is "this"? Most people arguing against some of the more fervent predictions and promises of "inevitability" are people who are using these models in day to day - they see what the models can do, and what they struggle at.
> Now if you have millions of instances running in parallel, all "probabilistic", working on frontier AI research I really don't see the blocker (and believe me I wish I did).
My genuine prediction is that you'll get a lot of early results simply because you're applying attention to some low hanging fruit of problems, but then it will drop off due to the cost of tokens and the low rate of return. This doesn't mean that the models are especially capable of novel thought, just that we haven't algorithmically brute forced a problem with known solutions.
We would be seeing more success cases if the promises were true, setting aside AGI, human replacement, etc. We would see more, better products with more features that people would use. We wouldn't be having any arguments. The human replacement presupposes the models work in ways that they don't, and until proven otherwise, can't. I've watched those who embrace it fully flounder around on projects, some have lost their mind from the constant LLM validation, and I've seen companies go all in and then pull back based on both cost and efficacy over the last year.
I'm still waiting for the success case examples applied on a scale that would make any of the predictions come true.
> This assumes that there aren't algorithmic breakthroughs which reduce training/inference costs by several OOMs.
Yes, if one must assume something it is generally fair to assume that things will continue as they are. Research breakthroughs do happen, but they are not something for which you can predict the timing.
You're assuming that frontier AI research can "kick off the exponential" - that there exists such an exponential to kick off, and that frontier AI research can reach it soon, rather than in a couple centuries. Those are massive assumptions that have exactly zero empirical data behind them.
> How much do these models need to do before people throw their hands in the air and say, ok this is happening.
It's quite interesting to me that your epistemology is basically "God in a box is inevitable" and not skepticism.
> Now if you have millions of instances running in parallel, all "probabilistic", working on frontier AI research I really don't see the blocker (and believe me I wish I did).
The blocker is lots of things, but for one? Compute. Millions of instances running parallel, all "probabilistic" testing different theories would require million of GPU clusters.
Two? For subjects like Physics - Atoms. You need to be able to run experiments. You need access to the real world.
Then there's the S curve. Algorithmic breakthroughs operate on orders of magnitude. Eventually we'll run out.
>As long as the term “AI” means by-and-large LLMs with additional features sprinkled on top, the answer is no. More likely (without careful vetting by the folks aggregating these models) is that the quality will go down as more and more AI-generated output gets subsumed into these models.
Nope. This isn't how it works.
AI progress has largely been synthetic and has produced leaps and bounds capabilities increases in the last couple of years.
>LLMs-as-AI can only give us probabilistic outputs based on inputs
I am not completely sure what you are saying here, but it sounds like a variation of the "it's just a stochastic parrot" argument, which is reductionist. The human brain is also just a bunch neurons firing.
> The problem for us isn’t that AI will take our jobs; it’s that snake-oil salesmen can sell the idea that AI will take our jobs, investors buy into it, companies try it, fire their folks, the snake-oil salesmen IPOs, the companies that bought into this idea implode in some form or fashion, and the salesmen have already taken the money and ran.
Or, it eventually becomes clear to enough people that the AI companies aren't going to make enough money to justify their valuations, so the asset bubble bursts, the economy crashes, and we lose our jobs.
Many jobs imo have been under threat for many yrs.
The tremendous growth in earnings meant some fake excess head count number was viable.
Google faced an activist investor who practically forced sundar to fire a bunch of people. This is what’s coming for big tech if this AI thing blows up. Apple Is safe because they are clever and saw this from a mile away.
Investors want their cake and they will eat it too.
This is all noise. The leaders of these companies are flip-flopping to whatever sounds best for their current agenda - hiring, fundraising, pre-IPO, etc.
The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential. Everything else is transient noise.
> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential
I agree with your sentiment (about the noise), however I think this over simplifies it a bit. We may get AI that is super-human at frontier research and dramatically accelerates the pace, and still have to wait decades before it disrupts the job market (or maybe never displaces all work).
For one, the answer may depend on material science and chip manufacturing that can take a very long to build out a supply chain for even with super AI help.
And we may just find that the human mind is way more capable than we thought and even with accelerating research it's just a harder problem than anyone expected, even algorithmically.
I expect it to be a bit of both, and from ~2015 - 2025 I was in the "AI is coming for all our jobs" camp. My perspective changed last year after doing a deep dive into latest science on the human brain. (I've kept a very close eye on AI dev progress for 12+ years.
> I agree with your sentiment (about the noise), however I think this over simplifies it a bit. We may get AI that is super-human at frontier research and dramatically accelerates the pace, and still have to wait decades before it disrupts the job market (or maybe never displaces all work).
I don't see why that's the case when you have super-human researchers on tap. There are indeed physical (supply chain-y) issues to deal with but isn't the whole point that:
1. Super-human at AI research + scaling to millions of instances will probably result in super-intelligence in everything which is not AI research. (a subset of which is white-collar work)
2. Use that super-intelligence to solve any supply-chain issues you might be facing.
> And we may just find that the human mind is way more capable than we thought and even with accelerating research it's just a harder problem than anyone expected, even algorithmically.
I hope so but whenever I do, I feel like I'm coping hard and not dealing with the facts.
I'm not saying we're there yet - I'm saying the trend lines are clear.
> Unlimited intelligence doesn't mean unlimited resources or instantaneous implementation.
