These are 50 of the hardest questions I put to AI assistants over almost four years. They come from a full screen of my ChatGPT and Claude history. Most are my own hypotheses about inference efficiency, latent reasoning, continual learning, agent memory, world models and evaluation integrity. I asked them to test each idea against the literature.
Mohamed El HarchaouiAI engineer · Data scientist, 10+ years
Score over time
Each dot is one question, placed by date and overall score. Filled dots are the top 50 and hollow dots are the other scored questions. Click a dot to jump to its entry.
Threads followed over months
Several questions belong to the same line of inquiry, revisited and refined over time. The chips give the rank of each entry in date order.
The questions
Questions are lightly edited for spelling and clarity. Expand Original wording on any entry to read the prompt exactly as I wrote it.
How this list was built
Extracted every message I sent in two exports: ChatGPT (Feb 2023 – Feb 2026) and Claude (Feb 2025 – Nov 2026).
Reused an earlier local topic clustering of the ChatGPT history (bge-m3 embeddings, UMAP, HDBSCAN) to find the AI-research clusters.
Pre-scored all prompts with research-vocabulary and question-form signals, then removed duplicates, code pastes and pure proofreading requests. This left 512 candidates.
Ran a semantic pass to catch short or unusually worded questions that the keyword filter missed: every prompt was embedded with bge-m3 and compared with the scored questions, and the 220 nearest prompts not yet shortlisted were added to the candidates.
Read each candidate in full, merged follow-ups from the same thread into one entry, and scored the survivors on the four-part rubric.
Wrote a short note for each entry that places the question in the current research landscape.
Criterion
Weight
Question asked
Depth
35%
Does it go to the mechanism: losses, activations, compute and data?
Originality
30%
Is it my own hypothesis or design, not a lookup?
Frontier
25%
Is it on an open, current research problem?
Stakes
10%
Does it touch safety, integrity, governance or society?
Each criterion is rated 0–10. The overall score is the weighted sum, scaled to 100. Dates are the dates the message was sent.