Extra Garlic

有没有想过，全球每天烧那么多 token，光 Google 就已经一个月 3.2 千万亿 token。到底都在做啥？ 据说55%都用在代码上，那为什么至今仍未看到有颠覆性的产品出现？？

I had a strange concurrency bug at work last week, one that made me question my whole understanding of distributed locking. a data race was happening that looked like the standard concurrency issues i'd seen before, two requests were modifying a group of rows at the same time. the standard solution is to find the rows they're both modifying, put a lock on them, and the issue is resolved. this covers a huge class of races, lost updates, dirty writes and other nasty stuff. my bug wasn't like that, the transactions were touching different rows that has share the same "group_id", and i could have tried to lock every row that is being transformed but... that's its own kind of problem, groups in this case can have hundreds or thousands of rows, holding locks on all of them creates contention everywhere and you end up with lock queues that grind throughput to nothing, postgres has a whole class of issues that come from exactly this pattern. what i actually had was something subtler, the rows were connected by a rule i didn't quite see. the book DDIA (chaper 8, v2) used an example that made the problem click for me: We have two doctors on call for the same shift at a hospital, both feel sick at the same time and both open the app to take themselves off call, the app does the responsible thing in each transaction and checks whether at least one other doctor will still be on call before letting them go. both transactions run that check at the same moment, both see that yes the other doctor is still on call, both decide it's safe to proceed, and both commit. the hospital now has zero doctors on call and not a single database rule was broken, no lock would have caught this because each transaction was modifying a different doctor's row. the race wasn't on either doctor's row, it was on this abstract rule that connects them (the invariant) that at least one doctor must remain on call at all times. This is a bit weird to reason about because the rule isn't a row in the database, it's this abstract relationship between rows. this was the gap in my mental model of locking that took me a week to figure out, we get taught to lock rows but sometimes races that exists aren't on rows. the technique described for solving this problem is called materialising conflicts and it sounds obvious once you've seen it. and you would have probably guessed it from reading the above. if the thing you need to lock doesn't exist as a row you create one, a row whose entire job is to be the lock target for that invariant. The hard part is reasoning about the problem and trying to find that invariant. continuing with the example for the doctors you'd create a row representing the on-call shift itself and lock the shift when anyone tries to go off call. the actual on-call data doesn't change but suddenly the race has a concrete row to fail safely on. the pattern to extract is: anywhere two entities share an invariant the race lives in the invariant not in either entity, and if the invariant isn't represented as something you can lock then your locks are pointing at the wrong thing entirely.

I'm done. I'm f***ing done.

watch ts go green end of day

Breaking: Korean court partially grants injunction against Samsung union’s “illegal strike,” effectively making a strike impossible.

The AI industry just invented a new job. Wait until you hear what it does.

They reinvented the hearing aid by studying the human ear Normal hearing aid: $4700 Theirs: $20

New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text.