What's new in LLMs

Anthropic released Claude 2 with improved performance across various metrics such as coding or answering standardized tests. It’s technically capable of processing 200k tokens although for now it still has a 100k token limit, which is more than anyone else. You can upload a document directly to it and ask questions about it, or you can upload a writing sample and ask it to follow a similar style.

Google released a few updates to Bard such as: support for 40 languages, image input, and an option to adjust the response style. I tried out the image input with a photo I had taken of a spider and it did well:

OpenAI ChatGPT - they still didn’t re-enable their Bing search integration since they didn’t want people copying entire web pages. But they recently released their code interpreter which lets you execute Python code within it. You can also upload files to it and have it run code on them. The interpreter is also able to figure out what Python libraries to import as needed, and it can output graphics.

Meta (Facebook) announced CM3leon although they didn’t release it. It’s able to produce both text and images using the LLM-style technique. To find out more about it, I uploaded their paper (available here) to Claude and asked for a summary. Seems like it did a good job!