Paying Attention #3: Physics is All You Need?, Codex as Meta-Agent and Big Week Ahead
I write this as I sit in the Atlanta airport awaiting dinner, drinking a Heineken 0.0%, before I hop a flight to London. I head there to meet with a client this week. Yes, the client will use our AI agent.
I really love international travel. When I was 27, I flew with only a backpack to Dublin. Over the course of 3.5 months I made it as far east as Belgrade. I’ll tell you more about that some day.
This is my first international trip as a Dad. I am excited about it, but I know it’s hard on my wife. She’s shouldering the full parenting burden. I will make the most of it!
1. Physics is All You Need?
What happened
A physicist used Claude Code for 12 days as he worked on a cosmology model. Claude was quite useful when it came to the mechanical elements of building the model: writing equations, debugging and optimizing against the test suite. Claude even correctly calculated the right metric. The problem is Claude had no idea why it had arrived at the correct metric.
Why it held my attention
It’s the perfect demonstration of what today’s AI systems can and can’t do. I’ve thought a lot lately about the distinction between intelligence and intellect. To me, intellect is a facet of intelligence. It’s the ability to carry out computations. Computers have obviously been capable of performing computations for a while. The difference for AIs is that they can do it in the realm of ambiguous natural language too. They can do qualitative and not just quantitative computations.
AI is really artificial intellect rather than intelligence. The humans needs to supply the intelligence. I’m working on a longer essay on this topic.
What I’m carrying forward
The fact AI may end up being a superhuman intellect, but not necessarily a superhuman intelligence, is a very good thing. It means humans will be able to work in partnership with this powerful force we are unlocking and not necessarily be replaced. It means AI will have its thing it’s good at, and humans will have ours.
2. Codex as Meta-Agent
What happened
There was a Codex feature that went unannounced. Codex can now spin up its own threads and worktrees. It allows Codex to orchestrate itself, essentially working as a meta-agent controlling its own agents.
Why it held my attention
Engineers from the top AI labs are talking about a new type of workflow. They say they rarely prompt the agent now. What they do is have an orchestrator that they give instructions to, and the orchestrator prompts the agent.
What I’m carrying forward
It’s unclear if you need to have unlimited tokens because you work at a frontier lab to make this work. I imagine you don’t. A ChatGPT Pro sub is probably enough for now.
3. Big Week Ahead
What happened
The rumor mill is churning. There’s talk that Claude Mythos, GPT-5.6 and Gemini 3.5 Pro could be released this week. It’s the potential for a sort of AI trifecta. I still feel like a kid on Christmas when this happens. If the kid was a giant nerd.
Why it held my attention
I believe it will be a real inflection point. If either Mythos or GPT-5.6 are the clear leader, it will be the first time Anthropic or OpenAI had a clearly more capable model since probably o1. It feels like things are really ramping up, and it always feels like that. The infinite ramp up to the Singularity.
What I’m carrying forward
I hope I have a lot to write about in next week’s edition of Paying Attention. Let’s see what happens this week. It has the potential to be a big one.