Something is always missing when you talk to an AI. It writes like a person and argues like one, but the conversation still feels off. For a long time I couldn't say what it was.
Last summer I argued about this with a coworker. I said I want my agent to write my messages, and I want the other person's agent to read them and pass on what matters. They thought that was a bad idea. Their point was that human conversation carries a signal an AI can't pass on, and conveying that signal is a real part of the job.
I think they're right that the signal exists. I disagree about whether we should keep it.
Status
Impro by Keith Johnstone is a book about improv theater. It gets read at Palantir because the FDE job is mostly walking into someone else's problem, pulling them into yours, and getting them to act on it while looking confident.
The part that stuck with me is status. Johnstone argues that every exchange between two people raises or lowers one of them. It isn't social status. A junior can play high and a CEO can play low. It's something you do constantly, mostly without noticing, and it usually ends up close to even with small shifts back and forth.
Once you see it, it's everywhere. Every meeting ends with someone being the person who decides, and that isn't always the person with the best idea. A good test: is the CEO the smartest person in the company? Usually not. They played their cards well.
How visible the game is depends on the people. At some companies almost nobody seems to notice it happening, so the best idea rarely wins. At Palantir people are smart enough to know what they're doing when they talk, and they do it on purpose. Either way, the game decides.
The people who are very good at this treat communicating an idea as part of the problem. What they say matters less than how they move status while saying it. Over a long career that wins. They get promoted and they go far.
They're also the people who hate the idea of agents talking to agents, because you can't play status through one. Whatever the sender's agent writes, the reader's agent pulls out the facts and drops the rest.
So that's what I think is missing when you talk to an AI. There is no status game to play. I don't see that as a flaw.
HAAH
I call it HAAH: human to AI to AI to human. I tell my agent what I want to say and it writes the message. The other person's agent reads it and gives them the parts that matter to them.
The usual objection is that this is brain rot. You make an AI formalize a message just so another AI can deformalize it again, adding tokens for the sake of tokens.
That assumes both sides do the same thing. They don't. If you manage your context properly, your agent knows you better than anyone. On the way out it knows which facts I need to get across. On the way in it knows which facts matter to you. The message in the middle is just transport.
The colony
The way I'd want people to work is closer to an ant colony. Ants coordinate through pheromone trails, and no ant has to convince another one. Only the outcome matters. The best idea wins, and it doesn't matter who had it or how well they performed it.
This only holds for factual work. In entertainment the person is the point, and you'd never want an agent in between. But most of what we say at work isn't entertainment. It's ideas that should be judged on what they are.
Midway
HAAH is a middle stage. Models can't come up with the best ideas yet, but they're already good enough to carry ideas between people. If they end up having the best ideas too, humans aren't needed on the factual side at all.
Until then, talk to my agent.
Feedback
Disagree with something, spot a mistake, or just liked it? Tell me. Anonymous unless you leave an email.