I think it’s pretty obvious that we’re in a bubble. Almost no one rational denies this. In fact, I think even Sam Altman and Dario Amodei would acknowledge it. They’re just very good at using it to their advantage. And when the bubble pops, they won’t be the ones closing shop — they’ll have built the next Google or AWS.
At the same time, there are people who don’t understand how computers work and have made it their shtick to swear loudly and proclaim that LLMs have no redeeming qualities at all.
Personally, I subscribe to a variant of Karpathy’s and Sutton’s worldview. LLMs are wonderful, and we’re all enjoying the “$3 Uber rides” phase — but we also have a long way to go. A hill taller than most checkbooks, including Masayoshi Son’s. Do you really think Microsoft would pull back on CapEx if they believed all knowledge work would be automated by 2030? And why would OpenAI build an AI-powered recruiting platform for pesky humans?
I think there are currently only three and a half categories of AI products with true product–market fit (PMF):
Software development is forever changed, and this is true across the skill spectrum. Most highly motivated people can now prototype with AI — a much bigger deal than most realize. It means we can build entirely new platforms that turn these prototypes into something substantial. The next generation of tools and infrastructure will meet the needs of the vibe coders. In fact, as much as Vercel wants to call itself the “AI Cloud,” it’s really the vibe coding cloud. Lovable and Replit are similarly inclined.
On the elite developer side, we saw more initial skepticism. That makes sense: they’re more demanding and have more of their personal identity tied to “being a good developer.” We saw similar headwinds with Heroku and ops teams before DevOps became standard practice. But, when folks like Mitchell Hashimoto and Armin Ronacher find great utility in coding agents, you don’t have much ground to stand on.
Most importantly, I don’t see why models won’t keep getting better at coding. Coding is inherently more verifiable than most other domains. And, much of it fits perfectly into the “monkey see, monkey do” pattern of model learning. Coding is — and will remain — the single biggest and most valuable category where LLMs have PMF.
Not just in the explicit sense (though that too), but as a synonym for products whose sole purpose is to deliver the maximum amount of dopamine possible. Virtually all of Meta’s attempts, as well as Sora, Midjourney, and Character.ai, fit in this bucket — entertainment for a WALL-E-style sedated life. AI will get so good at this that it will make heroin look like Skittles. Social media with a heat-seeking missile: endlessly tailored content just for your amygdala. I personally don’t want anything to do with it, but it’ll obviously be huge.
It’s no secret that ChatGPT traffic drops when kids are on summer break. Well, guess what — the grown-ups cheat too. A lawyer asking ChatGPT to write a legal brief without reviewing the citations is just cheating on homework at work. For a lot of folks in tech, it’s hard to grasp that the vast majority of people don’t live to work — they work to live. Sure, they want to get promoted and have decent careers. But at the end of the day, it’s mostly about keeping the boss happy and not getting fired. ChatGPT is a godsend.
Malleable software has always been a dream for people who deeply understand computers. Many have tried and failed miserably to bring this worldview to the masses. LLMs finally make it possible to be a bit fuzzier with computer I/O. This isn’t about autonomy — it’s about making those pesky input fields behave the way most people always assumed computers should. That doesn’t require infinite scaling laws, just small, good models that are blazing fast.
Agents that autonomously, without human supervision or approval, perform lots of economically viable tasks are not close. And ironically, the moment they arrive, they’ll cease to be economically valuable — because I don’t believe in a fast takeoff where one lab races ahead. If I had to guess, we’ll see several open-weight models by next summer with “enough nines” of reliability that, unless you can add another one, it won’t make a difference. And as Karpathy said, adding a nine of reliability takes the same effort every time.
I have to pinch myself regularly that I get to work on software during this time in human history. What are the chances? Fucking brilliant. But just as the iPhone 1 and iPhone 17 are fundamentally the same kind of "thing," this will also be true for ChatGPT 7 and 8. Amazing technology, yes — but I’m willing to bet that this time, too, the phrase “this time it’s different” will age poorly.