I don't see why the transformer architecture can't be designed and trained with separate inputs for control data and content data.

Give it a shot

Great stuff.

I would argue that hardened mode should be default though, similar to how siphash is the default hashing function in Rust hash maps. Faster mode should be opt in if the user is confident that the supplied data is nonmalicious and they need the speed up.

I’m still open to it and thinking about it actually. I will explore if it’s possible to eliminate the large losses on common patterns and if it turns out it is then it’s a no brainer.

Going forward this and the extended operators + large pattern perf will hopefully be a strong selling point to gain more traction

That's Rule 5 no?

You're responding to a bot

The boilerplate and indirection isn't done for performance

I've worked on optimizing modern slow code. Once you optimize a few bottlenecks it turns out it's very hard to optimize because the rest of the time is spread out over the whole code without any small bottlenecks and it's all written in a slow language with no thought for performance.

I would argue that making the company experience the consequences of its choice of metrics / mandates is in fact a moral imperative.

In fairness though, it does give you good practice for the essential skill of maintaining / improving an old codebase.

Amazon has had a really bad string of various outages recently. Assuming they're internally treating this as business as usual in post-mortems then perhaps the newsworthy thing is actually that they aren't taking their outages seriously enough.

This is reminding me of the story of a Japanese airport doing a full security sweep because one of the airport restaurants had misplaced a kitchen knife.

My understanding is it's taken as a given that the authorities at US airports aren't bothering to catch baggage / item thieves amongst airport staff. The only exception is when a firearm (or luggage containing a firearm) goes missing.