I have little reason to believe this, and Anthropic have every reason to lie about it.

Another common annoyance is having a request go to a provider that dribbles out ~1 tps (even for small models like DeepSeek V4 Flash). If you cancel the request, you still get charged for the prefill and the handful of generated tokens. If you don't cancel the request, you might be waiting 10 minutes for the turn to finish.

The overall experience is pretty good, and it's the best way to try new models, but they don't appear to do any real vetting or apply any quality standards to their providers, and occasionally it bites you.

Hmm? You have the revenue already. I know it's awkward from an accounting point of view, but you already took my money. "Letting" me keep the balance in the account is not generous.

Edit: on re-reading this came out more combative than I intended, sorry. I think what you're doing is reasonable.

> The MHS driver also helps an AI agent understand how to use a device it has never seen before, giving it information about machine characteristics that may not be discernable from code alone (for example, the weight of a robot arm, which is important for knowing how to manipulate it safely).

I physically flinched when I read this. VLA, JEPA, sure, they make sense. Is connecting an LLM to a robot arm and say "perform this complex physical manipulation task by issuing text-based commands, make no mistakes" really the right abstraction?

> We have a working exploit (OOB access in the V8 heap), our security folks put one together based on the example I posted above (and they're cleaning it up to post it here). In general, we find that correctness issues like this are pretty much always exploitable with a bit of effort (not even that much effort normally, just gluing together a few gadgets), so we treat correctness issues as security issues until they are proven not to be, rather than the other way around.

The floating-point-to-heap-corruption chain here is... uniquely JavaScript, but in general getting two different implementations to disagree is the start of lots of interesting inconsistent behaviour.

Separately, local inference frameworks tend to expose all kinds of weird and wonderful gadgets on their HTTP interfaces, which can be a rich source of vulnerabilities even if the /v1/chat/completions API etc is reasonably hardened. For example llama.cpp has a custom API for saving and restoring KV checkpoints to disk, and I wouldn't be surprised if that could be used as an arbitrary disk read/write.

Using these APIs usually requires the API key (bearer token), but again, people think it's normal to run the agent's shell in an environment where it has both the API key and the necessary network access to use it.

First one of these I've seen using DOM manipulation and CSS transitions instead of canvas, so that's neat.

Not completely true: KV is a projection of the activation at each layer's input, so attention heads see (a representation of) all previous tokens' activations at that layer. The hard decision at the LM head doesn't change that.

In signage you can have one character per column (kind of like one word per row on some English signs), making it look like right-to-left text.