Yesterday I listened to the latest Hard Fork episode, which features an interview with Dario Amodei, one of a group of OpenAI employees who left to start Anthropic, OpenAI’s competitor. The interview is quite interesting for someone like me who doesn’t know this area well. But it made me want to check out Anthropic’s new [Claude 2 AI] (https://claude.ai/login).
Claude 2 is not available from Anthropic in Canada, but I got access through Quora’s Poe app.
I’m an anthropologist, a writer, and I’m interested in these big language models and teaching and writing. On the one hand, there is a lot of disruption. I’m on sabbatical, but I don’t know what that means for student essays. But I’m fascinated by AI as an editorial assistant. Many writers rely on editors, but for a long time those editors were unpaid, often female relatives. AI provides me with a free editor and critical reader, for the first time not a family member or friend.
Working with Claude 2-100K yesterday, I came to a couple of realizations.
First, 100K makes Claude much more useful and less stupid than ChatGPT. A few weeks ago I had the idea of using ChatGPT 4 to rearrange a book of fragments. Much of my work involved condensing fragments into summaries for ChatGPT 4’s short memory, or at least the short memory I have access too. With Claude 2-100k, I could give the AI the whole draft. We discussed several possible outlines. With the ChatGPT 4 model I have access to, it always felt like I was at the limit of the AI memory and capacity. With Claude-2-100k, the answer was much slower to come, but was feedback on the entire damn manuscript. My first book took 8 months. This took under a minute.
From there, I had a long, free-flowing conversation with Claude 2 about my book, its ideas and theories, and then we turned to writing and the differences between how Claude 2 writes and how I work. Frankly, Claude 2 just felt much more thoughtful and interesting to talk to than ChatGPT 4.
ChatGPT feels like it has less memory, is dumber, and frankly more prone to nonsense. It seems smart at first, but on further review it is often very formulaic.
Maybe Claude 2 will seem that way to me, but yesterday it seemed like a much more interesting interlocutor than I’ve had on this book, ever. What it says about me that I say this about a large neural net running on servers, converting text into tokens and then running complex vectors and matrices, leaves me very confused. What does it mean to be human, I wonder?