The influence of generative AI on the US stock market, business culture, and popular culture is huge. However mounting evidence suggests that none of the large AI companies are anywhere near profitable. As time goes on, AI stock prices look more like stock bubbles of the past. This begs the questions:, when will it pop, how much will it impact the stock market, how bad will the inevitable ripples impact the general economy and our retirement portfolios?
The Generative AI Bubble Is really Going to Pop - Part Deux
- Thread starter AndrewZ
- Start date
I still go there, but not for the comments.
Slashdot is dead.
But it's a news aggregator with an editorial focus that theoretically matches my interests. (Not nearly as much now, and their actual editors are idiots)
From one of those links:
Even before AI I would scream and pull my hair out watching otherwise intelligent people using a search engine. That hasn't improved even with the advancements of AI and natural language interfaces.
First are the keywords, many people just don't get them and even at the natural language level their questions are so nebulously phrased as to be nearly semantically meaningless. That's why there will likely always be a place for good doctors, to translate the vague conversations with the patients to meaningful diagnostic inputs for an expert system.
The second problem that bothers me even more, got a good set of keywords, near perfect results - then the user not recognizing the answer to their question in the top result.
It's a great tool but humanity may be too stupid to use it. Ultimately fully functional humaniform robots with infinite patience and gentle interrogation skills to elicit the desires of the human without needing ESP may be in order. A corollary question might be what's the point of the humans.
and
Even before AI I would scream and pull my hair out watching otherwise intelligent people using a search engine. That hasn't improved even with the advancements of AI and natural language interfaces.
First are the keywords, many people just don't get them and even at the natural language level their questions are so nebulously phrased as to be nearly semantically meaningless. That's why there will likely always be a place for good doctors, to translate the vague conversations with the patients to meaningful diagnostic inputs for an expert system.
The second problem that bothers me even more, got a good set of keywords, near perfect results - then the user not recognizing the answer to their question in the top result.
It's a great tool but humanity may be too stupid to use it. Ultimately fully functional humaniform robots with infinite patience and gentle interrogation skills to elicit the desires of the human without needing ESP may be in order. A corollary question might be what's the point of the humans.
Once saw this great quote from a park ranger about the challenge of designing trash cans for national parks. The problem being that there's considerable overlap between the intelligence of the smartest bear and the dumbest people.
Every new tool has some learning curve, and I think a lot of people just don't get how to use AI seinsibly. I joined a training we had for using AI to make meetings more efficient and it was literally just someone telling people to prompt "give me an agenda for a meeting." Yeah, that's not going to be useful. The whole point is not to end up with a generic template.
Every new tool has some learning curve, and I think a lot of people just don't get how to use AI seinsibly. I joined a training we had for using AI to make meetings more efficient and it was literally just someone telling people to prompt "give me an agenda for a meeting." Yeah, that's not going to be useful. The whole point is not to end up with a generic template.
Hey Skoop, the AI bubble is a huge thing. I listed just a few reasons why it's huge in the opening post. I know the first AI bubble thread got a little off topic, but the posts were still interesting and somewhat pertinent. I'm not quite sure why you shut it down when it could have kept going with a little moderation. In my tenure at AT this was the first thread I've ever had closed. Anyways, this is highly topical and relevant. I'm hoping it lives. Maybe you can elucidate me on the ways of the wise moderator.
It got more than a little moderation. It was off topic for most of its tenure. Most of the posts were pertinent only if you squinted real hard at them.
The business of AI is an important topic. If the posts this time stick to business, then it will have legs. But the topic at this point, in this forum, has two strikes.
The business of AI is an important topic. If the posts this time stick to business, then it will have legs. But the topic at this point, in this forum, has two strikes.
The hyperscalers seem to still be growing revenue and profit while sinking billions into GenAI, partly because of layoffs, but also because each has some core piece(s) of business that just keep on printing money - e.g. ads for Google and Facebook. As long as they keep growing, investors aren't going to care whether that growth actually came from GenAI or not. If users adopt GenAI en masse, they'll enshittify it with ads and keep printing money. If not, they'll figure out where the users are and serve them ads wherever that is.