Of course you're right. At the end of the day you need to deal with the bedrock which is the laws of physics. I could be wrong but I struggle to believe we are close to the edge of what is possible in getting the most out of our limited resources or time.
Without atomic physics, uranium would just be another shiny rock in the ground. Sand is just what covers beaches. With enough time and intelligence we've made the shiny rock power cities and persuaded the sand to solve long-standing mathematical conjectures.
> I hope so but whenever I do, I feel like I'm coping hard and not dealing with the facts.
Maybe I'm being naive, but as someone who spent 2015-2025 fully bought into the idea AI would replace all human labor within ~20 years, my updated take that it won't has a different flavor than coping.
A bit contributor to my perspective was studying the physics of the brain, and a deeper understanding of the very hard economic constraints in the current LLM supply chain.
Re the brain, we compute information in a fundamental different way that is over 1,000,000 more energy efficient than an LLM. We are also around 1,000,000 times more sample efficient at learning.
There is a lot of details in terms of specifics between complexity in human brain architecture (variety of synapses, diversity of neuron types, etc) that have no material equivalent yet.
And on the other side, understanding how hard it has been to build out euv lithography infra, and to completely retool manufacturing for a new paradigm of computing will require so much coordination, a super AI could have all the answers but we also have to deal with getting a bunch of stubborn people to align.
Some people are just coping yes, but if you really zoom into the details on both sides of the equations the gap between where we are and what it would take to dramatically surpass human intelligence is vast.
At the same time I think people overestimate how hard some kinds of knowledge work is (in absolute terms, since it's new and we haven't evolved into it as much), and how economically impactful human augmentation with weaker narrower AI can still be. (Lots of automation and prosperity but without such extreme disruption)
I think in the very long term all bets are off, but we will have plenty of time to evolve our economic systems before we get there.
> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential.
What if the answer is flatly: no? All that other stuff starts to matter a lot then.
Predicating your business decisions on a potential breakthrough that may never come is frankly insane. Imagine if at the dawn of the car industry Ford decided that it's actually a race to build the first flying car and nothing else matters.
> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential.
I think we know the answer to that already - LLMs show no sign of improving intelligence and instead providers are going down the ‘agentic’ rabbit hole.
There are too many things missing, like a world model, understanding, and taste (in the sense of knowing what is good and what is not good).
They are taught the difference through reinforcement learning with verifiable rewards. Pretending you've solved the task or making up a story about how you solved it won't do well in that training step.
> LLMs show no sign of improving intelligence and instead providers are going down the ‘agentic’ rabbit hole.
I'm not sure where you're getting this. I don't work at Anthropic but Fable (Mythos) seems demonstrably smarter than Opus for pretty much any definition of smarter and they claim that Opus was used heavily in Mythos development (yeah I know take this with a massive pinch of salt).
Either way if the models are indeed helping development, even on the engineering, you can iterate on models faster and even if they're not contributing to core research yet you still have a baby exponential by improving the engineering.
Anthropic's claims about their own products have almost zero value as evidence. They have lied to our faces about stuff before, and will again if it drives the hype cycle which delivers them money.
Sure, but let’s not pretend that people treated the statements of these ceos as strategic messaging. People very clearly treated what Altman, Zuck, Amodei etc have been saying as predictions, and it hasn’t been until they’ve been proven wrong that people have started with the counter-narrative.
It's important not to miss the fact that AI productivity was a useful excuse for companies looking to conduct layoffs. Did some companies buy the hype? Sure - but the biggest companies would have wanted that sweet stock price layoff bump anyways and AI was a readily available justification to get it.
1. Cutting edge LLMs developing ASI/AGI.
2. AIs doing general knowledge work
The second world will be achieved far before the first world is achieved. And as the first path gets develolped, the second path becomes cheaper and cheaper to run inference on along with being democratized which reduces the margins for the cutting edge companies. It seems like a mad dash to go as far as possible until 90% of general work can be automated with more cheaply available tech
> LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential
The cost is already outrunning the benefit to a massive amount, and the predicted expotential is not here yet. I predict it'll always be around the corner, a $1T model won't get there, but it will "look promising", but we'll sadly run out of money for the $10T or $100T model..
The only thing the $1T model needs to do is find some algorithmic speedup which allows it to be trained at $100B. I'm not saying that's easy or that it will happen but I just don't see why not.
I don't understand the logic here - a methodology to increase efficiency by 90% in training doesn't exist until it does - could you explain what you mean by "I don't see why it won't exist"? Are you seeing consistent gains by some process?
My read has been that a lot of leaders were trying to drive “being early” as the catalyst for future success. At the complexity scale of big orgs you’re mostly fiddling with the incentives that the system self-aligns toward. Firing a bunch of people does create an incentive to use AI, if you think it’ll help.
The more pernicious effect I’ve been seeing is that we’re living in the golden age of LLMs, but eventually that’ll fade. Tokens are subsidized and cheap, model capabilities leap forward regularly, and there’s competition driving it all. But even now there’s stories about frontier models suddenly becoming less capable, or providers switching to usage-based billing, and new model releases feel a bit more sluggish and less dramatic. (Fable/Mythos notwithstanding.)
Eventually the models are going to settle into a rut of being just “good enough” to earn a living rather than all this hoopla. A lot of people will be re-hired. And we’ll do it all again for the next wave.
