Yogurt, agents, and AGI

I wake up this morning, eat my yogurt, and read the news. I come across Bloomberg: Altman has told employees that OpenAI is open to slowing the development of its most advanced models and hopes other labs will join in.

I finish reading and go see what my agents have been up to overnight.

One was supposed to write seed data for a few database tables. The source material was ready, with detailed descriptions of the entities and the relationships between them. The kind of task any junior developer could do in two hours.

Some of it was fine. Where it couldn’t figure things out, it made things up. Now I get to work out where the source data ends and its imagination begins.

Another was supposed to sort files into folders. This one had the opposite problem: it got unsure about something and stopped at around two in the morning. Nothing else happened until morning.

Then I started merging pull requests. Mostly acceptable, bad in places. Several rounds of corrections just to get the agent to follow the guidelines it had already been given.

And yet I have one agent that makes me happy every morning.

It keeps track of the achievements of dogs I’ve bred. The job is to find social media posts, download and lightly edit a photo, prepare a post from a template, and publish it. But first it has to figure out which of those updates deserve a post at all.

Major wins call for a proper celebration. Minor achievements just need a mention. Results from some shows shouldn’t go on the website at all. That means understanding what happened, how much a particular title means, and what it takes to earn it.

For a while, I hired someone to do this. Without a background in dog breeding and showing, they kept struggling even with detailed written rules. They had to keep all of it in mind and apply it consistently to every new post.

Current models handle this beautifully. Give the agent the basic instructions, the regulations, and the title requirements, and in the morning it shows me the posts it found over the past day, already sorted into categories. It hasn’t made a single classification error yet.

So an agent can make sense of dog titles and apply the rules consistently. And another agent can fail to write seed data from a ready-made specification, inventing the missing data along the way.

That unpredictability is what puzzles me most. Even the difficulty of a task doesn’t always tell me whether an agent will do better than a person or leave me with work to redo.

This is roughly how my mornings go. Sometimes I’m pleased with the results. Sometimes I’m angry about the tokens I’ve burned, mostly at myself. Sometimes there are no results at all.

And every time I think: God, we are still such a long way from AGI.

I work with these systems a lot. They are useful, or I wouldn’t fire them up again every evening. But I still can’t confidently hand over even a straightforward job and expect a correct result in the morning. Where an agent should ask, it invents. Where it could keep going, it stops and waits.

Then I read another piece about superintelligence and the urgent need to slow down.

I understand that an unreliable agent with access to systems can be dangerous. I can readily believe that problems with keeping agents under control warrant real pauses. But that doesn’t explain why we should consider a general intelligence capable of independently handling a wide range of tasks to be close.

When I say AGI still feels a long way off, I’m thinking of all this experience. Including the successes. It’s hard to believe it’s close when every evening I’m wondering what I’ll get tomorrow: excellent work, made-up work, or no work at all.

Looking at the generative transformer models available to me, I really don’t see future conquerors of the world.

Maybe the labs already have substantially different models or architectures. Maybe it comes down to resources, training, and the conditions they work in. Maybe we’re looking at the same results and drawing very different conclusions. Or maybe there’s a lot more hype in these statements than I’d like.

But then I’d like someone to explain that gap. What exactly are you seeing that the people trying to get ordinary work done with your product every day aren’t?

Because I’m still finishing my yogurt and checking whether the future superintelligence made up the data for three tables.