AI Bubble talk
In contast to the other recent forum post, I wanted to bring up another important side of the AI discussion: people calling it a bubble.
Recently a MIT study entitled "GenAI Divide: State of AI in Business 2025" was just released. Among many other topics, it concluded that: People like ChatGPT for basic tasks and hate complicated enterprise systems, and companies that try to build their own AI usually fail.
The MIT study details that these issues are not model quality issues, but implemenation issues. It cites "the learning gap" as a core issue for both the AI tools themselves but also organizations themselves. (Imagine how much time it took you to get good with Google years ago, now imagine having to repeat that for each new tool). Articles elsewhere conclude that this likely implies that this isnt a failure of AI as a whole, but of corporate IT departments mistakenly believeing they can out engineer the major AI specialist corporations like OpenAI and do it themselves.
My question is this: How is this different than any other tech scenario in a major corp? Based on my experience, even with something as simple as a Kanban board platform, people inside of corporations using an internally created Kanban board are severely handicapped for many reasons. The money saved by using an internal tool ususally results in a 2-4x loss in productivity based on my rough anecdotal estimations.
Some examples are: its not nearly as feature rich as the commercial product, people that made it never fully documented it, people that made it have moved on and there is nobody who really acts as support for it anymore, etc.
So, Does this study pointing to enterprise failures involving AI point to a bubble? In my opinion, on this specific topic, I don't believe so. I don't think its much different than any other tech tool used by corporations. If it is different, I would love to be able to see that kind of comparison. Also comparing how the above cited figures change over time as models get more advanced would also be interesting.
What are your experiences with internal tools vs commercial products both within and outside of the AI realm?
Oh, this is a topic I can definitely talk about. So, for context, I majored in applied math and stats in college, and part of that was doing some ML classes and ML research. I also read most of the major ML papers coming out nowadays. With that aside, I'd like to comment a lot on what you're bringing up.
So, I went ahead and read the paper from MIT, I'll link it here
https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf But some key points I want to make here is about the learning of these models. Right now, the way Ai works is in 2 states: Training and deployment. During training phases, the model is given a ton of data, tokenized, and figures out how to weight certain responses. Say we want an LLM to give us data on valuing a firm, like Google. Well, we'll train this model based on a bunch of things. Past valuations and DCFs, news data, 10k's, corporate earnings, and based on what the model creates we can then say "Yes, this is correct" or "No, this is wrong". This is a fundamental piece of the transformer technology underpinning the current Ai revolution. Using supervised learning to validate all this is generally a good thing.
Here's the problem: Once a model enters production, it doesn't adapt. This is a key point about Ai, in the sense that it is still machine driven. Machine run on parameters that are hard to change. And a key point about machines is that you can actually modify the way they run. But anyways, this became a problem for many firms and users because the Ai tools would repeat mistakes. This makes sense, when you consider the fact the LLM is purely working off the training data. Let's say for example, you want to value a new firm, like Apple. Well, if that prior LLM is trained purely on valuing Google, then it will make the same assumptions about valuing Apple. We know as humans this is poor practice, and a great thing about humans is that we learn very well from our mistakes. Ai, on the other hand, will not actually understand it is making a mistake, because once again it is a machine. It is a model, and as someone who's spent years working with models(my wife included) I fully understand how they solely make outputs based on input training and assumptions. If you build a DCF, it won't actually think for itself. It's just a model. Same thing here. The LLM is mathematically designed to output the most optimal combination of words or numbers(with it altering sometimes going for the second or third based on random weighing as to seem more human) rather than actually thinking about the problem.
On your KanBan point(and internal tech point), the problem is shadow Ai. Most people using Ai are not using internal tooling. If you have used them before, you'd understand this. The facts are that OpenAi, Google, and other tech giants are currently paying 2-3X the amount finance companies are paying with better culture(for tech forwards people) and better hours, and so of course the best models will be developed there. Yes, complexity of these internal tools, like you stated, is oftentimes a reason not to use them. 2/3rds of internal tools fail, while 1/3rd of external partnerships fail.
I do have to disagree about this not being a failure to engineer Ai as a whole. I made a long comment a while ago in another thread that I think a lot of people would find interesting, on the state of Ai and why it isn't as present as you'd think, which I'll link here:
www.wallstreetoasis.com/forum/investment-banking/future-of-the-industry…;
TLDR is that the math of transformer technology runs in o(n^2) which means scaling models to massive inputs, like a whole entire investment banking team working a massive model introduces new costs that we have yet to see due to the massive subsidization of models. But on the point of Ai as a whole, I've been rather disappointed by most model releases nowadays. Sure, we're seeing video models improve a ton, but language is plateauing because we still have yet to figure out the math problem of transformers along with reinforcement learning, a key research topic that will 100% be integral to Ai in the future.
