By: [email protected]

Published October 24, 2025 at 01:09 AM

AI Bubbles and Bears

My latest rambling about the state of AI in late 2025 as the tide starts to turn a bit. Part 1 of this new “State of AI (Late 2025)” series.

Intro

Continuing to treat this blog as a glorified journal, I thought I’d write a post at what is really starting to look and feel like the peak of the AI bubble.

If you’re reading this blog odds are this isn’t a new story to you.

Over the past few months, now years, AI has continually improved and impressed on numerous fronts. From ‘reasoning’ to multi modal image / video generation even to pretty significant performance improvements even for the smaller models.

Despite that, over the past few months chatter of a ‘bubble’ has only increased and for good reason;

  • Lot’s of the recent frontier models have began to feel like incremental changes if at all - GPT 5’s release is a great example
  • Increasingly lofty promises from AI / Big Tech Executives often padded by defensive language
  • Incomprehensible amounts of (often cyclic) investment in massive data centers that leave many curious how we will ever meet this capacity

This all got me thinking - if we are in a bubble and just about everyone seems ready to admit it, what happens now? When does it pop and what does that look like? And maybe most importantly - what happens to “AI” after this?

In this post (and maybe another who knows) I want to explore the current state of AI (in late 2025) and make an argument that we are almost certainly in a bubble, how and why that happened and how I am looking at the long term future here

State of AI (Late 2025)

I wrote a whole 3 part series on the state of AI earlier this year, now it just gets this section in this one post so take that as you will. - For example this initial post about agents earlier this year - admittedly a lot has changed even since this.

I will say while this isn’t the first tech hype cycle I’ve lived through, even as an adult (see the numerous Cryptocurrency bull runs) it’s unquestionably the biggest I’ve experienced. It’s truly been fascinating to watch this play out in the world and even within my own life and career.

While I typically avoid making actual predictions, especially in public written form, I think now is an important time to talk about the state of things and where it looks like things are headed. Admittedly this is half for myself.

All to say - I do think we are in some form of a bubble. I think we have definitely overinvested and that these companies - especially outside of OpenAI - will almost certainly not live up to these insane valuations.

That said I also think AI (gen AI really) is here for good. You will almost certainly still be able to talk to your best pal GPT 9.32 Max or Claude Haiku Opus Magnum in 2028.

But Why?

Human nature I guess?

I think while lots of hype and overpromise I think has resulted in some ‘frothy’ valuations - I don’t think overhyping tech is some new phenomenon. Again maybe the scale itself is and there’s some interesting reading / opinions there but we’ve overbuilt entire industries of tech for centuries now (see the railroad bubble in the 19th century.)

Really I think the main driving force here is those in charge - Microsoft, Google, Amazon, etc. - see a huge opportunity here and simply the cost of missing out is WAY larger than the prospect of over investing into a bubble.

I think they view this a lot like the .com boom - where there were thousands of companies and investments and only a few survived. Those that did though, did really, really well and clearly the Internet has panned out (from a financial POV at least.)

Moreover, there was AI / machine learning before - this tech has real merit and use. Will every piece of software be behind an LLM / natural language interface? It’s looking increasingly like no, but that doesn’t mean (generative) AI won’t have a place in the world.

“When will I use this in the real world?”

Now that’s all fun and interesting but I’m not here to convince you we’re in a bubble, that’s not really my place and there’s many, many blog posts that will make that argument far better than I can.

I’m gonna make an argument here that while the above is correct, but focus more on that I think we’re looking at the question the wrong way.

The million dollar question here - is this stuff actually making people money?

The long and short of it seems to be “maybe”:

Overall the studies and reviews are mixed honestly - lots of propoganda from both sides you can find any narrative you want at this point.

To me though that shows that it’s definitely not super clear right now. That will almost certainly change at some point but I find that interesting.

Software Development

I think an interesting point here is how gen AI has worked with software development.

I’ve talked about this a lot so I won’t get too much into it but my thoughts here are basically that the way these models work is kinda perfect for software:

  • Tons of public code to train on
  • Software can be validated automatically via tests
  • Lots of common use cases and patterns - again often with public solutions
  • Everyone hates software peeps - what a good group to automate away or at least make cheaper

Not to mention these models are literally made by software engineers who know these problems very well.

Personally I find these tools helpful, they allow me to get some decent work done when I’m half paying attention which can be nice particularly in the evening. They’ve helped save me some Google searches without a doubt. I do not trust them to run a bunch of commands or operate unsupervised though.

That said there is increasing evidence that in the long run these aren’t actually saving developers time, primarily due to the slop they put out. I’ve definitely experienced this personally - you push the model more and more until it starts producing code that’s unworkable then you need to step back.

I’ve seen this study referenced a bunch online and even at work. TLDR software engineers using AI think they’re moving faster but are actually slower.

Again, I can see how this happens I didn’t think it’d be so bad but the numbers are there if you’re interested.

Will this change over time? Probably. But I’d argue this is kinda the best stomping grounds for these models and it’s clear they’re not gonna achieve the full automation dream quite yet.

No Money No Problems?

It’s a good point to remember what the point of this is - we want to reduce labor to save money or at the very least let people focus on harder problems.

The base assumption there is that AI will cost less to get to the same solution.

I.e., it will either be cheaper and run longer or be more expensive but quicker when compared to a human.

This does not seem to be panning out.

First off - all the frontier models are being heavily subsidized by the companies running them (OpenAI, Google, Anthropic) as they want to increase usage.

On top of that, with all the “reasoning” models usage is spiking and token usage is as well. Both of which are often the biggest factors in determining prices.

We can already see this playing out - even in software Cursor the IDE had to start charging for usage instead of a flat fee, Claude Code had to implement stringent rate limits and we’re seeing similar patterns across the industry.

I don’t think this means that these products are doomed, but I do think this means we’re gonna see:

  1. Diminishing returns on the increasing costs as again the models aren’t doing much better
  2. “Reasoning” (i.e., running stuff through the model a lot) will likely also see diminishing returns and we’ll have to find better ways to improve results
  3. We’ll likely see some people drop off these products as they get more expensive

All the data centers and investment might help but that’s pretty downstream - who knows how expensive it’ll be to run these models by then but these are years in terms of timeline.

Again, hard to say exactly how this plays out but it’s not looking like this tech is gonna get cheaper and better in the near future.

Conclusion

So is the dream dead?

I’d say no - again I compare this to the .com bubble. This tech is real, it will stick around, it seems to be a natural part of investment cycles that there will be winners and losers and this is a BIG cycle.

Which again raises the natural question - what do we do next? If this is a bubble, what can we actually do about it?

That’ll be the next post!

Originally this was just going to be one part but I am yapping too much to keep this short and pleasant.

Again, I hope by this point I’ve convinced you that you should exhibit some level of skepticism when Sam Altman is promising we won’t be working in 3 years and even day care workers will be AI by next school year.

To be clear too - I’m not a hater. I use this tech a lot and blow through my Claude Code quota nearly every day. I think these are useful tools but very much just tools in their current form and it’s looking less and less likely that there’s another radical jump in the near future.

Next time, we’ll go into what my longer term horizon looks like for this technology and what you can do today to stay competitive and quick on your feet.

What do you think? Am I totally off base? Has Claude fixed your code base, love life and finances and I’m just some voice on the internet? Would love to hear your thoughts in the comments!

Thanks for reading!

Comments

No comments available :/

Leave your thoughts! 💭


Tags

Agents

Gen AI

State of AI (Late 2025)

Say hi!