It Couldn't Do Maths. Now It's Launching Itself.

Format: Podcast debate Hosts: Host 1 (The Threshold) — the jumps already happened; the line is that the thing is now inside the loop of making the next one. Host 2 (The Measured Skeptic) — still a tool with a human on the handle; "building itself" is a headline that needs the small print.

Host 1: Four years ago you could sit and watch one of these things fail a sums question a ten-year-old would get right. Not a trick question. Seven times eight. People used that as the proof it was a toy. If it can't add up, it can't take anything seriously.

Host 2: And a lot of people still have that picture. They tried ChatGPT in 2023, it made something up with a straight face, and they walked away. Parlour trick. That memory is doing a lot of work.

Host 1: That memory is two years out of date, and two years in this field is ancient. The jumps aren't a feeling. They're a list. It couldn't do maths. Then it passed the bar exam. Then it started writing real software — not a suggested line, a job described in ordinary English, left running, come back to something that actually works. Then it opened the thing it had built, clicked the buttons, found what was broken, and fixed it. That's not a chatbot answering a question. That's a loop.

Host 2: Coding is the friendliest version of this argument. Rules, tests, either it compiles or it doesn't. A courtroom, a tax return, a conversation with someone who's frightened — that's a different kind of wrong when it misses. People I know in law still catch it on the specifics.

Host 1: Of course they do. I'm not saying it's finished. I'm saying look at the slope. If it shows a hint of a capability this year, the next version is good at it. Maths was the hint. The exam was the next one. Software was the one after that. The bit that isn't a parlour trick anymore is what the labs put in writing this year: the model is now inside its own build.

Host 2: "Building itself" is a headline. What did they actually say?

Host 1: OpenAI, in their own words, on the coding model that sits in the ChatGPT family: first model that was instrumental in creating itself. The team used early versions to debug its own training — that's the build. To diagnose its own tests. And to manage its own deployment — that's the launch. They used it to scale the computers that served it when the traffic hit. When people say ChatGPT is building itself, debugging itself, and launching itself, that is the thing they mean. Not a film. A post from the company that made it.

Host 2: And in the same paperwork they said it does not autonomously redesign itself. It can't start its own training run. It can't rewrite its own brain. A human still points it at the problem. That's a power tool on a bench, not a workshop with no one in it.

Host 1: I'm glad you said that, because that's the honest version of the line. Nobody serious is claiming it wandered off and invented the next ChatGPT in a cupboard. The line is this: the people building the next one are using this one to build it, debug it, and ship it. Anthropic's boss said the same shape the same week — Claude helping design the next Claude, not in every way, but in many ways, and that loop closes fast. Once the thing is in the loop of making the next thing, you don't get to keep the 2023 picture. That's the picture that stops people updating.

Host 2: So the danger isn't a robot waking up. It's that people are still judging this on a free-tier chat they tried two summers ago.

Host 1: Exactly. Judging what's running now by that is like judging a smartphone by a flip phone and calling the case closed. The free versions sit a long way behind what the paying side is using. If your last look was "it can't do maths" or "it made a case up," you haven't looked.

Host 2: Fine. Then what do you actually want someone to do with that?

Host 1: Sit with the list. Couldn't do maths. Then it could. Then the exam. Then the software. Then it clicks through its own work. Now it's in the room where the next one gets built, debugged, and launched. You don't need a forecast from me. You need the list to be in date. Spend time with the current thing, on real work, until the 2023 picture breaks. That's the whole argument.

Host 2: And if the next jump is slower than the last four?

Host 1: Then you've lost an hour a day getting good at a tool that already writes, checks, and ships. If it isn't slower — and it hasn't been — you're not the person still explaining that it can't do maths.