In saying all this, I'd like to make a bit of broader commentary on Ai right now. Ai is fundamentally not that different from a very complex statistical model. Ai is still a research intensive topic, and while we may make breakthroughs every few years that push us forwards, we're still so far from the true AGI we've been hearing about. Think of it kinda like curing cancer. Everyone hears about all the ways we could potentially cure cancer, as it is a pervasive and terrible disease, but we can't cure it. Why? Because research on cancer is incredibly difficult, expensive, and the problems of cancer are ever growing. Most cancer research is incremental gains on radiology, not massive breakthroughs in immunotherapy. Most research breakthroughs in Ai are slight model tweaks, not massive fundamental changes like the Attention Is All You Need paper.
On the topic of bubbles, I really don't see how this cannot be considered a bubble. I mean, seriously, Palantir is being valued more than Bank of America, Chevron, Samsung, and LVMH. Nvidia makes more than half of their revenue from just 4 customers: Microsoft, Google, Amazon, and Meta. That's absurd. That's absolutely insane, that Nvidia is basically being driven to 4.2T in market cap, which is more than the entire global pharma industry and nearly worth the entire US insurance market. Everyone can see that the valuations are insane, the payoffs aren't present, and the growth isn't happening. Public markets and VC firms continue to subsidize this, but for how long? How long until everyone says "We've invested 10 Billion dollars to make 2 billion dollars"? How long until we stop seeing chatbots and random Ai garbage being pumped into every software service in the hopes to capture a tiny sliver of a market that doesn't yet exist? I saw an Ai washing machine. A fucking washing machine. Don't get me wrong, cool idea, but really? Am I incapable of washing clothes? Does my machine need Ai?
No, these models remain the same constrained models due to the basic laws of physics and mathematics. And now we've pumped trillions into companies that cannot justify their valuation. Sure, Ai is amazing, and will 100% change the world, but it isn't changing now.
If you don't believe me, believe Sequoia. They are positing a 600B dollar question. That question is trying to be answered right now, by big tech firms. Ever wonder why big tech is so desperate for companies to replace workers with ChatGPT? Ever wonder why they keep talking about how Ai will do XYZ, just give it a few years? It's because they want companies to pull a Klarna. They want companies to layoff people, replace them with this crappy Ai, and then finally they jack up the prices to make these models actually profitable, because if you read the thread I posted you'd know why Ai firms are losing so much subsidizing this stuff. They want companies to be dependent now, while prices are cheap, so that once they increase prices, these companies are forced to pay, but companies are increasingly gawking at this. 55% of companies already say they regret replacing workers with Ai, and we're barely a few years into this crap. If we see that now, imagine how much it'll be in a few years.
I preface this by saying, I am a mere amateur on this subject, so apologies if my questions are silly.
Doesn't it seem like the chinese economy is ultimately more suited for AI adoption? They have an industrial base, they have large supply chains, dark factories, infrastructure, machines that would produce productivity gains far in excess, if they were to adopt AI vs a financialized and service economy like the US. AI's sole though principle from american leaders seem to be cost cutting and neo-religious babbling about AGI/God/whatever the fuck else their strange minds come up with. In China, the AI leaders there, seem to talk about integrating AI all throughout their supply chains, and their physical products, whether it be robotics, EVs, etc...
Would love your thoughts. I know this opens up a geopolitical angle, which is distinct from the mathmatical one you were operating from before, but still interested in anything you may have to add to the conversation.
In my opinion, we may be experiencing an AI bubble not because AI won;t be useful, but because the market seems to be pricing in radical productivity gains on a very short timeline, perhaps within just one to two years. For me, GPT-5 has highlighted that progress in these models is not as exponential as it was initially marketed to be. I;ve gone from believing AGI was likely within the next 5;10 years to now being more skeptical. It feels like cracks are starting to show in the narrative spun by figures like Altman, Musk, and Zuckerberg;a narrative that once seemed full of imminent breakthroughs but is now beginning to feel more like overextended promises.
First, and most importantly: are you saying your wife is a (fashion/ models & bottles) model, or that she also works on ML models?
