[HN Gopher] AI assisted search-based research works now
___________________________________________________________________
AI assisted search-based research works now
Author : simonw
Score : 105 points
Date : 2025-04-21 14:15 UTC (8 hours ago)
(HTM) web link (simonwillison.net)
(TXT) w3m dump (simonwillison.net)
| simonw wrote:
| I think it's important to keep tabs on things that LLM systems
| fail at (or don't do well enough on) and try to notice when their
| performance rises above that bar.
|
| Gemini 2.5 Pro and o3/o4-mini seem to have crossed a threshold
| for a bunch of things (at least for me) in the last few weeks.
|
| Tasteful, effective use of the search tool for o3/o4-mini is one
| of those. Being able to "reason" effectively over long context
| inputs (particularly useful for understanding and debugging
| larger volumes of code) is another.
| skydhash wrote:
| One issue I can find with this workflow is tunnel vision,
| making ill informed decision because of the lack of surrounding
| information. I often skim books because even if I don't retain
| the content, I can have a mental map that can help me find
| further information when I need them. I wouldn't try to
| construct a complete answer to a question with just this amount
| of information, but I will use that map to quickly locate the
| source and have more information to synthesize the answer.
|
| One could use the above workflow in the same way and argues
| that natural language search is more intuitive than keyword
| based search. But I don't think that brings any meaningful
| productivity improvement.
|
| > _Being able to "reason" effectively over long context inputs
| (particularly useful for understanding and debugging larger
| volumes of code) is another._
|
| Any time I saw this "wish" pop up, my suggestion is to try a
| disassembler to reverse engineer some binary to really
| understand the problem of coming up with a theory of a program
| (based on Naur's definition). Individual statements are always
| clear (programming language are formal and have no ambiguity).
| The issue is grouping them, unambiguously define the semantic
| of these groups, and find the links between them, recursively.
|
| Once that's done, what you'll have is a domain. And you could
| have skipped the whole thing by just learning the domain from a
| domain expert. So the only reason to do this is because the
| code doesn't really implement the domain (bugs) or it's hidden
| purposefully. So the most productive workflow there is to learn
| the domain first to find discrepancy (first case) or focus
| yourself on the missing part (second case). In the first case,
| the easiest approach is writing tests, and the more complete
| one is to do a formal verification of the software.
| jsemrau wrote:
| My main observation here is
|
| 1. Technically it might be possible to search the Internet, but
| it might not surface correct and/or useful information.
|
| 2. High-value information that would make a research report
| valuable is rarely public nor free. This holds especially true in
| capital-intensive or regulated industries.
| simonw wrote:
| I fully expect one of the AI-related business models going
| forward to be charging subscriptions for LLM search tool access
| to those kinds of archives.
|
| ChatGPT plus an extra $30/month for search access to a specific
| archive would make sense to me.
| sshine wrote:
| Kagi is $10/mo. for search and +$15/mo. for premium LLMs with
| agentic access to search.
| AlotOfReading wrote:
| What they're talking about is access to professional
| archives like EBSCOnet or Bloomberg, which usually don't
| sell to individuals in the first place and start at tens of
| thousands of dollars per seat for institutional access.
| hadlock wrote:
| o3/o4 seem to know how to search things like pypi, crates.io,
| pkg.go.dev etc and apply those changes on the first try. My
| application (running on an older version of code) had a
| breaking change to how the event controller functioned in the
| newer version, o3 looked at the documentation and rewrote it to
| use the new event controller. It used to be that you were
| trapped with the LLM being 3-8 months behind on package
| versions.
| simonw wrote:
| Huh, now I'm thinking that maybe a target for release notes
| should be to provide enough details that a good LLM can be
| used to apply fixes for any breaking changes.
| rd wrote:
| MCP maybe? A release notes MCP (maybe into ReadTheDocs or
| pypi) that understands upgrade instructions for every
| package.
| sanderjd wrote:
| This is the thing I don't really love about MCP: Why
| should it require a separate protocol, rather than just
| good readable documentation?
| navinsylvester wrote:
| Ironically, https://context7.com/
| TrackerFF wrote:
| I'm not a researcher, but don't most researchers these days
| also upload their work to arXiv?
|
| Sure, it's not a journal - but in some fields (Machine
| Learning, Math) it seems like everyone also uploads their stuff
| there. So if the models can crawls sites like arXiv, at least
| there's some decent stuff to be found.
