Short article this time. The new LLM from Anthropic, Claude Sonnet 3.5 , has just dropped and I’m impressed. It’s fast, it’s accurate, and the code it produces seems to work more reliably than ChatGPT 4.0.
Claude immediately built a simple tic-tac-toe game for me as a one-page JavaScript app. It was fast and worked perfectly the first time. I did, however, have to update my prompt to make the computer player “think” about its moves. After that, the game was unwinnable. Then I tried something more ambitious, building an escape-the-maze game. Claude was very fast, very reliable, and the code never crashed. I had to prompt it a couple of times to improve gameplay and at one point, it forgot to increase the number of monsters, but quickly fixed that when I pointed it out.
Unfortunately, I pay for an OpenAI account, not an Anthropic account, so I quickly hit its limits on free requests. However, in that time, it was amazingly fast and accurate.
To “finish” the game, I had to switch back to ChatGPT 4o and it was night-and-day different. I had to manually fix the orientation of the main character’s mouth and it was slow. Very slow. However, it did a decent job of creating appropriate monsters and nicely handled a bug where sometimes the game would pause permanently.
Historically, when things like this happen, OpenAI quickly releases an update of ChatGPT which reclaims the throne. Given how recently 4o was released, it will be very interesting to see if OpenAI matches this move. The general opinion around the web seems to be that despite Claude’s limitations (for example, it can’t run code), it’s one of the fastest, most reliable GPTs around. Anthropic has really done a fine job here.
If you want to try out the game, click here.