It sounds more like this was a predefined account recovery flow, rather than some LLM agent making use of arbitrary write access.

If someone is talking about "spinning up a separate machine" for Postgres, they're not talking about a desktop or mobile app...

Why go through the trouble of shoehorning SQLite into a cloud database by getting solutions for HA/failover and DR, when you can just use Postgres off the shelf?

So you can post about it on HN, obviously

Wouldn't the rest of the world encourage ASML to keep supporting the fab because they want the chips to keep coming?

Which rest of the World? ASML already has restrictions on China from Netherlands (where they are based) and the US (which provides some core IP).

My argument is that they would add an exception for TSMC in the event that Taiwan fell under Chinese control. The alternative would be an extreme supply shock to the industry that's responsible for most stock market and GDP growth in America.

Interesting project in general. I wonder whether it could be adapted to behave reasonably without relying on threading. E.g. run the GC only when *alloc is called.

EDIT: misread the post! Never mind

You even read the comment you’re responding to? They’re saying no threads.

You're right. I can't delete anymore unfortunately

I mean they brought that incident on themselves...

Yeah, all left pad incident showed was that NPM cares more about their corporate users than open source developers.

That factor was downstream of its complexity though. It's far harder to implement a compiler for Ada 83 than even for modern C.

What's a situation where one needs to use `eval` in benign way in JS? If something is precomputable (e.g. `eval("1 + 1")` can just be replaced by 2), then it should be precomputed. If it's not precomputable then it's dependent on input and thus hardly benign -- you'll need to carefully verify that the inputs are properly sanitized.

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

because it's all one (unexplainable) matrix of weights.

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