| btbuildem wrote:
| It's a relevant question about the economic model for the web. On
| one hand, the replacement of search with a LLM-based approach
| threatens the existing, advertising-based model. On the other
| hand, the advertising model has produced so much harm: literally
| irreparable damage to attention spans, outrage-driven
| "engagement", and the general enshittification of the internet to
| mention just a few. I find it a bit hard to imagine whatever
| succeeds it will be worse for us collectively.
|
| My question is, how to reproduce this level of functionality
| locally, in a "home lab" type setting. I fully expect the various
| AI companies to follow the exact same business model as any other
| VC-funded tech outfit: free service (you're the product) -> paid
| service (you're still the product) -> paid service with
| advertising baked in (now you're unabashedly the product).
|
| I fear that with LLM-based offerings, the advertising will be
| increasingly inseparable, and eventually undetectable, from the
| actual useful information we seek. I'd like to get a "clean"
| capsule of the world's compendium of knowledge with this amazing
| ability to self-reason, before it's truly corrupted.
| fzzzy wrote:
| You need a copy of r1 and enough ram to run it, and a web
| searching tool, or a rag database with your personal data
| store.
| qwertox wrote:
| I feel like the benefit which AI gives us programmers is limited.
| They can be extremely advanced, accelerative and helpful
| assistants, but we're limited to just that: architecting and
| developing software.
|
| Biologists, mathematicians, physicists, philosophers and the like
| seem to have an open-ended benefit from the research which AI is
| now starting to enable. I kind of envy them.
|
| Unless one moves into AI research?
| bluefirebrand wrote:
| I don't think AI is trustworthy or accurate enough to be
| valuable for anyone trying to do real science
|
| That doesn't mean they won't try though. I think the
| replication crisis has illustrated how many researchers
| actually care about correctness versus just publishing papers
| simonw wrote:
| If you're a skilled researcher I expect you should be able to
| get great results out of unreliable AI assistants already.
|
| Scientists are meant to be good at verifying and double-
| checking results - similar to how journalists have to learn
| to derive the truth from unreliable sources.
|
| These are skills that turn out to be crucial when working
| with LLMs.
| bluefirebrand wrote:
| > Scientists are meant to be good at verifying and double-
| checking results
|
| Verifying and double-checking results requires replicating
| experiments, doesn't it?
|
| > similar to how journalists have to learn to derive the
| truth from unreliable sources
|
| I think maybe you are giving journalists too much credit
| here, or you have a very low standard for "truth"
|
| You cannot, no matter how good you are, derive truth from
| faulty data
| simonw wrote:
| Don't make the mistake of assuming all journalists are
| the same. There's a big difference between an
| investigative reporter at a respected publication and
| someone who gets paid to write clickbait.
|
| Figuring out that the data is faulty is part of research.
| bluefirebrand wrote:
| Figuring out that data is faulty is one thing
|
| There is still no possible way that a journalist can
| arrive at correct information, no matter how good, if
| they only have faulty data to go with
| simonw wrote:
| That's what (good) journalism is: the craft of hunting
| down sources of information, figuring out how accurate
| and reliable they are and piecing tougher as close to the
| truth as you can get.
|
| A friend of mine is an investigative reporter for a major
| publication. They once told me that an effective trick
| for figuring out what's happening in a political story is
| to play different sources off against each other - tell
| one source snippets of information you've got from
| another source to see if they'll rebut or support it, or
| if they'll leak you a new detail because what you've got
| already makes them look bad.
|
| Obviously these sources are all inherently biased and
| flawed! They'll lie to you because they have an agenda.
| Your job is to figure out that agenda and figure out
| which bits are true.
|
| The best way to confirm a fact is to hear about it from
| multiple sources who don't know who else you are talking
| to.
|
| That's part of how the human intelligence side of
| journalism works. This is why I think journalists are
| particularly well suited to dealing with LLMs - human
| sources lie and mislead and hallucinate to them all the
| time already. They know how to get (as close as possible)
| to the truth.
| barbazoo wrote:
| Same with using AI for coding. I can't imagine someone
| having the expectation to use the LLM output verbatim but
| maybe I'm just not good enough at prompting.
| simonw wrote:
| Using AI for coding effectively involves getting _very_
| good at testing (both manual and automated) and code
| review:
| https://simonwillison.net/2025/Mar/2/hallucinations-in-
| code/...
| bluefirebrand wrote:
| Manual testing, automated testing, and code review
|
| All three of those things are things that software
| engineers rather reliably are bad at and cut corners on,
| because they are the least engaging and least interesting
| part of the job of building software
| simonw wrote:
| Yep. Engineers who aren't willing to invest in those
| skills will have limited success with AI-assisted
| development.
|
| I've seen a few people state that they don't like using
| LLMs because it takes away the fun part (writing the
| code) and leaves them with the bits they don't enjoy.
| bluefirebrand wrote:
| > Engineers who aren't willing to invest in those skills
|
| Are bad engineers
|
| > AI-assisted development
|
| Are also bad engineers
| parodysbird wrote:
| Biologists, mathematicians, physicists, and philosophers are
| already the experts who produce the text in their domain that
| the LLMs might have been trained on...
| sshine wrote:
| The article doesn't mention Kagi: The Assistant, a search-powered
| LLM frontend that came out of closed beta around the beginning of
| the year, and got included in all paid plans since yesterday.
|
| It really is a game changer when the search engine
|
| I find that an AI performing multiple searches on variations of
| keywords, and aggregating the top results across keywords is more
| extensive than most people, myself included, would do.
|
| I had luck once asking what its search queries were. It usually
| provides the references.
| simonw wrote:
| I haven't tried Kagi's product here yet. Do you know which LLM
| it uses under the hood?
|
| Edit: from https://help.kagi.com/kagi/ai/assistant.html it
| looks like the answer is "all of them":
|
| > Access to the latest and most performant large language
| models from OpenAI, Anthropic, Meta, Google, Mistral, Amazon,
| Alibaba and DeepSeek
| dcre wrote:
| Yep, regular paid Kagi sub comes with cheap models for free:
| GPT-4o-mini, Gemini 2.5 Flash, etc. If you pay extra you can
| get the SOTA models, though IMO flash is good enough for most
| stuff if the search result context is good.
| intended wrote:
| I find that these conversations on HN end up covering similar
| positions constantly.
|
| I believe that most positions are resolved if
|
| 1) you accept that these are fundamentally narrative tools. They
| build stories, In whatever style you wish. Stories of code,
| stories of project reports. Stories or conversations.
|
| 2) this is balanced by the idea that the core of everything in
| our shared information economy is Verification.
|
| The reason experts get use out of these tools, is because they
| can verify when the output is close enough to be
| indistinguishable from expert effort.
|
| Domain experts also do another level of verification (hopefully)
| which is to check if the generated content computes correctly as
| a result - based on their mental model of their domain.
|
| I would predict that that LLMs are deadly in the hands of people
| who can't gauge the output, and will end up driving themselves
| off of a cliff, while experts will be able to use it effectively
| on tasks where verification of the output has a comparative
| effort advantage, over the task of creating the output.
| gh0stcat wrote:
| You've perfectly captured my experience as well, I typically
| only trust it and have good experiences with LLMs when I have
| enough domain expertise to get to at least a 95% confidence the
| output is correct. (Specific to my domain of work, I don't
| always need "perfect"). I also can mostly use it as a first
| pass for getting the idea of where to begin research, after
| that I lose confidence that the more detailed and advanced
| content it is giving me is accurate. There is a gray area
| though where a domain expert might have a false sense of
| confidence, and over time experience "Skill Drift", where they
| lose expertise because they are only ever verifying a lossy
| compression of information, rather than re-setting their
| context with real world information. I am mostly concerned with
| that last bit.
| saulpw wrote:
| I tried it recently. I asked for videochat services like the one
| I use (WB) with 2 specific features that the most commonly used
| services don't have. It asked some clarifying questions and
| seemed to understand the mission, then went off for 10 minutes
| after which it returned 5 results in a table.
|
| The first result was WB, which I gave to it as the first example
| and am already using. Results 2 and 3 were the mainstream
| services which it helpfully marked in the table as not having the
| features I need. Result 4 looked promising but was discontinued 3
| years ago. Result 5 was an actual option which I'm trying out
| (but may not work for other reasons).
|
| So, 1/5 usable results. That was mildly helpful I guess, but it
| appeared a lot more helpful on the surface than it was. And I
| don't seem to have the ability to say "nice try but dig deeper".
| simonw wrote:
| That sounds like a Deep Research query, was that with OpenAI or
| Gemini?