I'll admit my current knowledge of AI is limited compared to yours. My only real connection is that I designed and outlined trading models back in 2013 or so where you basically: brute force all combinations of variables until you get the best prediction accuracy (which you predefine with time span and profit/loss percentage) with the optimal set of variables, obviously limited by your degrees of freedom aka dataset size. (please forgive me getting stat terms wrong, I havent looked at that stuff in over 10 years)
To my surprise years later, when I was doing some learning for continued growth by taking a basic AI lesson on kaggle, thats the exact algorithm that is the basis of AI.
I'm not sure that growth in AI's abilities are strictly limited by n^2 aka parabolic resource needs. There are often advances in algorithms etc that can net you a X increase in performance with a Y decrease in cost, and those have nothing to do with simply scaling up resources and costs.
Random thought/question: Are AI models mistakes largely updated and "learned from" when the next model is completed and then released? If so, it might be analogous to how humans update their memories etc during the night when we are asleep.
I'm not sure anyone is saying that AI is fully able to replace full human educated workers yet, but I do believe that it can currently streamline a team. Even if AGI isnt possible, I think current advancements in models, in addition to people becoming more adept at using it, will allow teams to get more and more streamlined, or at least output more and more for a given labor input.
Personally, it seems like the current state of AI should allow for construction esque robots that can fully replace workers, but not as much money and resources are being put into that because the cost of that labor is low and thus the ROI isnt as high.
fwiw, "Sure, Ai is amazing, and will 100% change the world, but it isn't changing now." These are speculative stocks, and stocks are inherently forward looking to begine with. Isnt Nvidia's P/E multiple 3x lower than Teslas?
This final argument seems to be based on the assumption that models will get worse over time. I'm not sure thats the case. "55% of companies already say they regret replacing workers with Ai, and we're barely a few years into this crap. If we see that now, imagine how much it'll be in a few years."
Oh yeah, I meant that interal tools not being as feature rich is a negative. Anytime you have do do anything beyond basic, things get super complicated very quickly and you have to start jumping through more and more hoops to get everything to still jive correctly.
That was just me complimenting my wife
Yeah, so this is an astute observation that is definitely true. What you're describing is most close to an RNN, or recurrent neural network. What an RNN does is parse over every word in a sentence or every number in a sequence. This creates issues. 1, you need the models to remember things. For example, lets take the sentence "The black cat likes the tree, it will go there at noon". Now, an RNN will struggle a lot with this, because it will will be able to match black and cat, but it might fail to match that cat to liking the tree, or the time it will go to the tree, or something else. So oftentimes, these models would just fail to process basic information. These models also were not able to be parallelized, meaning you could only have 1 GPU processing at a time. This created major constraints, as training was slow, processing was slow, and data was inaccurate.
What a transformer does is take this fundamental concept of an RNN and add in a self attention mechanism. What this basically does is pair every token against every other token in the sequence. So, for example, "The black cat likes the tree, it will go there at noon" would process through a transformer by taking every single word in that sentence and pairing it with another. This means the model all of a sudden is able to pair the black cat, the tree, and the time, by using an attention score to each element. This is the biggest fundamental change to how Ai is done, as this technology innovation has basically driven Ai for the past couple years. Ironically, it's not even all that crazy. It's just taking RNNs and doing some mathematical weighings to bind key pairs together.
If you read what I wrote about transformers above or my entire post about the o(n^2) problem, you'd kinda understand why this is a problem. o(n^2) is fundamental to how these LLMs operate. They literally cannot operate any differently. This is because you cannot have a transformer that does not check every token against every other token input, or else you've recreated the same issues as the RNN in forgetfullness. You can try SSMs, but those have data loss issues, and a bunch of other potential problems along with being nowhere near production grade. This idea that algorithms expand is generally false. We've had CS as a study for decades, yet some things we cannot break because of the basic laws of mathematics. You cannot make a sorting algorithm faster than o(nlogn) because of comparison. Same thing applies here. Because of the basic fact transformers compare in a brute force fashion, you cannot create a model that doesn't run in o(n^2). Even as we've seen innovative new model techniques like DeepSeek's micromodels(instead of creating one monolith, create hundreds of specialist models and pick the best one for each task), we still can't escape math.
Not really, no. The problem is that this would require humans to manually report every error, and humans at Ai companies to check all those errors and manually fix them. They can fix themselves over time, but this is a big reason why for years models couldn't tell you how many R's there were in strawberry. It would require very manualized labor to specifically fine tune a model out of very specific negative behaviours. Reinforcement learning does fix some of these issues, but we're still pretty far out. We don't really know how to do RL in production models at this point. It's always going to be "error, update, fix, deploy" which is, well, exactly what a SAAS could already do. So basically we've made fancy SAAS.
