Episode — 2026-04-07

Jamie Dimon — the man who runs the world's biggest bank — says AI is about to hand you back a day and a half every week. Great. I'd settle for Tuesday afternoons, personally.

Right, let's get into it. Dimon published his annual letter to JPMorgan shareholders yesterday, and it's a good one if you like that specific flavour of Wall Street optimism that comes wrapped in a mild warning. The headline prediction: within 30 years, the working week in the developed world will shrink to three and a half days — driven almost entirely by AI productivity gains. He also threw in curing cancers and people living to a hundred, just for good measure.

Now, Dimon isn't a starry-eyed tech idealist. He's openly acknowledged AI will kill certain jobs — he's said so before. But his read here is the longer arc: short-term disruption, yes, but long-term abundance. More time, better health, more meaningful work.

Here's the thing, though — 30 years is a long time. That's roughly how long we've been waiting for self-driving cars to be "five years away." The direction might be right. The timeline? Your guess is as good as his.

What to watch: whether the short-term job displacement starts outpacing the long-term uplift story that executives like Dimon are selling. If those timelines diverge too much, the optimism gets harder to sustain.

Meanwhile, over at Meta, something quietly significant is happening. The company has long been the open-source good guy of frontier AI — releasing their Llama models for anyone to use, modify, and build on. That posture made them popular with developers and a useful counterweight to the closed shops at OpenAI and Google.

Well. That might be changing. Meta has a new model in development, codenamed "Avocado" — which I accept is the most Meta name possible — built under Scale AI's Alexandr Wang, who joined to lead their AGI effort. And according to reports this week, when it lands, it'll launch as a closed model first. Open-source versions may follow, but they're not the lead.

Why? Partly safety concerns, partly competitive pressure. Because Avocado, reportedly, is not currently beating Google and OpenAI on the things that matter — coding, reasoning, writing. Meta is trying to catch up, and apparently they'd rather do that quietly, behind closed doors, than publicly release something that loses.

The bit that matters: if Meta abandons the open-source lane, that lane gets significantly less crowded. And for anyone who's built workflows or products on Llama models, it's worth paying attention to whether the direction of travel just shifted.

And then there's the plumbing. This one's less flashy but arguably more important. Broadcom announced today that it's signed a long-term deal with Google to develop and supply custom AI chips — specifically Google's TPUs, which are their alternative to Nvidia's GPUs — all the way through to 2031. And in a related deal, Anthropic gets access to 3.5 gigawatts of AI computing capacity drawing on those same chips, starting in 2027.

A couple of numbers worth sitting with. Anthropic's run-rate revenue has gone from nine billion dollars at the end of last year to thirty billion now. That's not gradual growth. That's almost vertical. Claude is clearly doing something right.

The deeper story here is that the chip layer is where the real infrastructure wars are being fought. Nvidia still dominates, but Google's TPUs are becoming a serious alternative. And when you have Anthropic — a company running one of the most in-demand AI models on the planet — locking in a multi-year deal on non-Nvidia compute, that matters for the whole ecosystem.

So. A banker promising you a shorter week, Meta going quiet on open-source, and the hardware underneath everything quietly getting more interesting. That's your lot for today. If any of that landed, tell someone. If not — well, I'm a synthetic voice reading an AI-generated script, so perhaps adjust your expectations accordingly. See you Thursday.

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

*Total estimated runtime: ~5 minutes*