And for Microsoft, the big revenue growth driver is Azure. Lots of people using lots of cloud compute. Same for Amazon. GenAI demand obviously contributes to this, so they have an incentive to hype it. But if massive demand for GenAI doesn't materialize, will cloud compute usage decline, or will we humans keep finding things we want technology to do that rely on data centres?
IDK. I've been a US stock market skeptic for a minute, there are signs of distress all over, but the hyperscalers just keep making number go up.
And for Microsoft, the big revenue growth driver is Azure. Lots of people using lots of cloud compute. Same for Amazon. GenAI demand obviously contributes to this, so they have an incentive to hype it. But if massive demand for GenAI doesn't materialize, will cloud compute usage decline, or will we humans keep finding things we want technology to do that rely on data centres?
IDK. I've been a US stock market skeptic for a minute, there are signs of distress all over, but the hyperscalers just keep making number go up.
From what I see inside businesses - Gen AI usage will likely increase because people will start using GenAI for regression or classification problems that can be solved with less computationally expensive models like XGBoost because it is fast and easy to put together.
It has ever been thus. Hardware makers make more capable hardware; software people find ways to use it that may not be efficient, but increase capabilities all the same. I see no end to the demand for compute. That's not terribly specific to AI, AI models are simply the most recent big thing. Whether or not AI itself is a "bubble" I'm not sure. I see enough value that I believe these kinds of models will continue to be used and have an appropriate business case, but it may be overhyped.
I think It's easy for most objective and informed observers to agree that AI is a bubble, but the much more interesting question is what will be the trigger that finally causes it to pop?
The problem with the AI bubble is that yes, it is a bubble and you can slap AI on something that has 0 AI and get tons of seed funding. But there are also some seriously impressive things AI is doing so it gets recognized hype that inflates the bubble.
I mean, even ars' own Lee H. Used it to write a perl script and he admitted it was totally acceptable compared to half a year ago.
I mean, even ars' own Lee H. Used it to write a perl script and he admitted it was totally acceptable compared to half a year ago.
Fortunately the FP has provided on-topic fodder to keep the lounge at bay
https://arstechnica.com/information...bble-while-seeking-500b-valuation-for-openai/
Not that the overall track record of major IT projects is good by any means but yikes. Once the hype bubble springs a leak, it will collapse quickly
https://arstechnica.com/information...bble-while-seeking-500b-valuation-for-openai/
Not that the overall track record of major IT projects is good by any means but yikes. Once the hype bubble springs a leak, it will collapse quickly
That OpenAI seems to have their for-profit conversion on hold. A bunch of funding, primarily from Softbank, is dependent on that, plus penalties on other funding like investment-conversion-to-loan by the end of the year if still held as a weird inside-shell of a non-profit. Has there been any news on restarting this reincorporation effort since the Musk offer and California complexity scuttled it this spring?
Also a general plug for Ed Zitron's analysis. Just the free parts give plenty of insight into how wacky and weird the financial positioning of the industry is. Including some frankly absurd creative accounting techniques that mean almost nothing about business sustainability especially when released only when they look good.
His analysis is wonky as hell lol, anyone with a bit of background reading can shoot tons of holes in it.
In the last free to read piece - he was on a crusade against annualizing revenue and dedicated a whole page of text to it. Wtf, who cares if it is 30 days -> 365 days or calendar months. Yeah, sure they want to "leak" a new bigger number every time they sign a new client, it's a race, why would you want until the next calendar month.
The "oh no Cursor no longer gives away tokens for under the cost price / model owners can charge whatever they want" part - yeah no shit, that's why the big money is with foundational models like Gemini, GPT and Claude. And of course the ones selling tools like every gold rush. But that's what you get for depending on frontier models for coding, not commodity like whatever is available on Groq and Fireworks.ai. Most applications work just fine on GPT-OSS-120b, during my test run of it, crafting system prompts and testing capability, it knows a ton about the (culinary) world, is multilingual, fast and cheap, and has test-time compute / CoT / whatever you call that.
Then a weird huge wall of text, many pages about who would acquire Cursor. Why? It's just a VSCode fork. Gemini CLI is proper Open Source, Claude Code is better though, and Alibaba has forked Gemini to Qwen Coder. Even Jetbrains has a half decent agentic plugin, Gemini added it in their VSCode and Jetbrains plugins, every provider will have their own in a few months.
