from Hacker News

Ask HN: How to Use "Deep Research"?

by muddi900 on 4/6/25, 6:17 PM with 17 comments

The issue I have had with llms since the start has been the slop. Contrary to popular belief, the open web was mostly slop before ChatGPT. We just used to call it SEO blogspam. And all llms are trained on them.

So when I tried google's Deep Research in Gemini, I ran into the same problem. It was basically a typical llm chat response, but a lot more verbose and citations to the same blogspam listicles that makes regular human research difficult.

How do I avoid this pitfall? So what I am asking is, how do I use deep research?

  • by addaon on 4/6/25, 9:34 PM

    Use it for what? If your goal is understanding… probably don’t. If you’re trying to produce a document for an audience that doesn’t really care and will just use an LLM to summarize it without any fact checking, then give the minimal prompting to get the token count you need. This is where LLMs excel —- sequences of tokens that are plausible use of human languages where the cost for verifying if they have semantic value exceeds five or ten minutes.

  • by GianFabien on 4/6/25, 10:56 PM

    My naive, yet practical, view is that all LLMs produce output that reflects what was most commonly found on the internet. So if your area of interest has a lot of slop, then that is exactly what Deep Research will return.

    So if you want to do "real deep research" into some topic, then you start with the various summaries and drill down. I use Google Scholar to locate papers, then I search for the cited papers. It is tedious, but the only way that I know which avoids the pitfalls of random regurgitation of easily found materials.

    Currently AGI is marketing hype. Critical thinking with real understanding, i.e. knowledge building hasn't yet been automated. You need to put in the intellectual effort.

  • by solardev on 4/6/25, 9:17 PM

    What topic or field are you researching? Is it something academic, programming related, or...?

    It generally helps if you ask it to focus on specific kinds sources (gov sites, academic, etc.) but it really depends on the topic. Easy for some topics (the general sciences) but not so easy for other ones (popular culture stuff or current events).

  • by leonidasv on 4/7/25, 1:12 PM

    For market research, Exa.ai may be a better option. It's a search engine that crawls the web and index pages in LLM embeddings. For my searches, it delivers good results that are not found in Google's first page because the SEO spam is burying it. You should give it a try.

  • by beefnugs on 4/7/25, 12:27 AM

    Hire an expert in the field who can create 100 sentences of "dont do this... dont do that..." then create your own business as a middleman who keeps this list a top secret and supplies the service to others.

    Then in a couple months go out of business when someone tricks the AI to spill its prompt

  • by tmaly on 4/9/25, 8:41 PM

    The one thing I had to do is ask for a References section at the end of the research. You can then take that and go through it to check things.

  • by Boristoledano on 4/7/25, 2:11 PM

    Integrate our web search agent (linkup.so) which has been trained to cut through the noise :)

  • by transformi on 4/7/25, 9:54 AM

    Searching for free coupons :)

  • by dangus on 4/7/25, 12:40 AM

    Do actual research like a researcher instead of being a lazy bum with LLMs?