4o mini has very impressive reasoning capabilities, especially on math. I mean just look at this! But given the price, would it be possible that vision requests are being routed to 4o? (Idea came from a discussion with author of blossom series)

Just upgraded my mac to 14.4 and found single click on desktop will clear all windows. Trying to toggle this off, only to find the first phrase that came into my mind is Exposé. One of those moments that make you miss 00s

Llama-3.1 is obtaining really cool results on math and reasoning, which is something I care the most. The paper is quite open to share lots of details. Here is my study note: 1. specialized recall/classifier over the entire corpus to find high quality math/code data. It's

huggingface.co/spaces/lmsys/g…

Do you want to learn intuitively about the impact of temperature, top k, and top p when using an LLM? 👀 Check out this interactive open demo to directly explore them hf.co/spaces/osansev…

Andriy Burkov I’d say it’s like a group of people sitting in a circle creating a story. Each person in turn has to add a word to the story based on the words provided by the people prior..,

Florence-2 + SAM2 - video processing I allowed for comma-separated class names in both video and image inference mode now, you can play with it in my Hugging Face space: huggingface.co/spaces/Skalski…

Very interesting. I thought prompts are less helpful when Lora tuned on specific task. Also, to what extend does telling llm not to hallucinate actually works

$0.19 per 1M tokens for input and output, not limit on output length.

A very straightforward explanation, perfect for some casual reading before bedtime.

Makes a lot of sense. The first vision models are really similar to other fine-tuning/post training. Just a bit at the end. Newer models (GPT4o, Chameleon) with early fusion are an entirely new pretraining stack. Easy to bet on the latter.

❤️❤️

Taking some well-deserved time off. Keeping up with fast paced tech can be overwhelming. Don’t forget to also take care of your mental well being, fellas.

I want ALL!!!

Let's go! 2 new code LLMs were released by Yi-01.AI Yi-Coder 1.5B and 9B under Apache 2.0. Yi-Coder 🚀 🧮 1.5B and 9B as Base and Chat Model with a 128K Context Window 💡Outperforms CodeQwen1.5 7B and CodeGeex4 9B and rivals DeepSeek-Coder 33B. 🥇 Achieves 23.4% pass rate on

Heads up: If you downloaded Yi coder chat GGUF today, it's probably broken. They were missing the im_start token in tokenizer_config.json. I updated it locally and it seems to fix it, LM Studio community already has the fix for 9b, 1.5b incoming Mine will be updated shortly :)

Tune in for some Yi-coder discussion x.com/i/spaces/1oyka…

❤️

lol nice meme, but yeah retrieval is usually the bottleneck