The argument that computational complexity has something to do with this could have merit but the article certainly doesn’t give indication as to why. Is the brain NP complete? Maybe maybe not. I could see many arguments about why modern research will fail to create AGI but just hand waving “reality is NP-hard” is not enough.

The fact is: something fundamental has changed that enables a computer to pretty effectively understand natural language. That’s a discovery on the scale of the internet or google search and shouldn’t be discounted… and usage proves it. In 2 years there is a platform with billions of users. On top of that huge fields of new research are making leaps and bounds with novel methods utilizing AI for chemistry, computational geometry, biology etc.

It’s a paradigm shift.

Where the current wave all falls apart is on the financials. None of that makes any sense and there’s no obvious path forward.

Folks say handwavy things like “oh they’ll just sell ads” but even a cursory analysis shows that math doesn’t ad up relative to the sums of money being invested at the moment.

Tech wise I’m bullish. Business wise, AI is setting up to be a big disaster. Those that aimlessly chased the hype are heading for a world of financial pain.

I've been programming for 30+ years and now a people manager. Claude Code has enabled me to code again and I'm several times more productive than I ever was as an IC in the 2000s and 2010s. I suspect this person hasn't really tried the most recent generation, it is quite impressive and works very well if you do know what you are doing

Yes, there is hype.

But if you actually filter it out, instead of (over) reacting to it in either direction, progress has been phenomenal and the fact there is visible progress in many areas, including LLMs, in the order of months demonstrates no walls.

Visible progress doesn’t mean astounding progress. But any tech that is improving year to year is moving at a good speed.

Huge apparent leaps in recent years seem to have spoiled some people. Or perhaps desensitized them. Or perhaps, created frustration that big leaps don’t happen every week.

I can’t fathom anyone not using models for 1000 things. But we all operate differently, and have different kinds of lives, work and problems. So I take claims that individuals are not getting much from models at face value.

But that some people are not finding the value isn’t an argument that those of us getting value, increasing value isn’t real.

Last week, when I was on PTO, I used AI to to a full redesign of a music community website I run. I touched about 40k lines of code in a week. The redesign is shipped and everyone is using it. AI let me go about 5-10x faster than if I would have done this by hand. (In fact, I have tried doing this in the past, so I really do have an apples to apples comparison for velocity. AI enabled it happening at all: I’ve tried a few other times in the past but never been able to squeeze it into a week.)

The cited 40% inaccuracy rate doesn’t track for me at all. Claude basically one-shot anything I asked for, to the point that the bottleneck was mostly thinking of what I should ask it to do next.

At this point, saying AI has failed feels like denying reality.

The current AI hype is fueled by public markets, and as they found out during the pandemic, the first one to blink and acknowledge the elephant in the room loses, bigly.

So, even in the face of a devastating demonstration of "AI" ineffectiveness (which I personally haven't seen, despite things being, well, entirely underwhelming), we may very well stuck in this cycle for a while yet...

But there is so much real economic value being created - not speculation, but actual business processes - billions of dollars - it’s hard to seriously defend the claim that LLMs are “failures” in any practical sense.

Doesn’t mean we aren’t headed for a winter of sobering reality… but it doesn’t invalidate the disruption either.

As with so much of Silicon Valley's output the past few decades, "AI" is not something people will pay for. It could be a new form of Trojan Horse for data collection, surveillance and ads, though, adopting a hyperfocus on "growth" at any cost (human, environmental, etc.) but this might not even be necessary^1

Time will tell

In order to raise funds in the absence of any clear business plan, "AI" cannot just be ordinary and useful. It cannot be "autocomplete on steroids". It must be "world-changing"

No one shall underestimate or discount the value of "AI". Overestimation is permitted, however

1. "AI" is not depending on being brought inside the gates the Troy by the city's inhabitants. It's being placed inside the gates by others. Companies with large numbers of users are installing "AI" on users' computers without the users' consent. Demand for "AI" is assumed

AI tooling has only just barely reached the point where enterprise CRUD developers can start thinking about. Langchain only reached v1.0.0 in the last 60 days (Q4 2025); OpenAI effectively announced support for MCP in Q2 2025. The spec didn't even approach maturity until Q4 of 2024. Heck most LLMs didn't have support for tools in 2024.