Yeah, I think that's obvious to anyone. It's a tool for use as of now.
Comparing anything to Tesla is a bad idea in the world of finance. The problem with your statement is "forwards looking" vs "speculation". When a company like JP Morgan or Mastercard makes a revenue guidance, we can generally rely on it. We can sustainably say that it's very likely those firms will be around, making plenty of money tomorrow, because they're in industries that are profitable and pay them plenty to keep the lights on. With Ai, all that needs to happen is for one of the big tech firms to say "Look, we're not making money from this, we're scaling back and focusing on profit" and all of a sudden, all those Nvidia revenue and earnings projections are gone. And the problem is, these stocks are priced as if all the speculation will 100% come through, and Ai will 100% make trillions of dollars and benefit only these few companies at the top.
The problem is still fundamentally that without a major breakthrough, the models won't get much better. The key issues still remain, with error prone models and massive cost runups due to the time complexity. If 55% of firms are regretting it now, just a few years in, when we're at the peak, imagine what companies in a few years will be saying?
Sure, models will improve, but by how much? People seem to not realize, the easy pickings from Ai are gone. Those massive leaps we saw from just cramming more data, optimizing more parameters, and making small alignment changes are gone. It's like quant trading in the 70's vs now. The easy gains are gone, and everyone is fighting over what gains are left to be had, competing for small chunks.
Just wanted to say that GPT-5 is materially worse than 4o. It's absolute garbage. End rant.
One of my relatives works in the AI engineering side and he believes we're in the AI bubble because the lack of undertanding around what AI capabilities can do and what it cannot do.
He thinks that this will eventually lead to some type of tech bubble in the public markets, not at the scale of the dot com bubble burst though.
The primary purpose of internally-developed software is to promote the compensation/career of the individuals pushing the project, whether those are internal people or outside consultants.
There are always exceptions but most of the time, improving productivity wasn't the true goal (only the stated goal) so we shouldn't be surprised when it fails.
AI isn't a bubble, but anything can look like a bubble when viewed through the lens of 'why does the internal version suck'.
I've seen specialized tools, especially letting different software platforms talk and share data that had great rois for the company that developed them internally and deployed them. That isn't anywhere near the majority though...
I've also heard rumors of extreme versions of what you described. A companies internal data vis platform was half ready for deployment at best, but it was pushed on various groups within the corporation. It literally required a professional programmer to do basic charts and graphs that excel can do any day of the week, and you guessed it, the director pushing it had a team of professional programmers ready to fill that (fake) need. What once took any old excel dude 5 minutes, now took collaboration with an external group, and hours being billed by programmers to achieve basically the same results.
It's incredible the lack of common sense from observers who see what you described, and fail to figure out that the incentives drive the outcome. Starts at the top.
It just means slow to adopt companies will be crushed by companies that can adapt to tech faster.
We have tremendously slowed down hiring at all of our portfolio brands because we are replacing workloads with AI. It's amazing what you can do with N8N and a weekend.
The main thing I've found is that it takes a lot of ~1 - 3 hour tasks and cuts the time required to complete them by ~60% - 90%.
Examples:
-Creating SOPs -> feed a voice note and have it spit out a nice SOP. Turns ~1 hour tasks into ~10 min
-We replaced a lot of our customer service agents with AI. Almost all responses are now AI generated. For color, one co doubled and hired 0 additional CX agents lol.
-Our creative team at one brand was reduced by ~70% and output is WAY up because we simply generate what we need with AI + canva. Insanely fast. Again, rev is MUCH higher than last year.
I could keep going...
There's no way it's a bubble imo.
I guarantee your porfolio co's customers hate example 2 and 3 you gave. I can only hope the slop machine is actually accurate for example 1.
Our NPS scores + repeat purchase rates haven't changed. If they cared, those would look worse.
I'm not so sure about the creative team workflows because some things need a professional to edit/format and polish content. If you have PortCos using Canva maybe they aren't technical groups but once you look at more specialized applications AI is probably going to augment workflows to increase output as opposed to just obsoleting them. I think there is a lot of AI being oversold.
More early days of finance before electronic trading.
Less final days of mail in catalog shopping.
I'm talking about team members that have no graphic design skills being able to just whip something up that they need instead of asking the creative team for it.
