My original comment was poorly worded. Ghana gained independence 69 years ago, not 200.

The kidnapping and exploitation was in full swing 200 years ago, but it started earlier than that and ended way later, if ever.

Top tier post. Can't wait for the experts at awful.systems to come explain. The question by OP is a very good question so I am sure we will get expert explanations from users of the awful.systems instance.

They said they're working on Orthus for Qwen 3.5. It'll be amazing!

My oversimplified and possibly wrong understanding: this is like speculative decoding, but instead of a separate draft model (which does its own prompt processing), they use some diffusion thing strapped on top of the main model. The diffusion reuses the high-quality prompt processing result of the main model.

The 7.8x faster claim sounds almost too good to be true. But even if we get like 3x then this is still a huge revolution in localLLMing.

1.5 hours runtime for like half a liter of gasoline?? That's unbelievably inefficient. A half-liter of gasoline is like 15MJ, should power a laptop drawing 30W for a week.

Maybe it would be better with a fuel cell.

If you would rather not trust any AI company with your data, consider heading to !localllama@sh.itjust.works where self-hosted LLM are discussed!

I see. Mistral was the favorite in self-hosted LLM circles back in 2023-2024 but general opinion is that they have since been far surpassed by Chinese and American models, hence my question.

Good to know they've found a market with their online offering.

Is there a use case where Mistral still beat Qwen or Gemma? If you're using Mistral, which model and what do you use it for?

Reinforcement learning makes the model better over time, so why should there be fewer and fewer good results?

If you're talking about the rate of improvement going down, then yes, of course. That's bound to happen (unless you have an actual intelligence explosion, but in that case you won't know what "good results" even mean anyway).

Yes :)