This is ultimately the problem with using LLMs in education. Students will spew information back at you with absolutely no understanding of what it means. Did you READ the material? Obviously not. You just asked ChatGPT to write you a paper on Pride and Prejudice or Reconstruction.

Show Post

An alpaca farm in New Zealand, to be precise.

Show Post

Yeah especially with spare attention models like Qwen3.8 - even just going to Q8 on the KV completely ruins the quiality. I've also got a clever script which writes KV caches to disk so different sessions can interleave without having to prefill for half an hour - that's been a huge win.

Show Post

Good tip on leaving the KV cache unquantized. Quantized KV tends to hurt long-context recall first, which is the whole reason to run a 260K window.

Show Post

It is another market segment they could monetise. You can even advertise a higher tier of service that has "expert human oversight" with ridiculous per hour charges for the human. The Monkey's Paw better not fucking move...

Show Post