Hugging Face reposted this
Absolutely wild! 🤯 Google DeepMind Gemma 2B outperforms OpenAI GPT-3.5 on LMSYS Chatbot arena with a score of 1130! 20 months ago, "ChatGPT is a revolution, the most powerful model ever made," and today, you can run a model more preferred than this literally on a toaster!🍞 🚀 Gemma 2B It also ranks higher than: > Microsoft Phi-3 Medium (14B version) > Mistral AI 8x7B Instruct > Mistral AI 7B fine-tunes > Meta Llama 2 70B Test it on Hugging Face: https://lnkd.in/dbn4ZGjg Leaderboard: https://lnkd.in/dA-2CiEi
I gave it a spin and I have to admit that I am impressed with Gemma 2 2B! Ran a few simple tasks like polishing emails, making documentation concise, fixing grammar and tone in text, simple python scripts and debugging, queries using open webui web search enabled, cooking recipes etc and it performed better than I ever expected it to, for a model of its size. Blazing fast as well! Should be a decent daily driver. Excited to see how it lends itself to fine tuning, RAG and code completion.
we have a name for that we name it "Fast follower Principle" The 'fast-follower' theory takes the position that in the world of technology, marketing execution is more important than being first or even being the best. With a thorough understanding of the market, latecomers can—and often do—reach and win that market long before the pioneers think to do anything of the sort. I my self for example prepared everything only for the last step to AGI so as soon as the fundamentals are in place i will have AGI long time before the rest of the world will have it.
It is easy to overfit a model to perform well on known and available (during the training) test!
https://github.com/google-deepmind/gemma#system-requirements > Gemma can run on a CPU, GPU and TPU. For GPU, we recommend a 8GB+ RAM on GPU for the 2B checkpoint ... What was the toaster designed for that got 8GB+ RAM? :) But I have to give you this: the clickbait worked perfectly. :D P.S.: there're internet connected "smart toasters" with big LCD touchscreens, but I doubt any of them has that much RAM (or even anything close to 8 GB). :) https://www.google.com/search?q=smart+toaster
Gemma 2B achieves its performance by being optimized to run on lightweight devices while still outperforming larger models like OpenAI GPT-3.5, Microsoft Phi-3 Medium (14B version), Mistral AI 8x7B Instruct, Mistral AI 7B fine-tunes, and Meta Llama 2 70B. This makes advanced AI more accessible and efficient.
incredibly impressive but a small part of the story is also, that gemma 2 was trained on lmsys, which was not available for gpt3.5
would like to know why are you comparing it with previous versions when the top 3 are still gpt 4, 4 mini, Claude sonnet and Gemini adv
I’m imagining gemma 2 running on toaster 😱 Human: hey toaster! toast this bread little harder, I like it roasted. Toaster with Gemma: As an AI language model, I’m not allowed to roast anyone as it is against my safety and responsible AI policies. #humor 😝
Gemma 2B outperforms OpenAI GPT-3.5 on the LMSYS Chatbot arena with a score of 1130. It also ranks higher than Microsoft Phi-3 Medium (14B version), Mistral AI 8x7B Instruct, Mistral AI 7B fine-tunes, and Meta Llama 2 70B. This model can run on lightweight devices, making advanced AI more accessible.
Generative AI Global Blackbelt @ Google ◆ Founder of NLP London
1moPhilipp Schmid - what kind of toaster do you have?? 😳