Other focus more on the next level - currently, these are like advanced driver assist systems, and you are still in control. Next level is executing tasks autonomously, like Github Copilot sending pull requests (but it's not good enough for that yet), OpenAI has their own too, and Google Junie, all of them very new and at the edge of breakthrough - but not good enough for anything complex yet. Note: YET. The "L2" driver assist style AI coders were shit too just 6 months ago, extrapolate from there and these "L3" autonomous bug fixers could be there in a year.
Then the next another wall of text is titled "Get Acquired, Go Public, Or Die". Never mind, close tab. Sure if you aim for that, AI isn't the gold mine / easy path to riches as it was before a while back. But if Cursor and all the stupid low code LLM wrappers disappear, who cares. Frontier models still perform better every few months with a new release, current capability is plenty to x5 a senior's productivity. And open weight models make same big jumps, DeepSeek was the king with no equal, now it has plenty of competition.
What a waste of space and time to read, but to untrained eyes, this looks like a perfect doom piece to extrapolate from, to why all AI would fail.
I for one enjoy testing the capabilities of commodity models and check how it would help me do my work. The new open weight GPT-OSS-120b in particular sets a new benchmark in efficiency and there are plenty of tasks it could do well. That price/quality is crazy, and if you build a service around it - be it invoice recognition, customer service assistance / automated hinting / translations (we get tickets in every language, it would be great if it always normalizes to English first)*, it's a nice value added service with tiny running costs, but huge boon for productivity. And the token price war, as it continues, will keep pushing capability / accuracy up and cost down. I still can't believe a $0.15/MTok model (GPT-OSS-120b) performs this well at everything I threw at it.
It's boring, hard work and definitely not worth a billion dollar market cap though, so people who want to get rich quickly with the latest buzzword will be disappointed.
* Our human agents already copy paste it into a Gemini powered tool / workbench to scan for FAQs, translate, and and even suggest answers back in original language, it saves a ton of time while improving quality - you spend more time on the real cases, write a nice Jira ticket with screenshots, and less on simple AI solvable issues. But it's not tied together yet, we should spend time and glue it all together with Zendesk API now tooling seems to be stable - it was moving too fast, from the Workspace included Gemini App, to NotebookLM, to various playgrounds. GPT-OSS-120b will probably be the model of choice, and either Groq / Fireworks as provider / serverless API endpoint.
Hmm thinking about this more, the examples of customer service failures are all from companies skipping the "driver assist" phase of customer service tools. That works just fine / great even, but you need a human driving the car. Skip forward to fully automated chatbot or phone service, and no, there are plenty of edge cases it cannot handle, trigger hallucinations or tricked to say things it shouldn't. That's just aiming too high, the model can handle it just fine, but the glue around it is lacking. It's easy to blame AI / the model and not the consultant writing shit code and failing to integrate it properly or overpromising.
Ok, to shoot back: where's the revenue?? Correlated, are there enough GPUs to realize the revenue from substantially replacing real workers' total comp with a substantial part of their former salary going to an LLM company/provider, net? Who's got a model that can do even customer support?
So the straightforward uses might be too hard to integrate, but it's worth tons of money to deploy alone, somehow? What's the business case for the integrator to be able to do this well and turn their profit while also having the processing required be profitable for the model operator and/or provider?
How does that result in sustainable business for "frontier model" advancement? Is it estimated (because none of these companies really say) that their analysis rates are sustainable? Against opex? Against inference capex? And, of course, against training/refinement/downscaling expenses in general? No one has reason to tell you any of these things backed by GAAP, yet.
There's a lot of uses that are sensible on individual level but do not create any net value on the whole. For example using generative AI to fake grant applications is great for the grifter, but it's a net negative and doesn't create any long term surplus to pay for AI with.
Then there's deep learning's diminishing returns. The diminishing returns is why when top-of-the-line LLMs pass any "good enough" barrier, within less than a year, much-cheaper-to-run LLMs pass it too. So even if it is creating surplus, it is highly dubious it can capture that surplus.
What will happen if, hypothetically, diminishing returns thing became false? What is the biggest practical consequence of diminishing returns right now?
