Leave it to a Brit to deliver satire with a straight face in the most deadpan style. This had me in stitches, even my wife joined in to watch the whole thing.
Update: For anyone curious, presenter is Steve Sanderson and a little bit of context in this LinkedIn Post.
Nine things, Space, Rust, Pi? Oh my!
Asked about backlash to AI, Superhuman Mail CEO Rahul Vohra — whose company makes an AI-powered email assistant — seemed unfamiliar with the premise of the question. After hearing about poor polling around AI, he responded: “We don’t really see that.”
Keep in mind that AI has not increased the number of people who need software, or the number of maintainers who can review it. It has mostly increased the amount of code and the number of projects competing for attention. Some of that is healthy, but a lot of it fragments effort that should be shared.
Replacing a 3 GB SQLite db with a 10 MB FST
I do wish to point out, of course, that the whole reason it was possible to experiment cheaply and come across this serendipity was because 9 months ago, faced with the choice to either do the bad easy thing or the good nothing, I chose to do the bad easy thing.5
Ask an Astronaut: 333 hours of Q&A footage with astronauts : I’m at a loss for words, just see it for yourself. Saving it for Lin Lin and myself.
Watch a neural net learn to play Snake : Mesmerising to watch, it does seem to abruptly stop occasionally. Refresh the page if it does.
Our final finding is straightforward but important: training on a broad set of safety-relevant environments improves alignment generalization. Capabilities-focused distributions of RL environment mixes are changing and increasing rapidly; it is not sufficient to assume that standard RLHF datasets will continue to generalize as well as they had in the past.
Idempotency is easy until the second request is different
Leave Me Behind : Adam McNeilly’s quiet end-of-an-era post ruminating about the good old days. While I understand their sentiments, and the human mentorship angle, I don’t agree with the overall negativity. You can certainly leverage AI tools to fast-track your learnings. Not everyone had access to great seniors and mentors. The job was never solely about writing code, although that was the allure. We were paid to collaborate and build to generate value for our users and employers. The need for that collaboration and thus human touch, has not gone away.
Torrix, self hosted LLM observability, no Postgres, no Redis
Simon Willison’s annotated lightning talk from PyCon US 2026.
OpenAI and Anthropic had spent most of 2025 running Reinforcement Learning from Verifiable Rewards to increase the quality of code written by their models, especially when paired up with their Codex and Claude Code agent harnesses.
In November the results of this work became apparent. Coding agents went from often-work to mostly-work, crossing a quality barrier where you could use them as a daily-driver to get real work done, without needing to spend most of your time fixing their stupid mistakes.
The supposedly best model changed hands five times between the three providers. Coding is where most of the action is. Simon’s calm summary remains the best way to figure out what you missed since you last looked up.
The argument from the title: Python’s ergonomics were optimized for humans, and humans are increasingly not the proximate authors. If the model is happy writing Rust, the case for the slower-runtime, weaker-typed alternative gets thinner.
The last twenty years of language choice were shaped by a single constraint: humans write the code, and humans are slow at low-level languages. That constraint is gone. Stack Overflow’s 2025 survey had Rust as the most-admired language for the tenth consecutive year at 72%, with Gleam at 70%, Elixir at 66%, and Zig at 64%. The stated preference was always there; the tooling finally caught up to the preference.
This reads like the strongest argument until you start unpacking it. What happens when the agent halts halfway through a refactor? When your context window runs out, or the provider doubles the price? Or for those 20% of times where it just doesn’t do the job?
The model writes what I want to read, and I still want to read the simpler thing. Use what you know best, optimise when the need actually arises. I say this as a Rust programmer and advocate, not everything needs to be in Rust. Rewrite It In Rust is enough of a meme already.
Why don’t we go a step further if we’re all-in on AI? Let models and agents invent their own language and let it optimise based on our preference.
Koshy John on what the tool is good for and what it is not. The argument is small and worth keeping: use the model to extend the questions you can ask, not to skip the part where you think the answer through. His analogies are spot on.
Going back to the analogies: This is like copying answers through university and then showing up to a job that requires independent thought. It is like using a calculator for every arithmetic task and never developing number sense. It is like relying on self-driving features before learning how to actually drive. The support system may make you look functional, but it does not make you capable.
The piece I now send when someone shares a chat transcript instead of a position.
Three things.
WebMCP is available for early preview : The browser-as-MCP-host idea. Tools that pages can offer, sandboxed, gated.
As the agentic web evolves, we want to help websites play an active role in how AI agents interact with them. WebMCP aims to provide a standard way for exposing structured tools, ensuring AI agents can perform actions on your site with increased speed, reliability, and precision.
A Japanese glossary of chopsticks faux pas (2022) : Hashi-watashi, neburi-bashi, yoko-bashi. Each one named, with the reason. Should have learned about these before our trip.
Profiling Hacker News users based on their comments : Mostly a tool note. But also an uncomfortable reminder that profiling tools are accessible to masses now.