| saulpw wrote:
| This was OpenAI.
| Gracana wrote:
| You _can_ tell it to try again. It took me a couple rounds with
| the tool before I noticed that your conversation after the
| initial research isn 't limited to just chatting: if you select
| the "deep research" button on your message, it will run the
| search process in its response.
| baq wrote:
| > I can feel my usage of Google search taking a nosedive already.
|
| Conveniently Gemini is the best frontier model for everything
| else, they're very interested and well positioned (if not best?)
| to also be the best in deep research. Let's check back in 3-6
| months.
| throwup238 wrote:
| IMO they're already the best. Not only is the rate limit much
| higher (20/day instead of OpenAI's 10/month) but Gemini is
| capable of looking at far more sources, on the order of 10x.
|
| I just had a research report last night that looked at 400
| sources when I asked it to help identify a first edition Origin
| of Species (it did a great job too, correctly explaining how to
| identify a true first edition from chimeral ones).
| jillesvangurp wrote:
| Google has two advantages:
|
| 1) Their AI models aren't half bad. Gemini 2.5 seems to be
| doing quite well relative to some competitors.
|
| 2) They know how to scale this stuff. They have their own
| hardware, lots of data, etc.
|
| Scaling is of course the hard part. Doing things at Google
| scale means doing it well while still making a profit. Most AI
| companies are just converting VC cash into GPUs and energy. VC
| subsidized AI is nice at a small scale but cripplingly
| expensive at a larger scale. Google can't do this; they are too
| large for that. But they are vertically integrated, build their
| own data centers, with their own TPUs, etc. So, once this
| starts happening at their scale, they might just have an
| advantage.
|
| A lot of what we are seeing is them learning to walk before
| they start running faster. Most of the world has no clue what
| perlexity is or any notion of the pros and cons of claude 3.7
| sonnet vs. o4 mini high. None of that stuff matters long term.
| What matters is who can do this stuff well enough for billions
| of people.
|
| So, I wouldn't count them out. But none of this stuff
| guarantees success either, of course.
| energy123 wrote:
| > The user-facing Google Gemini app can search too, but it
| doesn't show me what it's searching for.
|
| Gemini 2.5 Pro is also capable of search as part of its chain of
| thought but it needs light prodding to show URLs, but it'll do so
| and is good at it.
|
| Unrelated point, but I'm going to keep saying this anywhere
| Google engineers may be reading, the main problem with Gemini is
| their horrendous web app riddled with 5 annoying bugs that I
| identified as a casual user after a week. I assume it's in such a
| bad state because they don't actually use the app and they use
| the API, but come on. You solved the hard problem of making the
| world's best overall model but are squandering it on the world's
| worst user interface.
| loufe wrote:
| There must be some form of memory leak in AI Studio as I'll
| have to close and open a new tab after about 2 hours as it
| slowly grinds my slower computers to a halt. Its ability to
| create a markdown file without escaping the markdown itself
| (included code snippets) is definitely my first suggestion for
| them to fix.
|
| It's a great tool, but sometimes frustrating.
| oulipo wrote:
| The main "real-world" use cases for AI use for now have been:
|
| - shooting buildings in Gaza https://apnews.com/article/israel-
| palestinians-ai-weapons-43...
|
| - compiling a list of information on Government workers in US
| https://www.msn.com/en-us/news/politics/elon-musk-s-doge-usi...
|
| - creating a few losy music videos
|
| I'd argue we'd be better off SLOWING DOWN with that shit
| esafak wrote:
| Programming is not real world?
| das_keyboard wrote:
| Yeah right. We also got "vibe coding" out of it.
| swyx wrote:
| > Deep Research, from three different vendors
|
| dont forget Xai grok!
| M4v3R wrote:
| Which, at least in my experience is surprisingly good while
| being much faster than others.
| fudged71 wrote:
| you.com is surprisingly good for this as well (I like the
| corporate report PDF export)
| softwaredoug wrote:
| I wonder when Google search will let me "chat" with the search
| results. I often want to ask the AI Overview follow up questions.
|
| I secondarily wonder how an LLM solves the trust problem in Web
| search. What's traditionally solved (and now gamed) through
| PageRank. It doesn't seem ChatGPT is easily fooled by Spam as
| direct search.
|
| How much is Bing (or whatever the search engine is) getting
| better? vs how much are LLMs better at knowing what a good result
| is for a query?