It's that you and I can use AI - it makes more sense to "split" a datacenter into a bunch of idiot-instances and rent them out than to attempt to build 1 slightly less stupid idiot that takes up the whole datacenter.
If diminishing returns thing became false, this all goes poof - it becomes far more profitable to use the datacenter as one, or very few large entities. The scifi AI apocalypse scenario.
If there was ever an LLM that is somehow "PhD in STEM" equivalent, nobody would be trying to persuade us to pay for it, they would be able to do something useful with it themselves, like e.g. build a Von Neumann asteroid mining robot or whatever (edit: as opposed to, checks notes, windows updates that corrupt your hard drive and break streaming).
Last edited:
I'm not really surprised there've been a bunch of AI failures. How to use GenAI effectively are still unsolved across most domains, and it's pretty frustrating to have to lift your entire context into a chatbot to use it. Seeing ChatGPT do something cool or useful is a lot different than productionalizing something that isn't garbage, but I've found executives to fall victim to "I can imagine it, so it must be easy" more than other classes of PMC worker. So they YOLO an AI program that's way, way premature.
In general, it's a lot more effective to bring the robot into your context, and I think tools like Claude Code are beginning to show a lot of promise because they do just that. (It's actually increased my happiness significantly, too.)
Combining genAI with MCP is also pretty powerful, but it's super, super new, and security is largely unsolved. At my company, we're doing this pretty well:
I think there's a good business to be had in agentic SRE things. For example, I regularly look at the slow query log in RDS, and I send the query to Claude or ChatGPT, and I get something pretty useful back. In most cases, I can just copy and paste, and open a PR, and our performance issue is solved. Sometimes greater context is needed but "Missing an index on column Foo" happens a lot more often than people think. I can only imagine that this is true to an even greater extent at non-tech companies.
We're just scratching the surface for useful ways to bring robots into pre-existing contexts instead of artificially lifting the context to the robot.
--
On the economics of tokens, this chart is pretty interesting. One of the claims I made on the first page of the OG thread was that the cost of tokens will fall as the field advances. I'm not sure whether this is good news for the makers of robots or not, especially as they seem to commodify within a year or two of a release. Should be interesting spillover effects for society, though.
In general, it's a lot more effective to bring the robot into your context, and I think tools like Claude Code are beginning to show a lot of promise because they do just that. (It's actually increased my happiness significantly, too.)
Combining genAI with MCP is also pretty powerful, but it's super, super new, and security is largely unsolved. At my company, we're doing this pretty well:
- Certain people can query our quantitative environment to prototype new quantitative ideas using natural language, and without having to know the schema of the available datastores. We are in the process of launching two new features whose R&D started that way.
- We're using MCP + Claude to do interesting things with observability. We're using Claude as the interface to build and publish groq parsers in new relic to turn unstructured log messages into structured, queryable things we can build dashboards around.
I think there's a good business to be had in agentic SRE things. For example, I regularly look at the slow query log in RDS, and I send the query to Claude or ChatGPT, and I get something pretty useful back. In most cases, I can just copy and paste, and open a PR, and our performance issue is solved. Sometimes greater context is needed but "Missing an index on column Foo" happens a lot more often than people think. I can only imagine that this is true to an even greater extent at non-tech companies.
We're just scratching the surface for useful ways to bring robots into pre-existing contexts instead of artificially lifting the context to the robot.
--
On the economics of tokens, this chart is pretty interesting. One of the claims I made on the first page of the OG thread was that the cost of tokens will fall as the field advances. I'm not sure whether this is good news for the makers of robots or not, especially as they seem to commodify within a year or two of a release. Should be interesting spillover effects for society, though.
Last edited:
One of the weird tensions is the cost of tokens falling as smaller models get more capable, but the biggest models demand more and more electricity because the underlying compute is hitting a power wall in the last year-ish.
The hyperscalers have enough money to build on site power for their compute needs, if they can get it built fast enough (some are reaching for the natural gas easy button and finding out it's 5 years and 2x+ the cost to buy a turbine for a combined cycle gas plant now).
Continually increasing costs to stay at the frontier while costs to consumers decrease due to competition and commodification isn't a recipe for long term profit growth, and I'm not sure how that circle gets squared without some kind of level reset. But maybe the hyperscalers are just too big to fail now, IDK. They could light hundreds of billions on fire and be fine.
