11 Comments
User's avatar
Matt Ferguson's avatar

Love the insight into your process here. Especially given your experience in olde timey human research.

Expand full comment
Natalia Cote-Munoz's avatar

Thank you!

Expand full comment
Maurice Blessing's avatar

Inspiring read. I am on a similar journey using several (paid) services and learning tons of new stuff. I’d like to read more details about your workflow. For instance: do you use the models apps/web interfaces through subscription or their API’s in some other app? And have you tried services like Consensus to quickly gather and compare academic publications? I’ll keep an eye on your publication and will consider a subscription as soon as I’ve weeded out my present ones!

Expand full comment
Natalia Cote-Munoz's avatar

And thank you for reading!!!

Expand full comment
Natalia Cote-Munoz's avatar

I use them directly through their subscriptions but NotebookLM, my favorite, is free for up to 50 sources per notebook.

Expand full comment
Maurice Blessing's avatar

Thanks for the answer! I guess you also have Google One which includes both Gemini Pro and NotebookLM Pro plus I think Google Cloud 2TB (and some other stuff I don’t use). Best AI-deal at this moment I think, especially since I noticed yesterday that Gemini 2.5 pro delivered very impressively in my ‘deep research’ on Middle Eastern politics, a subject I know quite a lot about so I was surprised by the depth and breadth of it’s analysis. It performed much better than Claud Opus 4, although not with the exact same prompt (still learning that technique also).

Expand full comment
Natalia Cote-Munoz's avatar

Yes I’ve been underwhelmed with Claude’s research. Claude’s strengths are elsewhere

Expand full comment
Stephen Fitzpatrick's avatar

This is great - super helpful. Have you used Claude (other models can work as well, though I have found Claude to be best) to build out very detailed and specific deep research prompts using the conversational tone from Phase 1? I have a Claude Prompt Project whose only purpose is to develop prompts for Deep Research models. It excels at specifying exact kinds of sources, output formats, or whatever else you might be interested in.

Expand full comment
Natalia Cote-Munoz's avatar

Yes, I definitely use Claude for that, it’s quite good!

Expand full comment
Jenny Ouyang's avatar

Thank you for putting together such a comprehensive article, Natalia.

Sounds like Gemini 2.5 is truly one of the top models out there. I’ve mostly been using ChatGPT, but your piece is a strong nudge to test Gemini more rigorously.

Best of luck with the hackathon, looking forward to reading about how it goes!

Expand full comment
Natalia Cote-Munoz's avatar

Thank you! Let me know how it goes!

Expand full comment