Meta's Muse Tops the Charts, Harvard Drops a 6.12B-Request Dataset, and Markdown Quietly Becomes Source Code
Meta's agentic app Muse hit #1 on the App Store this week, and the stock moved with it. That alone tells you where the consumer AI mood is right now — people aren't just chatting with models anymore, they're handing over tasks. Booking flights, sending emails, browsing the web on your behalf. Muse is basically the "let the AI do it" pitch, and apparently it lands.
To be fair, topping the charts isn't the same as being good. App Store rankings reward novelty and hype as much as utility, and agentic apps are the new shiny thing. But Meta's timing is interesting — they waited, watched OpenAI and Anthropic burn money on enterprise deals, then shipped a consumer-facing agent that just... works for normal people. I've seen demos where it books a flight in under a minute, and honestly the friction-free part is what sells. The skepticism? That comes later, when an agent books the wrong flight.
Meanwhile, if you're into the nerdy side of things, Chutes AI and Harvard just released a public dataset of 6.12 billion LLM requests spanning 9,174 models over a year. The stat that jumps out: 99% of requests are repeats within 15 minutes. Think about what that means for infrastructure — most of what we're paying for in inference is re-computing the same stuff over and over. Caching, prompt reuse, speculative execution... this dataset basically hands researchers a map of where the waste is. For anyone building AI infra, this is gold. For everyone else, it's a reminder that the "AI revolution" is still mostly people asking the same question twice.
Then there's Carson Gross — the htmx guy — making a point that's been rattling around my head since I read it: Markdown is becoming source code, not documentation. His argument: in agentic coding workflows, the prompts and specs written in Markdown are the real ground truth, and the generated code is becoming the "low-level implementation detail," like machine code you don't bother reading. He's right about the direction, even if the analogy to compilers makes me uneasy. Compilers keep their original source. LLM workflows mostly don't — the prompt session evaporates and the generated code becomes the only truth. That's backwards, and Gross is one of the few people saying it clearly. If you write specs for AI coding agents, this essay is worth your time.
On the legal/industry side, an unsealed court filing shows Microsoft and OpenAI employees privately acknowledged their AI products compete with the news outlets whose articles trained them — one Microsoft director apparently warned the public would see AI training as "astonishing theft." None of this is shocking to anyone who's followed the copyright wars, but seeing it in writing, in a filing, is a different weight. The "training data as fair use" argument gets harder to maintain when the companies themselves knew the products would eat the publishers' lunch.
And across the pond, UK MPs are calling Meta, Google, OpenAI, and Anthropic leaders to an urgent hearing on October 13 over AI safety concerns. PauseAI activists were already protesting outside Downing Street. The vibe: governments are done waiting for self-regulation to work. Whether that produces sensible rules or a mess of compliance theater is still an open question — history says a bit of both.
Quick add-on note for the builders: NVIDIA posted a solid piece on evaluating AI agents from tool calls to task completion. If you're shipping agents, the real question isn't "does it answer well" but "can it recover when a step fails mid-chain." That's the difference between a demo and a product.
From my perspective, the through-line this week is trust and truth. Consumers are trusting agents with real tasks, researchers are getting the data to make inference cheaper, and the courts and parliaments are circling the training-data question. The next 12 months are going to be messy — but messy in a productive way.
If you're juggling budgets and timelines while all this plays out, I've been using PayCalc to keep my own project math straight. Small thing, but it helps.
Views are my own. I write about AI, LLMs, open-source vs closed-source, and what actually ships.
Top comments (0)