I’ve learned this lesson over the years. It is quite common that users make a screenshot of the error with their phone, and send it on to support with hardly any details. The fact that errors become recognizably different is also an improvement: the user and support staffer can recognize recurring errors, and notice patterns.

It reminds me of facebook. Staff was locked out of the office due to the outage they were supposed to fix.

Luckily the plasma torch and bolt cutter didn’t require logging in with Facebook.

Phish? It turns out to be a band: https://phish.com/

In a perfect world, there would be a stable version of chrome, that would get fixes, but would crucially not get the new features that introduce new vulnerabilities. Not a fun job, I know, but with today’s coding agents it wouldn’t even be an unreasonable ask.

My favorite USB ethernet adapter is a lowly 100 MBit one that works everywhere without requiring driver downloads.

The nuclear option is to disallow it to store your watch history. You get only a search bar and your subscriptions.

I switched to this. It is less fun, but also less of a timewaste.

For Anthropic, it is valuable that they control the scheduling, so they can move jobs around to use the infa when it is relatively quiet. If you let customers choose the time, a lot of work will start at whole hours.

Might as well plug this recent Criminal Podcast episode: https://thisiscriminal.com/episode-358-the-formula-3-27-2026

TLDL: During prohibition, US government required adding 5% methanol to industrial alcohol, hoping that this would stop bootleggers from selling it as liquor. It was sold anyway, resulting in many deaths.

To be fair, we still add 5-10% methanol to industrial alcohol. But also a bunch of bitterants to discourage use.

Are adding it or just distill both because it's cheaper?

Yeast microbes make ethanol, not (much) methanol.

Adding. That majority of things we ferment and distill will not produce anywhere close to 5-10% methanol.

Both of you are right. There is one more edge case: if you commit to buying electricity in advance it might cost you extra to not consume it. It would still be in your interest to use the power at a net marginal loss rather than not using it and paying a fine for failing the contract.

Is this something you can share in more detail? Did you document a skill for the LLM to use? And with what tasks do you see most improvement?

The point of this is to reduce a complex tool surface to a single sql query tool without losing the richness of the underlying representation.

In practice this allows for me to combine multiple, complex data sources with a constant number of tools. I can add a whole new database and not add a new tool. My prompts are effectively empty aside from metadata around the handful of tools it has access to.

This only seems to perform well with powerful models right now. I've only seen it work with GPT5.x. But, when it does work it works at least as well as a human given access to the exact same tools. The bootstrapping behavior is extremely compelling. The way the LLM probes system tables, etc.

The tasks this provides the most uplift for are the hardest ones. Being able to make targeted queries over tables like references and symbols dramatically reduces the number of tokens we need to handle throughout. Fewer tokens means fewer opportunities for error.