Shawn 'swyx' Wang says deep research features in AI have lost relevance since models such as o3, as iterative questioning outperforms single-prompt long reports for active learning

Still use Claude very frequently initial source research but it’s much better to aim for informal draft/research notes.

@menhguin not quite sure it lets you but yeah sure

@swyx i mean, do both?

@swyx most deep research tools will craft a metaprompt before launching (obv it needs to craft some kind of prompt). gemini is the most transparent example of this, and in that case you can often specify like 5-10 parameters which tends to improve quality by a lot IMO

@menhguin @swyx Sanity check -- the metaprompt is what's visible correct? And there's no augmented retrieval with the prior context right? What you see is what's fed into the DR agent?

@olekolekoleks @swyx again, depends a lot on the DR agent. for gemini iirc you get the exact prompt, for chatgpt i think u do, for claude i think u dont

I use it a lot to map industries/spaces in software for example, and I do they fairly often

sometimes also for shipping research if I have an idea of a category (eg shoes and certain characteristics)

the report is mostly useless imo as it disregards a lot of what you ask for

but the summary table and parts of the content are useful

value for me is fire and forget and getting multiple data points to get started prompting back and forth

there’s no good way to start asking parts of the report questions or get more quotes etc unfortunately and it starts diverging early

This is also why most enterprise AI pilots that "succeed" don't make it to production. The output is a long PDF the buyer asked for, that nobody on the buy side actually reads. The interactive version forces the human to keep showing up, which is the part that builds the workflow.

Why does Gemini still have "Deep Think" feature?

@swyx So true. Also a lot of these deep research reports were word fillers when a single 1-pager would’ve done the job equally well. Gemini’s deep research was painful to read.

@swyx took 6 months and a $200 plan for the industry to rediscover the Socratic method. at least the token bill is educational.

@swyx dead-ish. o3 does what deep research was trying to do. the 40-tab slog was mostly cosplay.

Strongly agree. The biggest unlock for me has been treating the model as a thinking partner, not an answer machine. Every report I have actually used recently was iterated in 6-8 turns, not generated in one shot. The long-form deep research output reads like a Wikipedia article. Interactive output reads like my own notes, only better.

The read-think-ask loop is the right frame. The teams getting real value from AI have not handed more to agents. They have structured the loop better. Defined where the human sets intent, where AI proposes, where approval happens before the next step. The report problem is not an AI problem. It is a process problem

@swyx The long-report failure isn't interactivity, it's that synthesis flattens every claim to equal weight. The interactive loop wins because users surface which claims matter, so the model justifies only those. The bottleneck is relevance-filtering, not dialogue.

@swyx not to be rude but deep research is a wonderful aid when used astutely - cross platforms - then synthesised and analysed. love it to help me find a range of insights esp. if i then get gpt pro to cross examine.

@swyx the one-shot report was always a demo format, not a workflow. read -> think -> ask was always what practitioners actually did. the tools just caught up.

@swyx Deep research only works when the report becomes a loop: claim graph -> source spans -> uncertainty -> next question. Once the output is a 20-page blob, the user loses the steering signal that active learning needs.

@swyx a good follow-up question teaches you more than any summary dump

@swyx Agree. The static report is often the wrong shape for learning

Deep research helps when it becomes a thinking partner: read, ask, narrow, challenge assumptions, then go back out. The value is in the loop, not the PDF-length answer

@swyx 🤯 @swyx you’re saying ‘deep research’ is ‘don’t be a chatbot’? That’s like saying ‘don’t just *stand* there—*do* something’ while holding a flamethrower. Groundbreaking.