Notes

I often get asked whether I think AI is a bubble. Usually because I work in an AI-native company. I took Andrew Ng’s machine learning course back when he still used Octave, and started experimenting with agentic frameworks 2 years ago. My primary thought: this is a revolution and we don’t yet know what the effects will be. No, “AI” is not a bubble. But…

Yes, there is an AI bubble.

The bubble isn’t inherent to the technology; the allocation of capital is in line with the potential. The transformer revolution has irrevocably changed the scope of what we can use digital computers to do.

The capex spent on GPUs and data centers is questionable in cases, but we’re encountering some (healthy) limits on growth and I am optimistic we’ll make use of any spare computing power. We’re just scratching the surface of what’s possible with inference, and training will continue being critical for businesses even if open-weight models provide a capable enough foundation.

The bubble I see is opinions presented as wisdom.

AlexNet was almost 14 years ago, but most of the “AI experts” on LinkedIn didn’t notice the thaw even 9 years ago when the nerds realized what was becoming possible with natural language.

I don’t mean to be a graybeard that discounts newcomers, and I certainly don’t claim much wisdom myself. While most people only started paying attention when OpenAI showed that autocomplete was good enough to retire the Turing test, some began paying very close attention and experimenting. We can all benefit from following their critical thoughts on how machine learning and cheap, capable agents may shape the future and when.

What I do discount, and hope you will, is those who are happy to present with confidence the right way to deploy AI in products, the best methods to adapt workflows to agents, and predictions for who will win and lose in the uncertain future.

I’ll indulge myself in a prediction, though: this bubble will not pop suddenly. It may shrink, but only when a more attractive trend surfaces.

I wish more product teams were asking themselves: should we be offering the user more “AI”, or should we be ensuring our product works well when they are using agents alongside it or even as an intermediary?

The recent Dwarkesh Podcast episode about AI “learning on the job” has no news or discoveries but provokes some thoughts:

The true believers of AGI probably are also adherents to the Great Men of History view. Of course this view tends to correlate with seeing the PayPal Mafia boys as once-in-a-generation geniuses, meat-based AGI.

I find it more optimistic to see them as competent-enough nerds who happened into the right context to dent the universe. (For better or worse.) I could hold this POV because I’m just a less-advanced intelligence than them, but maybe it’s because I learned to appreciate the fundamental attribution error before my own ego was calcified against recognizing the value of context.

Anyway, context matters, perhaps more than anything! To animal and synthetic intelligence alike. For the foreseeable future, the efficacy of agents will be defined by our ability to give them the right context, not only wiring them to godlike artificial geniuses.

This is why I think the business model of firms like Asteroid is much more exciting than most wrappers. They have targeted a meaningful set of tools in a corner of industry that’s ripe to benefit, and put in the work to assemble and maintain the context needed for agents to become skilled users. Any firm in the healthcare/insurance space that puts in their own work bringing their data together can thus deploy those agents and meaningfully improve their operations today. Without waiting for some mythical god-like model that could “just figure it out” and save any of us from having to bother.

When I first started using coding agents: “Recent changes have caused an error rendering the template file /path/to/file.html. Please review the following error message and suggest necessary corrections. [relevant portion of stack trace]”

Today: “Oops! [lazy copy-paste of stack trace]”

More efficient circuits that require proteins grown with bacteria? I’m looking forward to seeing how we manage to connect bioreactors and chip fabs.

SpaceX is doing amazing work, and is full of engineers I really want to celebrate.

I would find the usual regulatory capture antics gross, but that’s not why I cannot enjoy their achievement. It’s that every success of these brilliant people is directly empowering their actively-fascist oligarch. And I despise him all the more for this.