You're Not Ready For Self-Building AI

Today on Briefly AI. Anthropic says it's very close to AI that builds itself. Salesforce pays two point seven billion dollars for a call-centre robot specialist. And one famous investor admits his AI bill could buy a house.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. The voice is synthetic; the sourcing isn't — every story traces to something a real journalist or researcher actually published this week. Someone human still has to do the reporting. For now.

That's the shape of the day. Let's dig in.

Somewhere right now, an AI is quietly doing its own homework, and the grown-ups are starting to ask who's marking it. Right, let's get into it.

First up. According to TechCrunch, Anthropic has warned that it is very close to what the field calls RSI. That stands for recursive self-improvement, which is a mouthful for a simple idea. It's the point where an AI gets good enough at the work of building AI that it starts training and improving itself. Each version helps build a better next version, and the loop speeds up.

Here's why people care. Right now, progress in AI is paced by humans. Researchers have ideas, run experiments, wait for results, argue in meetings. If the machine can do that work itself, the pace is no longer set by human working hours. TechCrunch describes RSI as the point that spawns a dramatic acceleration in capabilities. And it notes that for the people who worry about this sort of thing, it's the cataclysm there's no coming back from.

Now, a bit of care here. What's actually new is who's saying it. For years this was a thought experiment from the worrying end of the field. Now the company building one of the leading models is saying, in its own words, that it's close. That's not a researcher on a podcast. That's a lab describing its own product.

It isn't the first time the idea has come up on this show. Anthropic published research on self-improving AI back in late August, and a self-improving model called Recursive Flint was in the news late last month. So this has been building. What's different now is the word "close."

What I can't tell you is how close. The reporting is a warning, not a date, and "close" from a company with a lot riding on its models being impressive is worth reading carefully. But as warnings go, it's a notable one. A company saying out loud that its own technology is approaching a point it flags as dangerous is rarer than you'd think.

On a rather more earthbound note. Salesforce has announced plans to buy Cresta, according to The Information, via Techmeme. The price is reported at two point seven billion dollars. Cresta is nine years old and builds AI software for contact centres. That's the call centre, the people and now increasingly the software who answer when you ring your bank, or your broadband provider, and say the word "representative" four times.

Cresta's pitch is automation for that world. Software that helps handle customer conversations, and increasingly handles them outright. The Information describes this as the latest stop in Salesforce's recent shopping spree, so it's one of several AI-flavoured purchases, not a one-off.

Why does it matter? Because customer service is where AI has gone from demo to daily reality fastest. It's high volume, it's repetitive, and the results are easy to measure. A company can see in a spreadsheet whether the robot resolved the call. That makes it the obvious place for the big business-software firms to plant a flag. Salesforce already sells the system that stores your customer records. Owning the thing that talks to those customers is a logical next layer.

And for the people on the other end of the phone, on both sides of it, this is the bit that touches real lives. Contact centres employ a great many people. A deal of this size says that Salesforce is betting that automation of that work is where the money is.

What actually came of it, so far, is an announcement. It's a plan to acquire, not a finished deal, and the reporting gives me the price and the intent, not the integration plan. What I can say is that two point seven billion dollars is a lot to pay for a thing that exists mainly to make "please hold" a bit less painful.

And finally, a human story, and a rather expensive one. TechCrunch reports that Garry Tan, who runs Y Combinator, the famous startup accelerator, says he spends up to two hundred thousand dollars a year on tokens. Tokens are the units AI companies charge by. Roughly speaking, they're chunks of words, and you pay for every chunk that goes in and every chunk that comes out. Run AI agents, which are programs that work through tasks on their own, all day, and the chunks add up fast.

Tan said this on Walter Isaacson's podcast, The Builders. He also jokingly admitted to what he called "cyber psychosis." He says he barely sleeps sometimes, because he's so excited to be working with his agents.

Now, it's one person, and a particularly enthusiastic one, so I wouldn't treat it as the average experience. But it's a useful data point on what heavy use actually looks like. The tools are good enough that some people find them compulsive. And the bill isn't a rounding error, it's a salary.

If you're wondering what an enthusiast is actually running up that bill on, the practitioners I follow are fairly clear. One developer said this weekend that nothing is writing better code than Claude Opus 5.5 inside Claude Code. Simon Willison, who writes about this a lot, said Opus 5.5 can now build video games that are actually fun. There are leaks about a Claude Opus 5.6 arriving soon, but that's an unconfirmed rumour, so treat it that way. Meanwhile, Willison's command-line tool now supports OpenAI's GPT-6.1 Sol and GPT-6.1 Terra, so the competition isn't standing still.

So the picture is this. The best tools are very good, they're being used obsessively by a few, and they cost real money. For the rest of us, the question is less whether we'd spend two hundred thousand dollars, and more whether we'd notice if we were spending two hundred a month without much to show for it.

Three stories, and between them a lab nervous about its own creations, a software giant buying the call centre, and a man who's stopped sleeping because the robots are too much fun. I'd say the robots are winning on enthusiasm, at least.

You can find more at harrysharman.com. Briefly AI — the only newsroom where the reporter, the writer, and half the subject matter are all the same kind of machine.