We need chatbots to bombard all our social media feeds with pro-western military propaganda. Otherwise, Putin and Wumao and Evil Korea and The Muslim Horde and Drumpf will win.

One of my favorite moments like this was a Reddit thread where some account was pretending to be human and arguing with people in favor of the CEO's actions during The Purge. Then one person asked it a question about making some dangerous thing or other, and it starting replying with things like "As an AI model, I cannot explain how to do that." and stuff. It was great.

The techbros who want to use AI and the developers of AI aren't quite the same group.

Good point, however, thinking about it, I would consider those rules to be closer to AI than LLMs, because there are logical rules based on "understanding" input data. As in "using input data in a coherent way that imitates how a human would use it". LLMs are just sophisticated examples of the dozen of monkeys with typewriters that eventually come up with the works of Shakespeare out of pure chance. Except that they have a bazillion switches to adapt and are trained on desired output, and then the generated output is formed with some admittedly impressive grammar filters to impress humans. However, no one can explain how the result came to pass (with traceable exceptions being the material of ongoing research), and no one can predict the output for a not yet tested input (or for identical input after the model has been altered, regardless how little). Calling it AI is contributing to manslaughter, as evidenced by e.g. Tesla "autopilot" murdering people. PS: I know Tesla's murder system is not an LLM, but it's a very good example how misnoming causes deaths. Obligatory fuck the muskrat

I can run a small LLM on my 3060, but most of those models were originally trained on a cluster of a100s (maybe as few as 10, so more like one largish server than one datacenter)

Bitnet came out recently and is looking like it will lower these requirements significantly (essentially training a model using ternary numbers instead of floats to reduce requirements, which turns out to not lower the quality that significantly)

They should do to AI what they make me do at work: More with less.

Basically Mistral, check /lmg/ in /g/, if you have a GPU newer than 2 years you can probably run a 32B quantised model.

Haha try the entire datacenter.

If LLM was practical on three servers everyone and their mum would have an AI assistant product.