Honestly, this is just the inverse of them all getting hyped on AI replacing all the jobs. Between both of these positions, RTO, crypto, VR, it's really shown just how much they're trend chasers.
The ultimate irony here is that the biggest jobs wipeout most likely to happen now is when all these “AI exploration lab” type teams that every company quickly created are blown up.
Most, if not nearly all, of these teams have little to show ROI wise and the music on the AI bubble is slowing dramatically. They went from seemingly unlimited budgets and headcount when CEOs said “get me some of that AI” to some really uncomfortable scenes playing out know as the same CEOs realize this has cost a fortune with little to show for it.
Took that nonsense to Capitol Hill, trying to tell a bunch of politicians who knew damn well they are only there as long as they can keep their voters employed. They could have asked their own AI what happens when employment reaches 40-50%. Hint: it's never good. They were going to become another problem the government had to solve.
Also, UBI is non-starter no matter what Sam Altman believes.
Of course not. This coming crash will be where we learn that tech is "too big to fail" in the same way that the financial system is. They'll let one player fail (likely Anthropic, due to the constant fighting with the government) and bail out the rest.
I guess I'm not sure why you'd expect it to be any different than the dotcom bubble, when this didn't really happen. "Too big to fail" is a thing in finance because of contagion factors that don't really apply to tech; there's no number of AI company bankruptcies that would force your bank to suspend withdrawals.
I'm not sure what you're trying to imply with this link? As the first line of this article says, Silicon Valley Bank failed. The shareholders were wiped out. (Some people got confused because the Twitter conversation was primarily driven by rich VCs and founders who were depositors of the bank, quite reasonably arguing that they should get their money back if at all possible.)
The reason is because AI is viewed by the powers that be as inexorably linked to national security. If you haven't noticed, every lab out there has the same excuse when pressed about the disruption caused by their technology - "we do it or else China does."
It won't work because it's unnecessary. If these boneheads are right AI will drive the price of goods down. Price of labor goes down. It's deflationary. UBI is a backstop against inflation. And really just there so that rich people's money is worth something.
UBI can also be a backstop against deflation, can't it? If no one has income, it doesn't matter if bread costs $1 per loaf instead of $3-5 now - no one will be able to afford even $1 eventually, after they've sold their house at a loss, sold their car, sold their laptop, iphone, etc.
Sam Altman doesn't BELIEVE in UBI - but he knows it's the only way to SELL his product, therefore he pretends he believes in it. None of these guys want one cent going to anyone but themselves.
Remember it was reported that OpenAI didn't think that ChatGPT would be successful? OpenAI thought that ChatGPT was yet another toy before its launch. Yet once ChatGPT became an overnight success, Altman started to talk about how AI would be dangerous, how it would displace or even replace jobs. In contrast, Amodei seemed to always believe in what he said. So, can we say that Altman is a opportunistic businessman, and Amodei is a cult leader?
Individual firms can cut jobs while the overall economy sees more of that job pop up. That’s what happens when individuals become more efficient, new projects become more economical and more companies end up requiring these jobs.
I kind of wish there was a way to flip the script on the companies that gave up on humans and tried to switch to AI. Make them suffer for their idiocy in the same way that workers suffered or continue to suffer.
If that makes me a bad person, fine. If a few CEO's wind up working at 7-11 to make rent money, all the better.
It will affect jobs because the investment strategy works to degenerate the economy, but not because of the tech. Like a dog that barks at a truck and thinks he scared it away.
article misses an important point that these big tech companies are all listed on the public market, any narrative about their decisions should weigh that reality and why suddenly its being disseminated.
personally, I am collecting 3 salaries working remotely. For one of the jobs, I am tasked with hiring other devs but i dont put the effort in as i dont see a point. i just say i can't find a decent engineer and why should i when a frontier models can do most of their work? in our job postings we see thousands of applicants in a very short period of time, i just do these multi stage interviews with a rotation of candidates to basically buy time while i work on another job
i see that things are getting very desperate and i feel for those that are still struggling to find SWE jobs, AI is absolutely doing a number and the gap is going to increase not decrease.
decade+ ago AI agents were not a thing, it was only last year when it started picking up and there's less pressure to hire as frontier models improve massively
The damage is done. The truly jarring realization isn't that AI companies made these predictions, but how eagerly countless corporations were willing to sacrifice their own people in pursuit of profit – even at the cost of economic collapse.
After watching the AI roll out for a couple years now I'm much more confident that it's just a scam. There is no net positive ROI on AI. It's not good enough, and the "mass job destruction" scenario offsets any marginal gains by eliminating the market for basically all products. That doesn't mean that mediocre C-suites won't try but it only takes 1-2 quarters to feel the burn and back track.
It's a rather useful tool... and it absolutely has been overhyped repeatedly and sold as a panacea.
If you believe AI will 10x you're developers you've drunk the kool-aid, if you believe AI will have no impact on your developers then you're being stubbornly ignorant.
What we're seeing is that AI is enabling us to reduce product headcount, but increase the scope of ownership for teams. We used to require each product keep at least 3 people on it to ensure that knowledge doesn't get lost with a departure, but now we're comfortable allowing an individual engineer drive it and allow AI to accelerate onboarding for a replacement when needed. This means that the teams can now handle 3x the products and our wish lists are getting shorter over time.