In 2-3 years a lot of these libraries will be part way through their roadmap towards v2.0.0 to fix many of the pain points and fleshing out QOL improvements, and standard patterns evolved for integrating different workflows. Consumer streaming of audio and video on the web was a disaster of a mess until around ~2009 despite browsers having plugins for it going back over a decade. LLMs continue to improve at a rapid rate, but tooling matures more slowly.

Of course previous experiments failed or were abandoned; the technology has been moving faster than the average CRUD developer can implement features. A lot of "cutting edge" technology we put into our product in 2023 are now standard features for the free tier of market leaders like ChatGPT etc. Why bother maintaining a custom fork of 2023-era (effectively stone age) technology when free tier APIs do it better in 2025? MCP might not be the be-all, end-all, but at least it is a standard interface that's at least maintainable in a way that developers of mature software can begin conceiving of integrating it into their product as a permanent feature, rather than a curiosity MVP at the behest of a non technical exec.

A lot of AI-adjacent libraries we've been using finally hit v1.0.0 this year, or creeping close to it; providing stable interfaces for maintainable software. It's time to hit the reset button on "X% of internal AI initiatives failed"

The reason is hype deflation and technical stagnation don't have to arrive together. Once people stop promising AGI by Christmas and clamp down on infinite growth + infinite GPU spend, things will start to look more normal.

At this point, it feels more like the financing story was the shaky part not the tech or the workflows. LLMs’ve changed workflows in a way that’s very hard to unwind now.

I think I will keep using it while it's cheap, but once I have to pay the real costs of training/running a flagship modell I think I will quit. It's too expensive as it is for what it does.

> Depending on the context, and how picky you need to be about recognizing good or bad output, this might be anywhere from a 60% to a 95% success rate, with the remaining 5%-40% being bad results. This just isn't good enough for most practical purposes

This seems to suggest that humans are 100%. I'd be surprised if i was anywhere close to that after 10 years of programming professionally

Tokens/week have gone up 23x year-over-year according to https://openrouter.ai/rankings. This is probably around $500M-1B in sales per year.

The real question is where the trajectory of this rocket ship is going. Will per-token pricing be a race to the bottom against budget chinese model providers? Will we see another 20x year year over the next 3 years, or will it level out sooner?

exactly - just like 99.9873% of all codebases currently running in production worldwide :)

for example, fictional stories. If you want to be entertained and it doesn’t matter if it’s true or not, there’s no downsides to “hallucinations”. you could argue that stories ARE hallucinations.

another example is advertisements. what matters is how people perceive them, not what’s actually true.

or, content for a political campaign.

the more i think about it, genAI really is a perfect match for social media companies

AI just got better and better. People thought it couldn't solve math problems without some human formalizes them first. Then it did. People thought it couldn't generate legible text. Then it did.

All while people swore it had reached a "plateau," "architecture ceiling," "inherent limit," or whatever synonym of the goalpost.

If AI models can deliver measurably better accuracy than doctors, clearer evaluations than professors and fairer prosecutions than courts, then it should be adopted. Waymo has already shown a measurable decrease in loss of life by eliminating humans from driving.

I believe, technically, moderns LLMs are sufficiently advanced to meaningfully disrupt the aforementioned professions as Waymo has done for taxis. Waymo's success relies on 2 non-llm factors that we've yet to see for other professions. First is exhaustive collection and labelling of in-domain high quality data. Second is the destruction of the pro-human regulatory lobby (thanks to work done by Uber in the Zirp era that came before).

To me, an AI winter isn't a concern, because AI is not the bottleneck. It is regulatory opposition and sourcing human experts who will train their own replacements. Both are significantly harder to get around for high-status white collar work. The great-AI-replacement may still fail, but it won't be because of the limitations of LLMs.

> My advice: unwind as much exposure as possible you might have to a forthcoming AI bubble crash.

