Deepnote

Did prompting get unalived again? We’re already seeing that models like Fable do better without the user’s explicit preferences on how the task should be completed. In the very near future, the models will proactively propose three directions, and you pick one. You spend zero time framing requests. The agent already knows your codebase, your project, and your last fifteen decisions. It just asks "A, B, or C?" This is the cleanest RL signal. Every pick is a labeled preference.

Nobody: Absolutely no one: Chrome: let me prebundle my own local LLM

Sonnet 5 just became Claude's new default at less than half the price of Opus. One day after US export controls hit Anthropic, Zhipu open-sourced a rival at a fifth of the cost. And Midjourney pivoted into full-body medical ultrasound. Inside this issue: 🔵 Fable 5 returns: back for all users after a 2.5-week export-control freeze over a jailbreak weaker models could reproduce anyway 🔵 Jalapeño: OpenAI's first custom inference chip, taken from design to tape-out with Broadcom in nine months 🔵 Meta's Watermelon: 10x the compute of its predecessor, while Meta starts renting out spare GPUs 🔵 Funding frenzy: Baseten nears $1.5B at a $13B valuation, five months after raising at $5B 🔵 Ornith-1.0: an open 397B model that matches Opus 4.7 on SWE-Bench by learning to write its own scaffolds Read the full rundown below.

One European at the table and suddenly it's a two-hour lunch.

Which university produces the most data leaders? It’s not Stanford, MIT, or Harvard. We studied ~40K data leaders across 140+ countries, then ranked every school by the share of its alumni who reach VP or C-suite in data. The schools that produce the most data leaders by volume don't even make the top of this list. Can you guess what the top 3 schools are? * Data taken from State of Data Leadership in 2026 report by Deepnote. Access it at deepnote[.]com/research/state-of-data-leadership

I deleted half of our eslint config last week. Linters were built to catch human mistakes, but most of the code out there isn't written by tired humans anymore. It's written by a model that writes better code than humans those rules were built for. So, half the config was guarding against a problem that no longer exists. What I kept: the rules that encode something the model can't enforce on its own (e.g., architecture boundaries, security constraints,…). Up next: tests. Everyone thought that we need more tests to constrain agents. But asking an agent to find issues in the code is already a more reliable way to find regressions. So, who are we writing these tests for?

We studied the careers of 39,765 data leaders. What drives their promotion, and how AI changed their mandate? A few insights from the report: - Tenure at a company doesn't predict promotion. - 1 in 3 data leaders now owns AI as well. In 2018, it was 1 in 10. - Only 4.5% of Heads of Data ever reach the C-suite. 81% just move into another Head role, with most CDOs being hired externally. Finally, data leaders are being asked to build the internal platform where humans and agents actually collaborate: shared context, open standards, systems people can trust. Want to access the free report? Comment "Leaders," and I'll send it over.

The next social network will be around apps. I suspect the next social layer on top of AI won’t be feeds of text / images / videos. It will be a feed of mini-apps. Tiny dashboards. Games. Calculators. Simulators. Instead of a reel, I’ll send you a Wordle I built in my chat. There are already some great founders, like Eugenia Kuyda, building Wabi in this space. Roblox for adults with AI in the loop.