I recently tried to get customer service. I think it was from the DMV for my state. The website asked me to use the AI bot first. Completely useless. It took a phone call with a human to somewhat resolve the issue.
Dunno, the pencil sound effect as they let you speak to AI is quite calming. Now we need and efficient way to penalise claimants getting impatient or even rude at AI.
Sorry, but you aren't familiar with frontier models if you think it's not "good enough". The latest models are quite a bit more capable than most people in many regards already.
No, they aren't. Not even at the one thing (programming) which they are supposed to be best at. People are always claiming "frontier models are so much better" but that is a false claim.
Which "frontier models" are you familiar with in a programming context?
I'm a staff engineer and the amount of work and even troubleshooting the latest crop of LLMs can do at my behest is jaw-dropping.
I perused your comment history, and you refer to models as 'ChatGPT', which doesn't really inspire much confidence that you know what you're talking about.
People keep telling you that you're behind and using old technology and that the most recent models are really good and a sea-change because, well - they are.
> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential. Everything else is transient
As long as the term “AI” means by-and-large LLMs with additional features sprinkled on top, the answer is no. More likely (without careful vetting by the folks aggregating these models) is that the quality will go down as more and more AI-generated output gets subsumed into these models.
Even without that particular problem, LLMs-as-AI can only give us probabilistic outputs based on inputs; and by definition they’re reliant on humans to provide the training data for their model. Without specialized knowledge or training on that knowledge (And even with it, viz. Meta’s engineering), we don’t have to worry about AI itself. We do have to worry what investors who are looking for outsized returns will do to get those returns, job market be damned.
The problem for us isn’t that AI will take our jobs; it’s that snake-oil salesmen can sell the idea that AI will take our jobs, investors buy into it, companies try it, fire their folks, the snake-oil salesmen IPOs, the companies that bought into this idea implode in some form or fashion, and the salesmen have already taken the money and ran. Of course, we still lose our jobs, but maybe (!) we get them back when this all fails?
It’s no longer true that AI tools primarily get knowledge from their pre-training input data. That gives them a baseline, but nowadays AI chatbots and coding agents routinely assume they need to get up-to-date information in other ways, via web searches and other tool calling.
So I don’t see accuracy declining at least for programming.
> nowadays AI chatbots and coding agents routinely assume they need to get up-to-date information in other ways, via web searches and other tool calling. So I don’t see accuracy declining at least for programming.
How do those chat bots discern that the ‘web searches’ they’re using are returning human generated information only that’s been vetted instead of LLM output?
It’s true that they are only as good as their input data, but the same is true if you do your own web searches.
I know within a second whether the search result I am looking at is obvious AI slop (search for almost any health condition, recipe, esoteric questions etc almost always have slop at the top of the result), but LLMs regularly source those sites as the basis for conclusions.
>I know within a second
I am sorry to be pedantic, but the correct phrasing would be "I think I know within a second", which seems like a pretty important distinction.
Sure and thats fair, there's probably many things that are good enough that I miss, but I'm talking about the most obvious possible websites being used regularly as sources in web search. Identifying poor sourcing or misinformation on the internet and what is a credible source has been a lifelong skill I've had to build from the early days of the web, and in the AI boom its only gotten more necessary to be able to not get taken by hucksters with billion dollar budgets.
When you do your own web searches, you learn to trust certain signals over others. You learn which sources are trustworthy and which are suspicious. LLMs don’t do that, and they present every information to you as being equally valid.
Not quite.
The difference is in rate of generation.
Today, the ratio of good data to slop has dropped. Previously, you couldn't mass produce websites and have content ready to go at a click.
So you had better hit ratios before, and you have worse hit ratios now. The more labour intensive to create content, you lived in a better world.
Hmm, The better the ratio of creating content to verifying content, the better the environment we live in.
Every lab is already training on synthetic data and has been for years now.
The same way you and I do: vibes
Welcome to the postmodern internet. It's vibes all the way down.
Upvoted you, of course; but it’s worse than that. It’s vibes being marketed as correctness. To the lay person (and unfortunately, to more than a few folks who should know better), computers don’t “make up” information. Maybe some good (in some weird way) that comes from all of this is that we stop using LLMs for recitation of facts.
They don't need to, they can use tools to validate their assumptions.
If someone has a vibe coded website with bad statistics but shows up on a web search, what tools will it use to check those statistics? How will it know what data it needs to validate? What tools will it use?
Humans face the same problem. So this at least shouldn't make AI perform worse relative to humans, even if AI slop degrades the performance of both over time.
I find it to not be acceptable if AI's trend is to degrade performance of both AI and humans at the same time, that's kinda not the goal, right? Why are we spending money to make us dumber?
AI isn't going to go away, and AI-generated content isn't going to go away. So while it's an open question the extent to which AI training will be hampered by the proliferation of AI-generated content, I think that the existence of such content is a reality that we'll have to accept, whether you like it or not.
> while it's an open question
I think theres research that shows that models collapse after being trained on their own content.
https://www.nature.com/articles/s41586-024-07566-y
Right, but we’re not talking about a worst case scenario where models are trained primarily on their own output. We’re talking about what will happen in the long run, as future models are trained on a realistic mix of content. It’s surely an open question what will happen.