Hedging when you have much at stake is always a good idea. Bubble or no bubble.

Winters are when technology falls out of the vice grip of Capital and into the hands of the everyman.

Winters are when you’ll see folks abandon this AIaaS model for every conceivable use case, and start shifting processing power back to the end user.

Winters ensure only the strongest survive into the next Spring. They’re consequences for hubris (“LLMs will replace all the jobs”) that give space for new things to emerge.

So, yeah, I’m looking forward to another AI winter, because that’s when we finally see what does and does not work. My personal guess is that agents and programming-assistants will be more tightly integrated into some local IDEs instead of pricey software subscriptions, foundational models won’t be trained nearly as often, and some accessibility interfaces will see improvement from the language processing capabilities of LLMs (real-time translation, as an example, or speech-to-action).

That, I’m looking forward to. AI in the hands of the common man, not locked behind subscription paywalls, advertising slop, or VC Capital.

Assuming these claims are even partially true, we'd be stupid—at the personal and societal level—not to avail ourselves of these tools and reap the productivity gains. So I don't see AI going away any time soon. Nor will it be a passing fad like Krugman assumed the internet would be. We'd have to course-correct on its usage, but it truly is a game changer.

Lemma: any statement about AI which uses the word "never" to preclude some feature from future realization is false.

Lemma: contemporary implementations have almost always already been improved upon, but are unevenly distributed.

(Ximm's Law)

Really? I derive a ton of value from it. For me it’s a phenomenal advancement and not a failure at all.

> People were saying that this meant that the AI winter was over

The last AI winter was over 20 years ago. Transformers came during an AI boom.

> First time around, AI was largely symbolic

Neural networks were already hot and the state of the art across many disciplines when Transformers came out.

> The other huge problem with traditional AI was that many of its algorithms were NP-complete

Algorithms are not NP-complete. That's a type error. Problems can be NP-complete, not algorithms.

> with the algorithm taking an arbitrarily long time to terminate

This has no relationship to something being NP-complete at all.

> but I strongly suspect that 'true AI', for useful definitions of that term, is at best NP-complete, possibly much worse

I think the author means that "true AI" returns answers quickly and with high accuracy? A statement that has no relationship to NP-completeness at all.

> For the uninitiated, a transformer is basically a big pile of linear algebra that takes a sequence of tokens and computes the likeliest next token

This is wrong on many levels. A Transformer is not a linear network, linear networks are well-characterized and they aren't powerful enough to do much. It's the non-linearities in the Transformer that allows it to work. And only Decoders compute the distribution over the next token.

> More specifically, they are fed one token at a time, which builds an internal state that ultimately guides the generation of the next token

Totally wrong. This is why Transformers killed RNNs. Transformers are provided all tokens simultaneously and then produce a next token one at a time. RNNs don't have that ability to simultaneously process tokens. This is just totally the wrong mental model of what a Transformer is.

> This sounds bizarre and probably impossible, but the huge research breakthrough was figuring out that, by starting with essentially random coefficients (weights and biases) in the linear algebra, and during training back-propagating errors, these weights and biases could eventually converge on something that worked.

Again, totally wrong. Gradient descent dates back to the late 1800s early 1900s. Backprop dates back to the 60s and 70s. So this clearly wasn't the key breakthrough of Transformers.

> This inner loop isn't Turing-complete – a simple program with a while loop in it is computationally more powerful. If you allow a transformer to keep generating tokens indefinitely this is probably Turing-complete, though nobody actually does that because of the cost.

This isn't what Turing-completeness is. And by definition all practical computing is not a Turing Machine, simply because TMs require an infinite tape. Our actual machines are all roughly Linear Bounded Automata. What's interesting is that this doesn't really provide us with anything useful.

> Transformers also solved scaling, because their training can be unsupervised

Unsupervised methods predate Transformers by decades and were already the state of the art in computer vision by the time Transformers came out.

> In practice, the transformer actually generates a number for every possible output token, with the highest number being chosen in order to determine the token.

Greedy decoding isn't the default in most applications.

