Yesterday’s note discussed how making LLMs bigger, with more parameters (training data), may be through. But that isn’t stopping Anthropic, AI21, Cohere, and Character.AI from trying. Even Elon has teased his own LLM called TruthGPT. And now Stable Diffusion has joined the fun.
Stable Diffusion is known for their open-source AI image generator. (Sidenote: their Reddit is really entertaining and the most inventive space.)
Stable Diffusion’s LLM, called StableLM, is completely open-sourced and uses a permissive license, which means you don’t have to work through an API or purchase a license to build with it. That’s a major contrast from everything else out there and could give it a long-term leg-up on GPT-4. It’s also much smaller in size, making it more efficient to run, and theoretically could show how smaller models can outperform bigger models.
Notably, the models StableLM released don’t include RLHF (Reinforcement Learning from Human Feedback). This is the feature that largely made GPT-3.5 so much better than GPT-3. StableLM plans to include this feature in their model in the next releases.
Ultimately, though, an LLM is only as good as its interface. The reason that OpenAI is winning and surpassing 100M+ users with ease is because of ChatGPT. That application gave ease of access to their GPT-3 (3.5 and 4) model.
I don’t think it’ll take two years for us to get a widespread StableChat competitor to ChatGPT. I think it’ll be just months away considering it’s an open-sourced model and thus any builder can work with it. I’ll be keeping my eyes on Twitter to see what people are creating with StableLM in the coming days.
I didn’t have time to test it myself, however, follow the Tweet below if you’re interested in trying the model out yourself:
Another option is following the steps below, which allows you to bring the LLM into Terminal on macOS: