Brainstorming requires several shots in the dialogue and is more difficult to evaluate. I just started playing with Google Co-Scientist, but it takes a long time (over 24 hours) for its multi-agent orchestration

I re-use the same test prompt for lit review and found that many LLM are limited by the initial retrieval step. When both ChatGPT and Claude cited the same bad paper, I surmised that both of them may be working off Bing API results.

Gemini 3 surprisingly got the correct answer.

I would much rather pivot LLM evaluation on Bloom’s Taxonomy of tasks, instead of Sam Altman’s money grab from companies about replacing people roles.

Although I’m a huge critic of LLM outputs, I found two perspectives helpful. First, I considered LLM as a Reddit summary, and I would pick query topics accordingly. Second, I came across the revised Bloom’s Taxonomy for cognitive capabilities. I only trust LLM with tasks on the 2 lowest capabilities

I sympathize. Have you tried RAG with Bing results? 😳😆

Did SBB stop the train at the border? I’ve heard that late trains from Germany are not allowed into Switzerland.

RLHF before its time!

I would also add two points to this excellent advice! First, consider a few leadership role models. Second, observe what transferable skills and behaviors that they exemplify, for your own reinforcement learning.

👆👆