> More likely (without careful vetting by the folks aggregating these models) is that the quality will go down as more and more AI-generated output gets subsumed into these models.
This assumes that there aren't algorithmic breakthroughs which reduce training/inference costs by several OOMs.
How much do these models need to do before people throw their hands in the air and say, ok this is happening. The Erdos unit distance problem, which as far as I understand was approached by multiple competent mathematicians was solved by a frontier model. Sure people argue there was no novelty there (I cannot comment as a non-mathematician) but it feels like they can draw lines laterally from deep knowledge in different fields (in this case combinatorics and algebraic number theory I believe) and solve problems.
Now if you have millions of instances running in parallel, all "probabilistic", working on frontier AI research I really don't see the blocker (and believe me I wish I did).
> How much do these models need to do before people throw their hands in the air and say, ok this is happening
What is "this"? Most people arguing against some of the more fervent predictions and promises of "inevitability" are people who are using these models in day to day - they see what the models can do, and what they struggle at.
> Now if you have millions of instances running in parallel, all "probabilistic", working on frontier AI research I really don't see the blocker (and believe me I wish I did).
My genuine prediction is that you'll get a lot of early results simply because you're applying attention to some low hanging fruit of problems, but then it will drop off due to the cost of tokens and the low rate of return. This doesn't mean that the models are especially capable of novel thought, just that we haven't algorithmically brute forced a problem with known solutions.
We would be seeing more success cases if the promises were true, setting aside AGI, human replacement, etc. We would see more, better products with more features that people would use. We wouldn't be having any arguments. The human replacement presupposes the models work in ways that they don't, and until proven otherwise, can't. I've watched those who embrace it fully flounder around on projects, some have lost their mind from the constant LLM validation, and I've seen companies go all in and then pull back based on both cost and efficacy over the last year.
I'm still waiting for the success case examples applied on a scale that would make any of the predictions come true.
> This assumes that there aren't algorithmic breakthroughs which reduce training/inference costs by several OOMs.
Yes, if one must assume something it is generally fair to assume that things will continue as they are. Research breakthroughs do happen, but they are not something for which you can predict the timing.
You're assuming that frontier AI research can "kick off the exponential" - that there exists such an exponential to kick off, and that frontier AI research can reach it soon, rather than in a couple centuries. Those are massive assumptions that have exactly zero empirical data behind them.
> How much do these models need to do before people throw their hands in the air and say, ok this is happening.
It's quite interesting to me that your epistemology is basically "God in a box is inevitable" and not skepticism.
> Now if you have millions of instances running in parallel, all "probabilistic", working on frontier AI research I really don't see the blocker (and believe me I wish I did).
The blocker is lots of things, but for one? Compute. Millions of instances running parallel, all "probabilistic" testing different theories would require million of GPU clusters.
Two? For subjects like Physics - Atoms. You need to be able to run experiments. You need access to the real world.
Then there's the S curve. Algorithmic breakthroughs operate on orders of magnitude. Eventually we'll run out.
etc. et al.
The only thing that matters is the economy of it all.
Open AI et al are hemorrhaging absurd amounts of money. It's not clear whether there will ever be a good balance between cost, value, and price.
Lots of companies are already questioning the value they get from LLMs at current prices which are obviously not enough to generate profits.
>As long as the term “AI” means by-and-large LLMs with additional features sprinkled on top, the answer is no. More likely (without careful vetting by the folks aggregating these models) is that the quality will go down as more and more AI-generated output gets subsumed into these models.
Nope. This isn't how it works.
AI progress has largely been synthetic and has produced leaps and bounds capabilities increases in the last couple of years.
Sorry.
>LLMs-as-AI can only give us probabilistic outputs based on inputs
I am not completely sure what you are saying here, but it sounds like a variation of the "it's just a stochastic parrot" argument, which is reductionist. The human brain is also just a bunch neurons firing.
> The problem for us isn’t that AI will take our jobs; it’s that snake-oil salesmen can sell the idea that AI will take our jobs, investors buy into it, companies try it, fire their folks, the snake-oil salesmen IPOs, the companies that bought into this idea implode in some form or fashion, and the salesmen have already taken the money and ran.
Or, it eventually becomes clear to enough people that the AI companies aren't going to make enough money to justify their valuations, so the asset bubble bursts, the economy crashes, and we lose our jobs.
Many jobs imo have been under threat for many yrs.
The tremendous growth in earnings meant some fake excess head count number was viable.
Google faced an activist investor who practically forced sundar to fire a bunch of people. This is what’s coming for big tech if this AI thing blows up. Apple Is safe because they are clever and saw this from a mile away.
Investors want their cake and they will eat it too.
Is model collapse still an unsolved topic? Once saw a method that determines if a training input is high quality.
This is all noise. The leaders of these companies are flip-flopping to whatever sounds best for their current agenda - hiring, fundraising, pre-IPO, etc.
The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential. Everything else is transient noise.
> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential
I agree with your sentiment (about the noise), however I think this over simplifies it a bit. We may get AI that is super-human at frontier research and dramatically accelerates the pace, and still have to wait decades before it disrupts the job market (or maybe never displaces all work).
For one, the answer may depend on material science and chip manufacturing that can take a very long to build out a supply chain for even with super AI help.
And we may just find that the human mind is way more capable than we thought and even with accelerating research it's just a harder problem than anyone expected, even algorithmically.