The bubble might gradually deflate rather than popping, in that case, which honestly would be better for everyone.
The hyperscalers have enough money to build on site power for their compute needs, if they can get it built fast enough (some are reaching for the natural gas easy button and finding out it's 5 years and 2x+ the cost to buy a turbine for a combined cycle gas plant now).
Continually increasing costs to stay at the frontier while costs to consumers decrease due to competition and commodification isn't a recipe for long term profit growth, and I'm not sure how that circle gets squared without some kind of level reset. But maybe the hyperscalers are just too big to fail now, IDK. They could light hundreds of billions on fire and be fine.
The bubble might gradually deflate rather than popping, in that case, which honestly would be better for everyone.
GPT-5 isn't a case of charity, they really optimized it to run on less hardware.
OpenAI also shook up the market for open models as well, GPT OSS 120b runs on a single 80/96GB card, by heavily using 4 bit math, and it results in the low $0.15 input, $0.60 output token price. I expect GPT-5 to use the same trick, it's crazy fast and cheap.
Frontier models don't focus just on making the craziest biggest thing that can run, they had two of those, o1-pro, $150/MTok, o3-pro, $20, to explore if throwing more compute at it made sense. It generally didn't, it was better at some specific cases but GPT-5 generally does better than o3, without the crazy hardware/power requirements. I have a feeling frontier models are now moving towards making it cheaper and faster, it's good enough and the improvements no longer feel significant. See the GPT 4.1 to 5 discussion, better? Maybe. Different? Yes.
Pure token sellers don't sponsor your project, they aren't funded to light money on fire unlike Cursor did, provide a loss leader, maximize annualized revenue to unlock next funding round. The focus shouldn't be "substantial part of their former salary going to an LLM company/provider", it's providing decent value and let people glue together, LLM assisted, things that provide real value.
Like for each incoming ticket, if not in English, translate. For each ticket, scan for FAQ, add a hint for agent.
CSR can then just click ticket, okay, auto accept, next. Or nope not quite, let's discuss this more -> Claude goes "You're absolutely right" and writes a better response -> accept -> done. You can just write a few words and the AI can expand that into prose in whatever writing style you want. Haiku, Shakespear, Star Trek officer, whatever you want.
The stupidity of most project failures is to delete the human from the loop. Then you get worse instead of better service, sometimes even impossible to get a real human on the phone or email. They need to learn to walk before running.
Again, it's not too hard to integrate, you just need to do "honest, hard work", look at requirements and improve the workflow by sprinkling some LLMs on top. People have been doing that for ages now, Expert System is a 1950's idea, https://en.wikipedia.org/wiki/Expert_system
At least LLMs are much easier to work with vs older, more stiff and formal approaches. But it is far from going sentient and the context window of ~100-200k usable is far too low to feed in a whole company's knowledge. This means you need to do "honest, hard work", plan and execute well, no magic button to push to replace a customer support department.
Why does it matter? OpenAI as an entity is now an entity to turn VC funds into models and lots of GPUs. Their rapidly multiplying revenue is okay for investors to keep burning money.
Google doesn't do funding rounds, but can easily pay the capex and opex from its own revenue and profit.
When investors get scared and pull out, all there is left is a bunch of models and GPUs. That's just fine, whoever picks up the assets will be able to offer tokens at a much discounted rate, if it was funded by debt before. But as far as I know, all the rounds are for equity and a minority stake, the founders / board can do whatever they want.
If they switch from investment / race mode to launching a sustainable business in the future, we'll see what frontier model tokens cost at opex parity price. But DeepSeek R1 costs $3 at Fireworks, with a decent profit margin and no sustained use / contract discounts, I suspect GPT-5 and Gemini at non ultrathink / high / max mode to cost about the same or less. And at the same time, everyone is pushing for better -mini or -flash models.
