Local news of sorts, with some repetition with what I have said since it's been a while since I wrote something, last I spoke on news was a bit after Gemma 2's release. As said, nothing noteworthy really has gone off like I predicted with Llama 3.1/3.2 and the supposed multi-modal with 3.2 being a fat nothing since other models in the vision space beat it handily and the minor releases aren't really improving the models enough. The finetunes like the ones with RPMax aren't bad and Nemotron 70B from Nvidia is my go to for smarts but otherwise, somewhat meh.
Mistral's releases have caused a bit of stir, but the most impressive have been their low end and high end releases. Mistral Large 2 or Largestral is generally recognized to be the top dog in creative writing and RP to the point it is surprising people actually went and did some finetunes like Magnum v4 but fuck everyone who is poor that can run it at a reasonable speed and quant so it remains inaccessible to the general public at large and it still isn't cheap to try out on OpenRouter but it is really good from my tests with the Magnum v4 finetune. I think now Mistral Nemo 12B have supplanted Gemma finetunes in general, with Magnum v4 proving that somewhat definitively where the 9B Gemma tune is beaten by the 12B based on Mistral Nemo. The only real advantage left is the multilingual abilities where Gemma still beats a ton of models in that regard.
Chinese models are getting better at benchmarks, Qwen from Alibaba just released a new coding model today. But because they are geared towards benchmarks, they are generally trash at RP and the few finetunes that tried have kinda sucked and I think people have thrown in the towel, even with Qwen 2.5 models having released in September. Same with Yi, only models people are kinda focused on is coding and vision.
Speaking of vision, I decided to try and give QwenVL2 27b and Molmo which are the leading vision models to try and transcribe Japanese manga pages. It's somewhat disappointing to see that they are still trash at it, they get like 50-70% of the text and then hallucinate and it gets worse if you ask them to translate. Oh well, more waiting needed, I guess.
>>89301213Because Silicon Valley and a lot of the companies aren't focused on consumer front-facing AI. A lot of AI is being focused on business cases in the background. But yes, really, the main issue is most of the startups and businesses are located in big cities.