I expect it to be a bit of both, and from ~2015 - 2025 I was in the "AI is coming for all our jobs" camp. My perspective changed last year after doing a deep dive into latest science on the human brain. (I've kept a very close eye on AI dev progress for 12+ years.
> I agree with your sentiment (about the noise), however I think this over simplifies it a bit. We may get AI that is super-human at frontier research and dramatically accelerates the pace, and still have to wait decades before it disrupts the job market (or maybe never displaces all work).
I don't see why that's the case when you have super-human researchers on tap. There are indeed physical (supply chain-y) issues to deal with but isn't the whole point that: 1. Super-human at AI research + scaling to millions of instances will probably result in super-intelligence in everything which is not AI research. (a subset of which is white-collar work) 2. Use that super-intelligence to solve any supply-chain issues you might be facing.
> And we may just find that the human mind is way more capable than we thought and even with accelerating research it's just a harder problem than anyone expected, even algorithmically.
I hope so but whenever I do, I feel like I'm coping hard and not dealing with the facts.
I'm not saying we're there yet - I'm saying the trend lines are clear.
> Use that super-intelligence to solve any supply-chain issues you might be facing.
I think this is where a lot of people's thinking goes awry. Unlimited intelligence doesn't mean unlimited resources or instantaneous implementation.
> Unlimited intelligence doesn't mean unlimited resources or instantaneous implementation.
Of course you're right. At the end of the day you need to deal with the bedrock which is the laws of physics. I could be wrong but I struggle to believe we are close to the edge of what is possible in getting the most out of our limited resources or time.
Without atomic physics, uranium would just be another shiny rock in the ground. Sand is just what covers beaches. With enough time and intelligence we've made the shiny rock power cities and persuaded the sand to solve long-standing mathematical conjectures.
“I don't see why that's the case when you have super-human researchers on tap. ”
Hahahaa this is what AI psychosis looks like
> I hope so but whenever I do, I feel like I'm coping hard and not dealing with the facts.
Maybe I'm being naive, but as someone who spent 2015-2025 fully bought into the idea AI would replace all human labor within ~20 years, my updated take that it won't has a different flavor than coping.
A bit contributor to my perspective was studying the physics of the brain, and a deeper understanding of the very hard economic constraints in the current LLM supply chain.
Re the brain, we compute information in a fundamental different way that is over 1,000,000 more energy efficient than an LLM. We are also around 1,000,000 times more sample efficient at learning.
There is a lot of details in terms of specifics between complexity in human brain architecture (variety of synapses, diversity of neuron types, etc) that have no material equivalent yet.
And on the other side, understanding how hard it has been to build out euv lithography infra, and to completely retool manufacturing for a new paradigm of computing will require so much coordination, a super AI could have all the answers but we also have to deal with getting a bunch of stubborn people to align.
Some people are just coping yes, but if you really zoom into the details on both sides of the equations the gap between where we are and what it would take to dramatically surpass human intelligence is vast.
At the same time I think people overestimate how hard some kinds of knowledge work is (in absolute terms, since it's new and we haven't evolved into it as much), and how economically impactful human augmentation with weaker narrower AI can still be. (Lots of automation and prosperity but without such extreme disruption)
I think in the very long term all bets are off, but we will have plenty of time to evolve our economic systems before we get there.
> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential.
What if the answer is flatly: no? All that other stuff starts to matter a lot then.
Predicating your business decisions on a potential breakthrough that may never come is frankly insane. Imagine if at the dawn of the car industry Ford decided that it's actually a race to build the first flying car and nothing else matters.
> The only thing that matters is if LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential.
I think we know the answer to that already - LLMs show no sign of improving intelligence and instead providers are going down the ‘agentic’ rabbit hole.
There are too many things missing, like a world model, understanding, and taste (in the sense of knowing what is good and what is not good).
As long as LLMs don't understand the different between regurgitating facts and making up stories, they're going to necessarily be limited.
They are taught the difference through reinforcement learning with verifiable rewards. Pretending you've solved the task or making up a story about how you solved it won't do well in that training step.
> LLMs show no sign of improving intelligence and instead providers are going down the ‘agentic’ rabbit hole.
I'm not sure where you're getting this. I don't work at Anthropic but Fable (Mythos) seems demonstrably smarter than Opus for pretty much any definition of smarter and they claim that Opus was used heavily in Mythos development (yeah I know take this with a massive pinch of salt).
Either way if the models are indeed helping development, even on the engineering, you can iterate on models faster and even if they're not contributing to core research yet you still have a baby exponential by improving the engineering.
Anthropic's claims about their own products have almost zero value as evidence. They have lied to our faces about stuff before, and will again if it drives the hype cycle which delivers them money.
Sure, but let’s not pretend that people treated the statements of these ceos as strategic messaging. People very clearly treated what Altman, Zuck, Amodei etc have been saying as predictions, and it hasn’t been until they’ve been proven wrong that people have started with the counter-narrative.
It's important not to miss the fact that AI productivity was a useful excuse for companies looking to conduct layoffs. Did some companies buy the hype? Sure - but the biggest companies would have wanted that sweet stock price layoff bump anyways and AI was a readily available justification to get it.
