Hej @mediano.nu, kan I ikke få Lars Hallengreen i studiet igen?

3. Use MCP Reasoner to guide the model into coding with less golden retriever energy: github.com/Jacck/mcp-re... - this MCP server can do both MCTS and Beam search.

2. "I see the issue", "Ah, now I see the issue", "Apologize for the confusion" = big reg flags. As soon as the model doesn't fix an issue in one or two iterations, the best thing is to revert a few steps back in your conversation and start from there with planning and reasoning.

Maybe it's the Silicon Valley ideology people oppose rather than the tools.

I just did a reasoning test on 4o on pro: chatgpt.com/share/67b10a...

Same question on a free ChatGPT account that I expect to be 4o as well: chatgpt.com/share/67b10b...

Bottom line is that 4o on pro got it right and (likely) 4o on free for it wrong. Could be purely random, who knows 🤷‍♂️

I highly recommend this episode of Latent Space with the guys behind: open.spotify.com/episode/1oEW...

👀 stackblitz.com
👂https://open.spotify.com/episode/1oEWNAYiJWoN7FBbYcZF7S?si=N-kZOoOaSIq7PaW1vvi2cA

Webcontainers 😅

Interesting test. Did you try Gemini 2.0 Flash Thinking Exp?

Some guy on Reddit tested it against Sonnet: www.reddit.com/r/ChatGPT/s/...

👋

Gemini is probably the best for vision tasks.