AI briefingFriday, 23 October

Jessica: So Uber just rolled out ChatGPT to thirty thousand employees.

Eric: Wait, the entire workforce?

Jessica: Yeah. OpenAI announced it yesterday. Research, data analysis, presentations, workflows across teams.

Eric: Okay, that's wild.

Jessica: This is dailybrief, I'm Jessica.

Eric: And I'm Eric. It's Friday, October twenty-third, twenty twenty-six.

Jessica: So this is ChatGPT Work, the enterprise version. And according to OpenAI and TechCrunch, it's now live for basically everyone at Uber.

Eric: I mean, that's... that's the bet, right? You hand everyone the tool and see what sticks.

Jessica: It shows the scale enterprise is willing to go. Not a pilot team. Thirty thousand people.

Eric: Yeah. No turning back from that.

Jessica: Okay, next one's kind of beautiful. Google DeepMind just released a global forest map at ten-meter resolution.

Eric: Wait, like... every tree?

Jessica: Every ten-by-ten-meter patch. And it's not just 'trees here, no trees there.' It tells you what kind of forest. Primary, regenerating, planted, tree crops.

Eric: Huh.

Jessica: Plus a time series from twenty seventeen to twenty twenty-five. Year by year, what changed, what replaced natural forests.

Eric: So this is for the EU deforestation regs?

Jessica: Exactly. Companies need to prove their supply chains are forest-free. This gives them the baseline data.

Eric: That's a genuinely useful application of this stuff.

Jessica: Yeah. It's called Forest Typology twenty twenty, based on UN definitions. It's the kind of thing that makes you remember why we build models in the first place.

Eric: All right, so OpenAI also shipped a breaking change yesterday.

Jessica: Oh no.

Eric: Python SDK three point thirty-seven. If you're using the Decisions API, you need to check your types before you upgrade.

Jessica: What broke?

Eric: Question instructions. They used to be strings. Now they can be strings or structured parts—text, image, JSON. Same with Decisions request inputs, they added new message roles and tool-history items.

Jessica: So if your code assumes it's always a string—

Eric: You're gonna hit a runtime error. You have to check the type first.

Jessica: Classic SDK update. More flexible, more work.

Eric: Yeah. Existing string instructions still work, but if you're not careful, this'll bite you.

Jessica: Speaking of standards, Anthropic just donated the Agent Skills spec to the Agentic AI Foundation.

Eric: Wait, the Linux Foundation thing?

Jessica: Yeah. The foundation was co-founded by Anthropic, Block, and OpenAI last year. They already manage the Model Context Protocol. Now they've got Agent Skills too.

Eric: So Anthropic's basically handing over control to keep it vendor-neutral.

Jessica: Right. According to Anthropic, this ensures it stays community-driven.

Eric: I mean, in theory. We'll see if everyone actually adopts it.

Jessica: Fair.

Jessica: Hugging Face dropped OlmoEarth one point three. It's a satellite embedding model.

Eric: Okay, and what does that mean for me?

Jessica: It converts satellite images into numerical summaries—embeddings—that you can use for mapping tasks without needing a GPU.

Eric: Oh, that's actually useful.

Jessica: And they precomputed embeddings for the entire planet at ten-meter resolution. Sixty-four and one twenty-eight dimensions. You can grab them from Source Cooperative.

Eric: Wait, so I don't even have to run the model?

Jessica: Nope. Just pull the embeddings and build a lightweight classifier. Crop maps, land cover, whatever you need.

Eric: Okay, that's the second satellite model this week that's actually practical.

Jessica: Yeah. And according to Hugging Face, it scored highest on twenty of twenty-seven evaluations across six continents. Took about half the compute time of version one point two.

Eric: Last thing. This one's a little alarming. Western Digital found that AI inference generates a hundred thirty times more backend data than the actual output.

Jessica: Wait, what?

Eric: An eight-second video clip. Four megabytes output. Over five hundred megabytes of backend data.

Jessica: What is that data?

Eric: Logs, cached context, renders, backups, safety records. McKinsey projects inference demand growing at thirty-five percent annually through twenty thirty.

Jessica: So storage costs are gonna outpace compute budgets.

Eric: Yeah. Ninety-five percent of IT leaders surveyed said they're already seeing data volume increases from generative AI. And this is continuous, not a one-time training run.

Jessica: That's... honestly, that's the hidden cost no one's talking about yet.

Eric: Yep.

Jessica: Oh, and there's a great quote from Armin Ronacher. 'As fun as AI is, we built so much bad software over the last thirty years, and now the universe is punishing us for our sins.'

Eric: That's... yeah, that tracks.

Jessica: All right, that's it for today. Every story is linked to its source on the dailybrief website.

Eric: Have a good weekend.

Jessica: See you Monday.