> The problem with this approach is that the model will always generate a token, regardless of whether the context has anything to do with its training data.

Absolutely not. We have things like end tokens exactly for this, to allow the model to stop generating.

I got tired of reading at this point. This is drivel by someone who has no clue what's going on.

Lol someone doesn't understand how the power structure system works "the golden rule". There is a saying if you owe the bank 100k you have a problem. If you owe the bank ten million the bank has a problem. OpenAI and the other players have made this bubble so big that there is no way the power system will allow themselves to take the hit. Expect some sort of tax subsided bailout in the near future.

Here's a few clarifications (sorry this is so long...):

"I should explain for anyone who hasn't heard that term [AI winter]... there was much hope, as there is now, but ultimately the technology stagnated. "

The term AI winter typically refers to a period of reduced funding for AI research/development, not the technology stagnating (the technology failing to deliver on expectations was the cause of the AI winter, not the definition of AI winter).

"[When GPT3 came out, pre-ChatGPT] People were saying that this meant that the AI winter was over, and a new era was beginning."

People tend to agree there were two AI winters already, one having to do with symbolic AI disappointments/general lack of progress (70s), and the latter related to expert systems (late 80s). That AI winter has long been over. The Deep Learning revolution started in ~2012, and by 2020 (GPT 3) huge amount of talent and money were already going into AI for years. This trend just accelerated with ChatGPT.

"[After symbolic AI] So then came transformers. Seemingly capable of true AI, or, at least, scaling to being good enough to be called true AI, with astonishing capabilities ... the huge research breakthrough was figuring out that, by starting with essentially random coefficients (weights and biases) in the linear algebra, and during training back-propagating errors, these weights and biases could eventually converge on something that worked."

Transformers came about in 2017. The first wave of excitement about neural nets and backpropagation goes all the way back to the late 80s/early 90s, and AI (computer vision, NLP, to a lesser extent robotics) were already heavily ML-based by the 2000s, just not neural-net based (this changed in roughly 2012).

"All transformers have a fundamental limitation, which can not be eliminated by scaling to larger models, more training data or better fine-tuning ... This is the root of the hallucination problem in transformers, and is unsolveable because hallucinating is all that transformers can do."

The 'highest number' token is not necessarily chosen, this depends on the decoding algorithm. That aside, 'the next token will be generated to match that bad choice' makes it sound like once you generate one 'wrong' token the rest of the output is also wrong. A token is a few characters, and need not 'poison' the rest of the output.

That aside, there are plenty of ways to 'recover' from starting to go down the wrong route. A key aspect of why reasoning in LLMs works well is that it typically incorporates backtracking - going earlier in the reasoning to verify details or whatnot. You can do uncertainty estimation in the decoding algorithm, use a secondary model, plenty of things (here is a detailed survey https://arxiv.org/pdf/2311.05232 , one of several that is easy to find).

"The technology won't disappear – existing models, particularly in the open source domain, will still be available, and will still be used, but expect a few 'killer app' use cases to remain, with the rest falling away."

A quick google search shows ChatGPT currently has 800 million weekly active users who are using it for all sorts of things. AI-assisted programming is certainly here to stay, and there are plenty of other industries in which AI will be part of the workflow (helping do research, take notes, summarize, build presentations, etc.)

I think discussion is good, but it's disappointing to see stuff with this level of accuracy being on front page of HN.

But that's me being a sucker. Because in reality this is just a clickbait headline for an article basically saying that the tech won't fully get us to AGI and that the bubble will likely pop and only a few players will remain. Which I completely agree with. It's really not that profound.

People have figured it out by now. Generative "AI" will fail, other forms may continue, though it it would be interesting to hear from experts in other fields how much fraud there is. There are tons of material science "AI" startups, it is hard to believe they all deliver.

What we should underscore though, is that even if there is a new AI winter, the world isn’t going back to what it was before AI. This is it, forever.

Generations ahead will gaslight themselves into thinking this AI world is better, because who wants to grow up knowing they live in a shitty era full of slop? Don’t believe it.