We have two worlds:
1. Cutting edge LLMs developing ASI/AGI. 2. AIs doing general knowledge work
The second world will be achieved far before the first world is achieved. And as the first path gets develolped, the second path becomes cheaper and cheaper to run inference on along with being democratized which reduces the margins for the cutting edge companies. It seems like a mad dash to go as far as possible until 90% of general work can be automated with more cheaply available tech
> LLMs with sufficient scaling can become frontier AI researchers kicking off the exponential
The cost is already outrunning the benefit to a massive amount, and the predicted expotential is not here yet. I predict it'll always be around the corner, a $1T model won't get there, but it will "look promising", but we'll sadly run out of money for the $10T or $100T model..
The only thing the $1T model needs to do is find some algorithmic speedup which allows it to be trained at $100B. I'm not saying that's easy or that it will happen but I just don't see why not.
I don't understand the logic here - a methodology to increase efficiency by 90% in training doesn't exist until it does - could you explain what you mean by "I don't see why it won't exist"? Are you seeing consistent gains by some process?
The leaders of these companies have no stake in it. They are beyond worldly concerns.
Napoleon got sent to Elba. Hitler ate a bullet. Your average tech CEO will still have more money than they can spend.
https://archive.is/Pn7GU
Text-only, no Javascript, no ads/tracking:
My read has been that a lot of leaders were trying to drive “being early” as the catalyst for future success. At the complexity scale of big orgs you’re mostly fiddling with the incentives that the system self-aligns toward. Firing a bunch of people does create an incentive to use AI, if you think it’ll help.
The more pernicious effect I’ve been seeing is that we’re living in the golden age of LLMs, but eventually that’ll fade. Tokens are subsidized and cheap, model capabilities leap forward regularly, and there’s competition driving it all. But even now there’s stories about frontier models suddenly becoming less capable, or providers switching to usage-based billing, and new model releases feel a bit more sluggish and less dramatic. (Fable/Mythos notwithstanding.)
Eventually the models are going to settle into a rut of being just “good enough” to earn a living rather than all this hoopla. A lot of people will be re-hired. And we’ll do it all again for the next wave.
In case anyone else read the "flipped on" as "flipped the switch on" rather than "reversed course on", no, it's the latter.
Honestly, this is just the inverse of them all getting hyped on AI replacing all the jobs. Between both of these positions, RTO, crypto, VR, it's really shown just how much they're trend chasers.
Only a Tech CEO speaks in absolutes, it seems.
The ultimate irony here is that the biggest jobs wipeout most likely to happen now is when all these “AI exploration lab” type teams that every company quickly created are blown up.
Most, if not nearly all, of these teams have little to show ROI wise and the music on the AI bubble is slowing dramatically. They went from seemingly unlimited budgets and headcount when CEOs said “get me some of that AI” to some really uncomfortable scenes playing out know as the same CEOs realize this has cost a fortune with little to show for it.
Can we just speed this up a little? I just want my $100 RAM prices back.
Not so long ago I felt I outsmarted the system by using my 10 years old computers and laptops, but slowly I'm starting to squeak like a little mouse.
Dumbasses the lot of them.
Took that nonsense to Capitol Hill, trying to tell a bunch of politicians who knew damn well they are only there as long as they can keep their voters employed. They could have asked their own AI what happens when employment reaches 40-50%. Hint: it's never good. They were going to become another problem the government had to solve.
Also, UBI is non-starter no matter what Sam Altman believes.
Now let’s see if there will be any real consequences for such reckless stupidity: my bet, nah.
Of course not. This coming crash will be where we learn that tech is "too big to fail" in the same way that the financial system is. They'll let one player fail (likely Anthropic, due to the constant fighting with the government) and bail out the rest.
I guess I'm not sure why you'd expect it to be any different than the dotcom bubble, when this didn't really happen. "Too big to fail" is a thing in finance because of contagion factors that don't really apply to tech; there's no number of AI company bankruptcies that would force your bank to suspend withdrawals.
https://en.wikipedia.org/wiki/Collapse_of_Silicon_Valley_Ban...
I'm not sure what you're trying to imply with this link? As the first line of this article says, Silicon Valley Bank failed. The shareholders were wiped out. (Some people got confused because the Twitter conversation was primarily driven by rich VCs and founders who were depositors of the bank, quite reasonably arguing that they should get their money back if at all possible.)
The reason is because AI is viewed by the powers that be as inexorably linked to national security. If you haven't noticed, every lab out there has the same excuse when pressed about the disruption caused by their technology - "we do it or else China does."
If there ever truly were consequences in the oligarchy of the US, the time for that is long gone now. There will be none.
Afer COVID and AI, the only buzzword left is WAR. Doesn't rhyme but we are well beyond funny.
> UBI is non-starter no matter what Sam Altman believes.
Do you mean it's a non-starter in the current political climate? Or that you personally just don't think it will work?
It won't work because it's unnecessary. If these boneheads are right AI will drive the price of goods down. Price of labor goes down. It's deflationary. UBI is a backstop against inflation. And really just there so that rich people's money is worth something.
UBI can also be a backstop against deflation, can't it? If no one has income, it doesn't matter if bread costs $1 per loaf instead of $3-5 now - no one will be able to afford even $1 eventually, after they've sold their house at a loss, sold their car, sold their laptop, iphone, etc.
Sam Altman doesn't BELIEVE in UBI - but he knows it's the only way to SELL his product, therefore he pretends he believes in it. None of these guys want one cent going to anyone but themselves.
