I made myself a podcast about which kind of startup makes sense now, in the AI era, when you build a startup and tomorrow a large language model has been taught to do the same thing.
So here is the gist. What I see now, having walked around various conferences in San Francisco, is that value is moving into the interface.
Meaning that soon all AI models will be identical and similar to each other. The volume of knowledge in the world is finite, and in a few years all the knowledge in the world will have been pumped out. But I mean the mass market right now, not narrow vertical models for medicine or law and so on.
And the mass market is just a language model for the average job, for an ordinary housewife, for an ordinary coder.
So all these models, Gemini, Codex and Claude, will hold one and the same set of knowledge. Accordingly, their value and their ability to reason will be roughly the same. Give or take, average across the board plus or minus 5 to 10 percent.
What happens after that? And what, it seems to me, is worth doing now?
Right now I have seen a big boom in harnesses. A harness is when you have several models at once inside some interface. And it seems to me that the harnesses which win popularity have the ability to compete with the models.
Take Perplexity, which I think you can call a harness, because essentially it is an aggregator of LLMs with a convenient interface. I do not use Perplexity much, because it is not very convenient for me, but for those users who need several LLMs in one bottle, for that housewife who needs to have several LLMs, it is a good model.
I am talking more about coding now, but not coding in the sense of you writing code, rather in the sense that any research work and so on, agentic work, is done through things like Claude Code or Codex.
So here is what I see: a lot of harnesses will appear with large numbers of users.
And accordingly there is no sense in using an inconvenient interface from Claude Code when I need to work with three LLMs.
The only question is that many LLM vendors will not be ready for their product to be inserted into somebody else's interface, and will ban you for it. Remember the OpenClaw example, where OpenClaw supposedly banned its own users and so on.
But by the way, I went up to people at an Anthropic gathering just now, they had a Founder House, at Terra, a kind of coworking place in San Francisco. So, walking up to 5 to 7 Anthropic employees who work in Applied AI, I asked them: I have about 4 subscriptions there, are you going to ban me? They said: no, dude.
They all laughed and said: dude, come on, until you have a thousand subscriptions you can relax. I said: and who needs a thousand subscriptions? They said: dude, the Chinese. Of course the Chinese buy a thousand accounts and do model distillation.
I do not know how model distillation works technically, but mostly they only block the Chinese who buy a thousand accounts and through those thousand accounts try to somehow distil their brains. No idea at all how that works, honestly. Well, presumably it works somehow.
I have heard the word distillation many times. Heard that word many times, distillation. But for me distillation is when I look through my own sessions.
Anyway, the point is that nobody gets banned with 4 accounts.
Again, some of my friends on Facebook will say they got banned, but maybe you just gave away that you are in a sanctioned country, maybe that is why. Otherwise, it seems to me, you can relax.
And when I say that I shove my account into various harnesses, they say that OpenClaw was some kind of special case where Anthropic got very angry. But in general they say: shove your subscription account wherever you like and do not worry, dude. Honestly, they say, do not worry. We have better things to do right now than block guys who put their account into some harness.
So there you go.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261022.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261022.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)