Many AIs, Many Humans: What Sam Altman Got Right on Cheeky Pint (and Where I'd Push Back)

Where we agree: the global software upgrade is real

The host asked a question I have been chewing on for a long time: "What do you think the global software upgrade looks like?" The host's example was blunt: is a PHP code base just a bad idea, and should you "rewrite a nice clean version in Rust"?

Altman's answer: he has seen new startups saying they're going to "make a whole new stack," and "that would have seemed insane two years ago." You would have said there is too much to write. "And now maybe soon you can."

This one is personal. Koru, our open source workspace, is built from scratch in Rust, with Loro CRDTs for data and iroh for peer-to-peer sync. It's native, not Electron, and runs on Windows and macOS, with Linux available as a nightly build. A few years ago, starting a new workspace app from an empty repo would have looked like a bad bet against a mountain of existing code. Today, AI is a big part of why we can even try. The tools changed the math on what's worth rebuilding.

More on that in how Koru is built.

A trillion AIs, not one giant brain

The line from the episode I keep coming back to, from Sam Altman:

"I would much rather intuitively have a trillion AIs of like slightly superhuman intelligence join society than one AI of one trillion IQ join society."

The reasoning is what makes it stick. He points out that "for all of the deep flaws of society, it is unbelievably robust," and that this robustness comes from lots of people keeping each other in check and collaborating. He had been sad that AI might trend away from that pluralism, and now he is excited about swarms of agents as a direction. The host summed it up as many agents pushing against each other in a dynamic equilibrium, and the trillion AIs line was the answer.

That's the world I want to build tools for. Not one central brain that everyone routes through, but many agents working alongside many people, each with their own context, checking each other's work. That future needs shared spaces where humans and agents can both read, write and see what the other did, and that is the direction we want Koru to grow in. More on the idea in What is multiplayer AI?

Personal agents are still waiting for their Macintosh moment

The host asked what will be true of the winning personal agent. Altman started with "intelligence, context, tools," said no one has figured it out yet, and that it is "a small number of ideas away from a super mainstream thing."

The host then asked whether personal agents have had their "Macintosh moment" yet. Altman's answer: "I think we have not." He expects the winner to look a lot like today's agents, except "it'll just work."

Agreed. Of those three, intelligence is improving fast and tools are getting standardized. Context is the hard part, and for most people the best context they have is already sitting in their notes: projects, decisions, meeting notes, half-finished plans. That's why Koru has a built-in MCP server, so the AI you choose can work with your workspace instead of starting from a blank prompt every time. See the MCP docs.

Ending fake work

The most relatable stretch of the episode was the host's web form rant: when you fill out yet another form, you're almost never "inventing new information for a web form."

Altman jumped in with my favorite line of the whole conversation:

"I would be so happy to never do like fake work at a computer again."

The host's conclusion: "any web form is basically fake." Altman agreed.

Same. Retyping your address, re-explaining your project to a new tool, copying the same context into five places. None of that is thinking. If agents take that off our plates, the hours go back into real work. For me, that means building.

Where I'd push back: local inference

Here is the part where I see it differently, and I want to be fair about it.

Altman argues that from "a global allocation perspective," it is "quite terrible" to have everyone holding a chunk of the world's limited memory supply in their own computer, "being used 5% of the time." You want maximum utilization in the cloud, especially while supply is tight.

On utilization, that is a fair point. Hardware on a desk sitting idle most of the day is not efficient, and in a supply-constrained world that matters.

But efficiency isn't the only thing people are optimizing for. For private notes, control matters more. Your notes hold your plans, your health questions, your half-formed ideas, your company's strategy. Plenty of people and companies will happily trade some efficiency to decide which model sees that, where it runs, and whether they need a connection at all to keep working.

That's why Koru is open source, works offline, and works with the AI you choose. If the cloud is the right call for you, great. If it isn't, you shouldn't have to give up your workspace to make a different choice. Both can be true: the cloud can be the efficient default for the world, and you can still keep the final say over your own notes.

How this ties to what we're building

Put the episode together and you get a picture: software gets rebuilt, AIs multiply instead of consolidating, personal agents get good enough to just work, and fake work starts to disappear.

In that world, the most important question is where humans and agents meet. Our vision for Koru is an open source workspace that works offline, where humans and AI agents collaborate on the same pages, with you in control of your data and your choice of model. That collaboration is the direction we are building toward, not something I am claiming is finished today.

What exists today: Koru is free and open source under MPL-2.0. It works offline, syncs peer to peer between your own devices for free, and includes a built-in MCP server. It runs natively on Windows and macOS, Linux is available as a nightly build, and mobile apps are coming soon. If you want an always-on copy, optional Koru Cloud sync is $9.99 a month for 100 GB.

A trillion AIs joining society will need somewhere to work with the people they are helping. I would like that place to be open, to work offline, and to belong to the people using it.

If you have thoughts on any of this, especially the local inference debate, I'd love to hear them. Watch the full Cheeky Pint episode, or download Koru, the open source workspace that works offline.

Frequently asked questions