For example, our wholesale team needed some assets for a deck when pitching a big box retailer. They made all the graphics themselves with Canva's AI tools > the graphic design team.
Or the ads team wants video versions of top performing static ads. Can do that in 10 seconds with VEO3 > shooting the raw footage & editing it. It's lower quality but it doesn't matter because the cost is SO much lower that you can test literally 100x+ the concepts lol
Would you care to share impacts to financial metrics (rev growth, margins etc) you are seeing? Im increasingly seeing sponsors implementing AI tools and get the sense the productivity gains are happening even if nascent but I don’t have solid financial metric.
8% of spend on people/salaries to 6% for one of our brands. I think that's a good example.
Our AI could cut down our workload by 95% if it was actually accurate at pulling information.
Short term, yes there is a bubble. Tech companies have laid off thousands under the guise of "AI-driven efficiency," but we all know it's just cost-cutting due to inflation and high interest rates. Speaking from personal experience at a tech company, I'm seeing no material impact on productivity post-layoffs. We've cancelled dozens of projects, and things are taking longer than usual due to a lack of engineers.
That being said, I agree that current AI limitations "are not model quality issues, but implementation issues." I said this in a previous post here that ChatGPT 5 is already a borderline AGI technology. If you disagree, what exactly is your definition of AGI? It's trained on the full corpus of human knowledge and can generate accurate, reasonable answers on virtually any topic.
If you don't like its answers, it's likely because the chatbot lacks the tribal knowledge that you've absorbed on the job without knowing it. Imagine you picked a random person off the street and fed them your company's entire knowledge base. Even a genius would get a few things wrong because they're missing contextual knowledge, which can only be gained after months of learning your company's quirks. Even separate teams within the same company have diverging priorities and communication styles. The barrier now is to simulate that same "on-the-job" experience for an AI chatbot. The chatbot's sheer IQ isn't the problem, it's the inputs.
This is correct, and already it can absolutely replace the vast majority of bankers, consultants, lawyers, or any profession where the work output is limited to Microsoft Office.
For the love of ray jay we are in an everything bubble since zero interest rate policy: stocks, bonds, real estate, private debt, and especially private equity including the venture capital tech mania which includes ai. Not to mention crypto. We’ve never seen money printing to this extent by the United States in our lifetimes trying to prolong the inevitable burst.
Is the federal funds rate still 4.25 to 4.50, or has it changed?
I had lunch with the Dean of Accountancy at my alma mater last week and some public accounting firms that would normally hire 5 new associates, he’s seeing them hire 2 this year. He thinks AI has a lot to do with it. He is very pragmatic and is trying to get AI learning into the accounting curriculum but legacy professors resist. They want to teach their recurring syllabus.
College education, job market, and corporate adoption of AI will be evolving rapidly (1,000 new jobs created in NY?).
I believe especially as a college student or soon to be, you have two choices 1) thinking this as a shitty time to be starting your career, or 2) thinking that this is the best time to accelerate your career, wealth, opportunities.
I think AI implementation is going to impact larger companies’ earnings first. Reminds me of how META rebounded in recent years using cost cutting. Could be a broader theme in the economy.
I am afraid of the impact to peoples’ livelihoods.
I could be off by a few years, but I think this prediction will happen.
I completed an MS in Accountancy in 2021 from the #2 program in the country. We definitely had some computer classes that were unfortunately totally worthless as well as some electives. There is definitely space to add AI training without totally re-doing the core courses. I think it's important, however, for the AI/computer classes to have someone who understands accounting/finance being actively involved; otherwise, you'll end up like us with totally pointless computer classes not applicable to the job.
Consequatur ipsam et facere quas inventore qui. Et vero suscipit temporibus non excepturi. Autem inventore quaerat possimus doloribus. Illum asperiores voluptatem ducimus eos est. Culpa saepe voluptatem ut maxime voluptatibus esse. Totam mollitia natus libero sunt.
See All Comments - 100% Free
WSO depends on everyone being able to pitch in when they know something. Unlock with your email and get bonus: 6 financial modeling lessons free ($199 value)
or Unlock with your social account...
Reprehenderit id magni et omnis quibusdam voluptas laboriosam hic. Consequatur quae quia sit consequatur.
Ex qui at sint sed neque. Harum provident architecto sint doloribus. Ut sequi enim sit fugiat. Nam ut praesentium vel eaque dolor sed. Est blanditiis ullam eum sequi odio quia.
Est facilis aut consequatur architecto. Ut sed non temporibus. Voluptatem at omnis dolorem natus.