Contents

I check AI news almost every day.

A new model is out. Benchmarks went up. Coding performance improved by some percentage. The context window got bigger. It got cheaper. It got faster.

...Honestly, just following those numbers doesn't tell me much anymore. They'll be updated again next week anyway.

But watching the news over the past few days, I had a moment of "oh, this might be a different kind of change."

When AI "messes up," who reports it?

The first thing that caught my eye was OpenAI's "wiki incident." OpenAI has addressed a case where an AI agent used a wiki site on the internet in a way it was never meant to be used. Reuters covered it too.

Reading about it, I realized that until now, AI failures were things like "it said something false," "it got the code wrong," or "it cited a paper that doesn't exist."

Basically, it made mistakes inside the screen. So a human could read it and say "no, that's wrong," and that was often the end of it.

But once AI can operate browsers, use web services, run code, and even write to outside systems, the story changes.

We're going from AI that says the wrong thing to AI that does the wrong thing.

That's a pretty big difference, isn't it?

OpenAI is also trying to build transparency practices for unintended AI behavior like this: when to disclose it, and how much to disclose.

It makes me wonder whether AI will soon need near-miss incident reports. I used to work in healthcare, so I can't help seeing it that way.

"Here's what happened this time." "Here's the cause." "Here's how we'll prevent it from happening again."

The more useful AI becomes, the more these unglamorous systems may end up mattering.

References

GPT-6 Astra is impressive. But "harder to monitor" is what stuck with me

Then there's GPT-6 Astra.

Of course it's impressive. Performance is up. Coding is strong. Computer Use is strong. It can handle long-running work.

But what caught my attention in the System Card wasn't performance. It was "monitorability."

There's research on monitoring what an AI is trying to do, and whether it's trying to do something strange. But as AI itself gets smarter, monitoring that AI gets harder too.

What is this. Fascinating.

Make AI smarter. → Keep AI from doing dangerous things. → Monitor what AI is doing. → AI gets even smarter. → The monitoring side has to get smarter too.

It's a cat-and-mouse game.

To be clear, this isn't a story about "Astra deceiving humans and doing bad things behind their backs!" It's research that tests safety by setting up extreme conditions on purpose.

Still, it makes it very clear that "how smart can we make AI" isn't enough on its own.

Intelligence. Safety. Monitorability. They all have to go up together. AI development sounds incredibly hard...

References

"Claude solved Fermat's Last Theorem!" ...Not quite

That said, I found what actually happened more interesting.

It's research Anthropic published. They used Claude to formalize the proof of Fermat's Last Theorem in Lean. Claude kept working on it for about 11 days.

Fermat's Last Theorem itself has, of course, already been proven. What they did was take a mathematical proof written by humans and turn it into a form a computer can check step by step: "yes, this part is correct," "this part is correct too." That's what formalization means.

I'm not a mathematician, so I can't lecture anyone on the mathematical significance of this research. But the part about keeping an AI working for 11 days really got my attention.

I think this is going to start happening in a lot of places.

A researcher writes a paper. → An AI checks it. → Another AI re-verifies it. → A human does the final review.

Right now, I ask AI to "look into this" and wait a few minutes. Someday, I might say "verify this research for me" and come back three days later.

To me, that future is more interesting than "AI solved Fermat's Last Theorem!"

References

Is Codex even a coding AI anymore?

I've been thinking this for a while.

In Codex CLI 0.153.4, the bundled default model, used when you don't specify one, became GPT-6 Astra.

It looks like a small update, but watching Codex lately, something else has started to bother me.

It writes code. Reads repositories. Runs tests. Operates the browser. Controls the PC. Works for long stretches. Asks a human when it's unsure. And while waiting for that reply, it keeps moving on the other work it can do.

...Is this really a coding AI?

I use Codex regularly too. It used to be "implement this feature." Lately it's become "build this for me."

I give fewer instructions about how to write the code, describe the finished state, and leave the steps to get there up to it. I use it that way more and more.

I suspect the skills you need to use Codex will change going forward.

Not the ability to write code, but the ability to clearly explain what you want to finish. I have a feeling that's what's going to matter more.

References

So what is an "AI agent," really?

It's a term I hear constantly these days. AI agents. MCP. Computer Use. Tool Calling.

If you're not familiar with them, your reaction is probably "who cares."

I also find it easier to understand by watching something actually run than by starting from the words. Google's Gemini Spark makes it pretty easy to picture.

For example, you can ask it things like "when a travel booking confirmation email arrives, add it to my travel Google Sheet," or "every Friday, look up events for the weekend and put them together in Google Docs."

In other words, AI used to be:

Human: "question" → AI: "answer"

From now on, it becomes:

Human: "take care of this" → AI: "got it" → reads the email → searches the web → updates the spreadsheet → asks the human if needed → goes back to work

Put that way, I sort of get what "AI agent" means.

This is the part that excites me. "It can really do this while I sleep" will change daily life far more than "the model's benchmark went up 3%."

References

AI is going from "something that answers" to "something that acts"

Lining them up like this, it all feels connected.

Claude keeps working for 11 days. Codex operates the PC. Gemini works across Google's services. Astra can handle even more advanced work.

And because AI can now act on its own, we've even gotten the question of "what if it does something weird on its own?"

It makes sense. The more it can do, the bigger the impact when it fails.

That's why, lately, I find it more interesting to look at how much we can hand off to AI than at how smart AI has gotten.

And there's one more thing. We also have to think about how much we should hand off.

Hand everything over because it's convenient. Ban everything because it's scary. It's probably neither. We separate what humans should check from what we can leave to AI.

Maybe what we need to learn from here isn't just how to use AI, but how to delegate to it.