|
| Or perhaps it has to do with the richer questions that get asked
| to chat vs search?
| vunderba wrote:
| > I wonder when Google search will let me "chat" with the
| search results.
|
| You don't hear a lot of buzz around them, but thats kind of
| what Perplexity lets you do. (Possibly phind too but it's been
| a while since I used them).
| dingnuts wrote:
| >I wonder when Google search will let me "chat" with the search
| results
|
| Kagi has this already, it's great. Choose a result, click the
| three-dot menu, choose "Ask questions about this page." I love
| to do this with hosted man pages to discover ways to combine
| the available flags (and to discover what is there)
|
| I find most code LLMs write to be subpar but Kagi can
| definitely write a better ffmpeg line than I can when I use
| this approach
| KTibow wrote:
| When AI Overview was called Search Generative Experience, you
| could do that. You can do that again now if you have access to
| AI Mode.
| 63 wrote:
| One downside I found is that the llm cannot change its initial
| prompt until it's done thinking. I used deep research to compare
| counseling centers for me but of course when it encounters some
| factor I hadn't thought of (e.g. the counselors here fit the
| criteria perfectly but none accept my insurance), it doesn't know
| that it ought to skip that site entirely. Really this is a
| critique of the deep-research approach rather than search in
| general, but I imagine it can still play out on smaller scales.
| Often, searching for information is a dynamic process involving
| the discovery of unknown unknowns and adjustment based on that,
| but ai isn't great at abstract goals or stopping to ask
| clarifying questions before resuming. Ultimately, the report I
| got wasn't useless, but it mostly just regurgitated the top 3
| google results. I got much better recommendations by reaching out
| to a friend who works in the field.
| xp84 wrote:
| From article:
|
| > "Google is still showing slop for Encanto 2!" (Link is
| provided)
|
| I believe quite strongly that Google is making a serious misstep
| in this area, the "supposed answer text pinned at the top above
| the actual search results."
|
| For years they showed something in this area which was directly
| quoted from what I assume was a shortlist of non-BS sites so
| users were conditioned for years that if they just wanted a
| simple answer like when a certain movie came out or if a certain
| show had been canceled or something, you may as well trust it.
|
| Now it seems like they have given over that previous real estate
| to a far less reliable feature, which simply feeds any old
| garbage it finds anywhere into a credulous LLM and takes whatever
| pops out. 90% of people that I witness using Google today simply
| read that text and never click any results.
|
| As a result, Google is now pretty much always even less accurate
| at the job of answering questions than if you posed that same
| question to ChatGPT, because GPT seems to be drawing from its
| overall weights which tend toward basic reality, whereas Google's
| "Answer" seems to be summarizing a random 1-5 articles from the
| Spam Web, with zero discrimination between fact, satire, fiction,
| and propaganda. How can they keep doing this and not expect it to
| go badly?
| CSMastermind wrote:
| The various deep research products don't work well for me. For
| example I asked these tools yesterday, "How many unique NFL
| players were on the roster for at least one regular season game
| during the 2024 season? I'd like the specific number not a
| general estimate."
|
| I as a human know how to find this information. The game day
| rosters for many NFL teams are available on many sites. It would
| be tedious but possible for me to find this number. It might take
| an hour of my time.
|
| But despite this being a relatively easy research task all of the
| deep research tools I tried (OpenAI, Google, and Perplexity)
| completely failed and just gave me a general estimate.
|
| Based on this article I tried that search just using o3 without
| deep research and it still failed miserably.
| simonw wrote:
| That is an excellent prompt to tuck away in your back pocket
| and try again future iterations of this technology. It's going
| to be an interesting milestone when or if any of these systems
| get good enough at comprehensive research to provide a correct
| answer.
| danielmarkbruce wrote:
| This is just a bad match to the capabilities. What you are
| actually looking for is analysis, similar in nature to what a
| data scientist may do.
|
| The deep research capabilities are much better suited to more
| qualitative research / aggregation.
| pton_xd wrote:
| > The deep research capabilities are much better suited to
| more qualitative research / aggregation.
|
| So it's just sentiment analysis at-scale, basically? Not
| really that interesting or useful compared to doing actual
| quantitative analysis. And usually qualitative analysis is
| borne out of data, so lacking that you can't even verify if
| the output is correct.
___________________________________________________________________
(page generated 2025-04-21 23:01 UTC)