Hey Giacomo

Sorry - I didn't check Bluesky for a long time. This semester of my masters is a bit busy. If you're still interested in a discussion about AI support for students: Let's talk. =)

Now for the second question: Sorry to disappoint, but I am not sure. I mean use gguf and exllama2. And I think exllama 2 uses the transformers library to accelerate multi gpu inf.

Sorry I need tp ise Bluesky more regularly. 🙈

I run Windows cause most viewers do. But I plan to add a third SSD with Linux as most workflows which aren't adopted by Pinokio Computer are easier to get running there.

is of utmost importance.
So let's not repeat the disaster with the reflection LM. 🙈

Me 😬🫡

But I am only a small Youtuber in the AI space woth only around 2700 subs. But I create AI Homeservers.

Vram isn't a big problem here. With such a thinking model it's more about the time it will take me to do it manually. 😅

But true, FP16 will surely deliver some improvement. But I guess I am going to write a python script to automate the "question-answer part".

Any guess on how bad degradation will be? I plannto test it locally with my own set of questions. I am not yet sure if I want to trst it in FP16, q_8 or q_4.

😂 👌 Poor clippy.

Would be funny to see what your AI assistant would think about beeing compared to clippy. 😅🙈

But yeah, computer use will be awesome once it's rate of failure gets lower. But I haven't yet tried the mode from anthropic to be honest. ✌️

Well, I mean no lecture really is guarded. You don't need to show a student ID to get into a lecture. And there are lectures for a broader audience on some evenings.

Just again dropping casually a source of educational gold. Thank you for sharing! Open education is a gift! 🫡

I can't afford any of these, but the L40S seems to be the better choice on the paper. Newer architecture, more tensor cores,more and higher clocked cuda cores, better memory bandwith(?)

In any case: Good luck! Not sure if any player in academia trained on such a huge set.

I can imagine that. 🙈

wow... Are you preparing to train on that amount of tokens? 😅

One more reason to switch to this platform. As it seems, even the API is open here. We can use the data to create our own open models (in theory).