Gemini Flash (0.30/MTok) is quite decent now, GPT-5 mini (0.25/MTok) is plenty smart for CSR role, you don't need GPT-5-high for that. GPT-OSS-120b (0.15/MTok) sets a benchmark in price/quality. These should be your workhorse. If you launch a Cursor clone (or see any other ai tool spam/promo post on Reddit), you need a frontier model or it would just write buggy code, but for prose and text, similar words with same meaning are fine. Style differences, who cares, just steer with system prompt. You can easily launch products on top of these 3 and not bleed money, $10/month/user buys a ton of tokens to parse / process emails. A standard incoming ticket, I parsed a few, is ~120 tokens in 550 characters. Not everyone lets Claude read / search / parse a ton of code to fill up 100k context window.
Finally at the hardware / software / libraries front:
Note that training is still done on H100/H200's right now, the Blackwell ecosystem is still starting up and when they get it up and running for training models, cost will drop by like 40% or more. Libraries and kernels also get better over time, on the same hardware:
Newer models spend significant engineering time to use FP4 wherever possible, to double throughput / half energy use. GPT-OSS-120b did it for a significant part of the weights, not all, some are too critical to quantize to just 16 distinct values + block based exponent.
Dodging the question by that much makes this an unproductive conversation. Microsoft and Google can enshittify their products by bundling services. Even with enterprise contracts, that's still not the same scale as OpenAI seems to be targeting revenue.
Remember, just OpenAI wants trillions spent (13-figure spend). Bouncing over to the source for this front page comment, all US labor, nationally, for May 2024 is 154,187,380 employed with an estimated mean annual wage of $67,920 each. That comes to ~$10.5T of traditional payroll (self-employed aren't covered in the survey) for the nation. Other than tearing back the intro programmer market that might actually just be tech companies covering for over-hiring, again, show me the money!
Remember, just OpenAI wants trillions spent (13-figure spend). Bouncing over to the source for this front page comment, all US labor, nationally, for May 2024 is 154,187,380 employed with an estimated mean annual wage of $67,920 each. That comes to ~$10.5T of traditional payroll (self-employed aren't covered in the survey) for the nation. Other than tearing back the intro programmer market that might actually just be tech companies covering for over-hiring, again, show me the money!
Hahaha if you believe that made up number, well, no wonder you're so stuck in your mindset.
And no, not "just OpenAI", no one else, Google, Meta, Anthropic, even Musk, has the same idiot idea in the next years. They will just spend whatever to get enough GPUs, but not "trillions".
Where would they spend it? Nvidia's annual revenue is $130B total in 2024.
Altman sells the idea that AGI is around the corner and the first to reach it will have an unstoppable lead. Investors that believe it let him set their money on fire. We don't even know if Altman really believes it himself, he's a good marketeer, but probably not this stupid. Softbank otoh, lol.
Still, OpenAI will first need to get the money. Their last round raised 6.6B for 2.3% of shares, making its market cap 157B, but no way anyone can exit a significant part without cratering the price so it's mostly imaginary. And that 1 trillion plan is imaginary as well.
You "just asking questions" doesn't mean everyone else has to do your homework. You can ask an AI for that yourself. This forums seems to frown at me pasting generated replies here, so I won't bother, but depending on provider they are well sourced and well thought out.
Research doesn't mean finding all the factoids that support your opinion. Try some AI chat and tell it to be critical. Your "show me the money" is bullshit because the trillions is Altman's bullshit, duh you can't show revenue to match that.
Last edited:
Some of them may have believed it early on, but I doubt any do now. I think smaller LLMs will cannibalize at least some of their market. The chat-buddy/therapy market for the big vendors should get vaporized. They are far too reckless, and are doing their best to push legislation so they couldn't be held accountable anyway.
More broadly, the "AGI is just around the corner" is priced into tech valuations in general, putting their stocks far above established major industries that aren't "tech", including actual tech. NVidia having higher market cap than TSMC and such. Even though TSMC has guaranteed moat due to lead times on the equipment, while NVidia has a very soft moat (there's so much money in this space, all the major players can afford custom hardware).
Other bullshit-quasi-analysis-providers echo Sam's sentiment as well, by the way, as a simple websearch could tell you. Automatic bullshit generators also agree (gemini example ), although they just try to guess the desired answer and then go maximum sycophant. Most people don't disable chat history so there's a plenty of clues to guess what the user wants to hear. (I had the history disabled for that example).
I think it is absolutely the case that valuations are inflated due to perceived closeness to AGI.
