xnzakg

I switched back to glm-5.3-flash from qwen3.8-flash-next.

qwen was faster in t/s, but feels like glm spends less tokens thinking and seems better at finding issues, to the point where it ends up being faster.

at least in Europe, diesel pumps have a larger nozzle, so they don't fit in gasoline cars. the other way goes fine though.

I guess they're the long term study subjects themselves at this point. honestly, maybe it's a positive thing? lots of testing across a variety of subjects, without all the ethics and paperwork slowing it down.

GLM-5.3-flash is open-weight though. you can run it on your own hardware for a few k$. yes, it will probably be slower, but you're no longer reliant on a subscription, a working network connection, or the servers not being overloaded.

I've had a very similar issue happening randomly. same pattern of the model outputting gibberish, seemingly trying to end its response and starting to panic when it couldn't, repeating a bunch of times.

turned out to be a bug in a patch, causing kv cache indices to be stored in too narrow variable types, making them wrap around.

fascinating, and slightly horrifying, that LLMs are able to notice that their output isn't what they meant to output. reminds me of the mirror test.

not a bad analogy imo. power tools vs hand tools. faster, but if you're not careful you can cause more damage before you notice. like a table saw, where you can cut too far or at an angle and the result will look bad. or an electric screwdriver, where you don't feel the torque as much, so it's easier to cross-thread.

and skills from building with hand tools help you judge the results, but not necessarily with actually using the tools.

Huh, almost as if the governments get something out of allowing the companies to abuse us...

Don't worry, the arbitrary software is probably running under Intel's ME or AMD's PSP.

If the kids are clever enough to have a second browser the parents don't know about, I'm pretty sure they'll find a way to bypass this. For example, who's stopping an older sibling giving remote desktop access to his browser?

It's the same as age limits on energy drinks, alcohol, tobacco etc. More of an inconvenience than a restriction if someone is determined enough.

We should focus on having parents teach their kids what they should and shouldn't be doing, instead of "blocking" them and pretending that solves the issue.

But in many cases, you could split it up into smaller, `inline` functions.

On the other hand the standalone models would all need "basic" knowledge, so I'm assuming this would be much less disk space-efficient.

How would this help against someone submitting an actual, non-compromised version bump, then adding malware once it's accepted?

Question is if this would be thorough enough for this attack? A package with a slightly more involved build process, maybe some patches because it was made to build on a different distro. Maybe you've already installed (and thoroughly inspected) it before, so you're only updating to a newer version, so you're not as thorough with your review. Or an xz-style backdoor.

> ultrasonic chirp authentication

Source? Can't find any good source on this.

I suppose using an LLM-powered autocomplete might be something for you then?

Then you can be writing the code, with the LLM doing the "boring" parts, in chunks small enough you can review them on the fly.