Of course they have. Fewer developers means fewer tokens sold.
Until AI no longer needs human supervision, it's more profitable to tax as many employees as possible.
You do not need developers to spend tokens, regular people can spend tokens just as easily.
This just tells me they don't know anything and they should probably just stop talking, and everyone else should stop listening.
Remember it was reported that OpenAI didn't think that ChatGPT would be successful? OpenAI thought that ChatGPT was yet another toy before its launch. Yet once ChatGPT became an overnight success, Altman started to talk about how AI would be dangerous, how it would displace or even replace jobs. In contrast, Amodei seemed to always believe in what he said. So, can we say that Altman is a opportunistic businessman, and Amodei is a cult leader?
IMO Demis has the most reasonable takes.
Yet layoffs excused by AI (expenses or claimed productivity gains) are continuing by the thousands each day, so... not sure what this is based on.
This is just based on Ramp’s data about their customers, but it seems that some companies are hiring more people after adopting AI?
https://econlab.substack.com/p/we-can-finally-say-ai-isnt-ki...
Individual firms can cut jobs while the overall economy sees more of that job pop up. That’s what happens when individuals become more efficient, new projects become more economical and more companies end up requiring these jobs.
I kind of wish there was a way to flip the script on the companies that gave up on humans and tried to switch to AI. Make them suffer for their idiocy in the same way that workers suffered or continue to suffer.
If that makes me a bad person, fine. If a few CEO's wind up working at 7-11 to make rent money, all the better.
> If a few CEO's wind up working at 7-11 to make rent money, all the better.
There are CEOs who have only ever failed abysmally their entire careers, and they generally only ever make more money. Accountability is for losers.
It will affect jobs because the investment strategy works to degenerate the economy, but not because of the tech. Like a dog that barks at a truck and thinks he scared it away.
article misses an important point that these big tech companies are all listed on the public market, any narrative about their decisions should weigh that reality and why suddenly its being disseminated.
personally, I am collecting 3 salaries working remotely. For one of the jobs, I am tasked with hiring other devs but i dont put the effort in as i dont see a point. i just say i can't find a decent engineer and why should i when a frontier models can do most of their work? in our job postings we see thousands of applicants in a very short period of time, i just do these multi stage interviews with a rotation of candidates to basically buy time while i work on another job
i see that things are getting very desperate and i feel for those that are still struggling to find SWE jobs, AI is absolutely doing a number and the gap is going to increase not decrease.
> there is an oversupply of SWE and a dwindling supply of jobs which pushes wages down
Which is exactly what every one on HN with a working brain predicted a decade+ ago.
decade+ ago AI agents were not a thing, it was only last year when it started picking up and there's less pressure to hire as frontier models improve massively
Any need to address the huge numbers of foreign workers in America putting direct, nonnegotiable pressure on those jobs?
It's time for the program to end.
I don't see the need and I don't think it has the impact you think it does
enterprises thrive when they have the freedom to choose who to hire without government interference
I don't want enterprises to thrive. I want me to thrive.
but if enterprises dont thrive you dont thrive either
they thrive the people in their own country struggle to survive
not seeing a problem with this is very suspicious
suspicious ?
The damage is done. The truly jarring realization isn't that AI companies made these predictions, but how eagerly countless corporations were willing to sacrifice their own people in pursuit of profit – even at the cost of economic collapse.
After watching the AI roll out for a couple years now I'm much more confident that it's just a scam. There is no net positive ROI on AI. It's not good enough, and the "mass job destruction" scenario offsets any marginal gains by eliminating the market for basically all products. That doesn't mean that mediocre C-suites won't try but it only takes 1-2 quarters to feel the burn and back track.
It's a rather useful tool... and it absolutely has been overhyped repeatedly and sold as a panacea.
If you believe AI will 10x you're developers you've drunk the kool-aid, if you believe AI will have no impact on your developers then you're being stubbornly ignorant.
What we're seeing is that AI is enabling us to reduce product headcount, but increase the scope of ownership for teams. We used to require each product keep at least 3 people on it to ensure that knowledge doesn't get lost with a departure, but now we're comfortable allowing an individual engineer drive it and allow AI to accelerate onboarding for a replacement when needed. This means that the teams can now handle 3x the products and our wish lists are getting shorter over time.
I recently tried to get customer service. I think it was from the DMV for my state. The website asked me to use the AI bot first. Completely useless. It took a phone call with a human to somewhat resolve the issue.
Dunno, the pencil sound effect as they let you speak to AI is quite calming. Now we need and efficient way to penalise claimants getting impatient or even rude at AI.
Sorry, but you aren't familiar with frontier models if you think it's not "good enough". The latest models are quite a bit more capable than most people in many regards already.
No, they aren't. Not even at the one thing (programming) which they are supposed to be best at. People are always claiming "frontier models are so much better" but that is a false claim.
Which "frontier models" are you familiar with in a programming context?
I'm a staff engineer and the amount of work and even troubleshooting the latest crop of LLMs can do at my behest is jaw-dropping.
I perused your comment history, and you refer to models as 'ChatGPT', which doesn't really inspire much confidence that you know what you're talking about.
People keep telling you that you're behind and using old technology and that the most recent models are really good and a sea-change because, well - they are.