If you believe that the investors are taking a sober look at all that and doing some modest calculation involving just the least-unproven applications*, accounting for the risk from smaller models becoming good enough, then I have a bridge to sell you.
* lets not forget that even for programming with all the "it made me 10x more productive", actual data based on fixed endpoints is scarce and conflicting.
edit: to be clear I think its a bubble driven primarily by "line go up" extrapolation - investors see line go up, they invest. Index funds automatically increase allocation, etc. The AGI hopes and trillions on datacenters and other such malarkey provide an imaginary way out other than a crash. Delaying but worsening the crash.
Last edited:
The AGI push is obvious in the coordinated news and PR campaigns. It's intensive and pretty cruel, frankly. A lot of young people believe that shit and throw in the towel on making future plans, drop out of university, etc.
Maybe because I use it all the time, but I haven't heard anyone talking about AGI lately. Admittedly, I haven't gone looking for it, as LLMs are really just a (very useful) tool to me at this point.
I don't think anyone believes LLMs are going to bring AGI at this point. Maybe the idiots at Softbank, but being credulous hype-believers is kinda their whole brand.
I actually feel like the LLM hype cycle is starting to turn? I spend a lot more of my time talking about it and using it than I did a year ago, but that's mostly because it's actually more useful across a wide variety of tasks I need to do. The people around me are talking about what they're using it for in their day to day existence rather than positing what might be possible. At least in my circles, it seems like a significant tone change over the last 6 months.
I don't think anyone believes LLMs are going to bring AGI at this point. Maybe the idiots at Softbank, but being credulous hype-believers is kinda their whole brand.
I actually feel like the LLM hype cycle is starting to turn? I spend a lot more of my time talking about it and using it than I did a year ago, but that's mostly because it's actually more useful across a wide variety of tasks I need to do. The people around me are talking about what they're using it for in their day to day existence rather than positing what might be possible. At least in my circles, it seems like a significant tone change over the last 6 months.
Honestly we need an "AI" soapbox thread, like "When did America stop dreaming big?...". These chatbots are not something apolitical or "neutral", the way AI used to be. We've already replaced NIH / HHS experts with "AI". It is a right-wing, fascist, anti-reality, anti-environmentalist, anti-art, anti-intellectual undertaking. On the environmental issues, the masks are completely off (edit: fixed link).
Also another example of AGI hype at an ACM event ("A Virtual Side Event at the 4th International Conference on Financing for Development, 30 June – 3 July 2025, Seville, Spain"):
View: https://www.youtube.com/watch?v=df5OhAvu6Hc&t=5267s
Saying that it is just Sam Altman is utterly at odds with reality.
I think the investors have the opposite perspective. They see you using it, they see a lot of crap being vibecoded, they worry that they will run out of new users soon, and then the growth phase valuations will crash.
If you just listen to what they're saying, it is quite clear that they understand they need continuous increase in capabilities to justify the valuations.
edit: Also this already happened with regular office workers. Anekdotally, our CPA and her friend were very excited about ChatGPT half a year ago. If they deemed that to be good enough to rely on for legal advice, it's going to be very difficult to sell them any kind of premium AI now, and harder still in half a year, and especially so given that the premium offerings are not even better at that kind of thing.
Hence the "AI agents" being the next big thing, the hope being that the business owner is a more discerning customer that would pay a premium for the better AI agent.
Last edited:
Maybe we'll start to see Professional Errors & Omissions and Malpractice insurance policies start to question GenAI usage by lawyers, CPAs, etc, once someone's unverified advice to a client turns out to be wrong and costs them real money
I think we should have seen that by now, give the really blatant failures that have been in the news. But maybe that news is just in other arenas from what we read.
Professionals, be it lawyers, software engineers or tax consultants, have lots of freedom to use and abuse tools. The idiot cases in the news are the rare exceptions.
Most who successfully use chatbots for research know which one does decent work, and to verify all output. OpenAI models hallucinaties a lot. Claude literally warns you when you ask something it is not well trained on.
The real successes come from hard work, like the RAG + 999 pages prompt from KPMG's TaxBot, not dumping a question in ChatGPT and copy pasting its reply.
But as a professional, you need to own your work. No matter who wrote it, yourself or Claude, once you send/submit it, it's yours, mistakes and all, no blaming the bot. If you read the lawyer fails this way, it's not ChatGPT's problem, it's the lawyer not verifying referenced cases exist or laws applied correctly, basic due diligence not done.
But people don't like hard work, they like no work, but still get all the rewards, that's why the AGI dream is so easy to sell to everyone. That AGI bubble will blow up when everyone discovers that Gen AI is garbage in, garbage out, there is a limit on what you can achieve with public data, the rest must be produced internally like those KGMP seniors writing tax advice feeding TaxBot. And that costs a huge amount of time and resources.
https://www.washingtonpost.com/education/2025/08/26/alpha-school-virginia-ai-education/
Bill Ackman apparently wants to start a teacherless AI-taught school that charges students $65,000 a year.
Bill Ackman apparently wants to start a teacherless AI-taught school that charges students $65,000 a year.
Yeah, this. It seems like a lot of users are rushing for lazy solutions and the people getting useful things from it are the ones investing a lot of careful effort.
My point is more that most people who thought it smart to rely on 2024 chatgpt for legal advice, are not going to become discerning customers now. These customers are largely lost to commodization of AI already.
It is also a pretty silly conjecture that more compute would be even useful for most of the work that LLM “AI” can do. Yes, with more weights you can store more information, but RAG is a far better solution than using weights to store information.
Anthropic's own lawyers submitting AI slop is a great example of what kind of "rare exception" it takes to get into the news: it has to be funny/ironic (AI company's own lawyers, ha ha!), high profile, and the judge has to notice.
Last edited:
It's companies that buy enterprise subscriptions and API keys that will develop and mandate properly working systems. You can't place the burden on each individual users.
Simple case, to edit a file, agents need to invoke the edit tool with exact text to search for and replace. It must be exact and correct. Hallucinations will return a tool error. This, and other (MCP) tools will provide safety in a walled garden.
I don't disagree. Scaling to GPT-6, 7, 8 using the same method is a dead end.
I suspect spending more time in post training would work better. Like supervised fine tuning, already available in production for minimal additional cost. https://fireworks.ai/blog/supervised-fine-tuning-tutorial
This way you can train the model to use tools better, like code graph search, at a deeper level than providing a bunch of instructions in the context. You nudge the weights in the direction of a prompt reply you prefer.
[edited to add example]
DeepSeek is a capable model, but LoRA improves it for your specific usage by a crazy amount. Example loss after training on a 800 messages set, 720 training 80 test split
View: https://medium.com/@rafaelcostadealmeida159/how-to-fine-tune-deepseek-r1-using-lora-7033edf05ee0
Nice, another anecdote... You can always find another case, and another, etc, of fuckups. When it stops being news it's a problem. Tons of bridges work just fine yet every now and then a high profile fuckup happens and gets the spotlight like the collapse a few days ago in China, you can't use anecdotes to prove your case.
What you can do though is point out the trend, just like bridges, fuckups/total going up significantly, past n sigma, is news. Figure out how to apply this to gen AI fuckup and write a blog post or something.
Last edited:
Paywalled but if it's as dumb as it sounds then the problem will solve itself. No one is paying that much especially as grade school is as much daycare as it is education
Nvida is profitable, Google is profitable, microsoft, meta, apple, etc. They are all profitable. Some of their investments in ai might not be bring in more return than investment, but the companies overall are profitable. AI hype can die tomorrow, and all these companies will still be profitable. This is why its not a bubble, because even if it "pops", these companies are just back to where they were before their ai investments, except for Nvidia, which would be trillions of dollars richer even if the ai investment train stops tomorrow.
But then what is a bubble? Was dotcom a bubble? Is anything ever a bubble?
Plenty of dotcom era companies were like this - profitable businesses, but their stocks were massively overvalued and did undergo a corresponding "correction". Intel went from, what, over $40 at the peak dotcom bubble, to $11 or so. That's a bigger relative move than the recent one caused by not being all that profitable. Cisco had a high of $52 and a low of $5.6 a couple years later.
I don't think anyone's expecting NVidia to go under here. They have a competent CEO who doesn't lay off the goose who laid the golden egg the moment he's trying to sell the second egg and finds the sale price to be somewhat less than expected.
Last edited: