You hear about this new programming language called "Frob", and you assume it must have a website. So you google "Frob language". You hear that there was a plane crash in DC, and assume (CNN/AP/your_favorite_news_site) has almost certainly written an article about it. You google "DC plane crash."

LLMs aren't ever going to replace search for that use case, simply because they're never going to be as convenient.

Where LLMs will take over from search is when it comes to open-ended research - where you don't know in advance where you're going or what you're going to find. I don't really have frequent use cases of this sort, but depending on your occupation it might revolutionize your daily work.

Just yesterday I was trying to remember the name of a vague concept I’d forgotten, with my overall question being:

“Is there a technical term in biology for the equilibrium that occurs between plant species producing defensive toxins, and toxin resistance in the insect species that feed on those plants, whereby the plant species never has enough evolutionary pressure to increase it’s toxin load enough to kill off the insect that is adapting to it”

After fruitless searching around because I didn’t have the right things to look for, putting the above in ChatGPT gave an instant reply of exactly what I was looking for:

“Yes, the phenomenon you're describing is often referred to as evolutionary arms race or coevolutionary arms race.”

Evolutionary arms race is somewhat tautological; an arms race is the description of the selective pressure applied by other species on evolution of the species in question. (There are other, abiotic sources of selective pressures, e.g. climate change on evolutionary timescales, so while 'evolution' at least carries a broader meaning, 'arms race' adds nothing that wasn't already there.)

That said, using your exact query on deepseek r1 and claude sonnet 3.7 both did include red queen in their answers, along with other related concepts like tit for tat escalation.

Firstly, "Evolutionary Arms Race" is not tautological, it is a specific term of art in evolutionary biology.

Secondly, "evolutionary arms race" is a correct answer, it is the general case of which the Red Queen hypothesis is a special case. I do agree with you that OP described a Red Queen case, though I would hesitate to say it was because of "equilibrium"; many species in Red Queen situations have in fact gone extinct.

https://en.wikipedia.org/wiki/Evolutionary_arms_race

https://en.wikipedia.org/wiki/Red_Queen_hypothesis

At any rate, almost NONE of these actual terms of art are about the sort of equilibrium that was the exact heart of the OP's query to the LLM, and thus nearly none of the broader umbrella 'arms race' is about why the plant doesn't have the evolutionary pressure to actually drive the parasite extinct. An arms race doesn't have to be in equilibrium. Armor vs weapons were in an arms race and indeed at equilibrium for millenia, but then bullets come along and armor goes exinct almost overnight and doesn't reappear for 5 centuries. Bullets win the arms race. Arms races have nothing to do, inherently, with equilibrium.

You seem to have misunderstood the nature of the equilibrium in a Red Queen scenario, which is the fundamental effect that the hypothesis is directly named for. That species that are in Red Queen relationships can go extinct is in no way a counterargument to the idea that two (or more) species tend to coevolve in such a way that the relative fitness of each (and of the system as a whole) stays constant. See, for example, the end of the first paragraph on the origin of Van Valen's term at your own wiki link.

Evolutionary steady-state is a synonymous term without the baggage of the literary reference and also avoids the incorrect connotation suggested by arms race that leads people to forget the abiotic factors that are often a dominant mechanism in extinctions as the realized niche vs the fundamental niche differ. Instead, Van Valen was specifically proposing the Red Queen hypothesis as an explanation of why extinction appears to be a half-life, i.e. of a constant probability, rather than a rate that depends on the lifetime of the taxa. This mechanism has good explanatory power for the strong and consistent evidence that speciation rate (usually considered as the log of the number of genera, depending on definition, see Stanley's Rule) has a direct and linear relation with the extinction rate. If Red Queened species didn't go exinct, Van Valen wouldn't have needed to coin the term to explain this correlation.

Or were you deliberately invoking Cunningham's Law?

GP was looking for a specific term that they had heard before. It was co-/evolutionary arms race, and ChatGPT guessed it correctly.

Also GPT-4o elaborated the answer (for me at least) with things like:

    > However, the specific kind of equilibrium you're referring to—where neither side ever fully "wins", and both are locked in a continuous cycle of adaptation and counter-adaptation—is also captured in the idea of a “Red Queen dynamic”.

    > You could refer to this as:
      * Red Queen dynamics in plant-insect coevolution
      * A coevolutionary arms race reaching a dynamic equilibrium
      * Or even evolutionary stable strategies (ESS) applied to plant-herbivore interactions, though ESS is more game-theory focused.

Or updated for the LLM age, "the best way to get the right answer from an LLM is not to ask it a question and use its answer; it's to post its response on a site of well-educated and easily nerdsniped people"

Time to pop some popcorn and hit refresh.

They’re very helpful for helping me ask more refined questions by getting the terminology correct.

I think of AI as an intelligent search engine / assistant and, outside of simple questions with one very specific answer, it just crushes search engines.

I use the LLMs to find the right search terms, and that combination makes search engines much more useful.

LLM by themselves give me very superficial explanations that don't answer what I want, but they are also a great starting point that will eventually guide me to the answers.

* Gemini lets you do this by actually streaming your screen through and verbally chatting about whatever is on screen. While interesting, I find the responses to be a little less thorough. YMMV.

Even with supposedly authoritative peer-reviewed research papers it is extremely frequent to find errors whenever the authors claim to quote earlier work, because the reality is that most of them do not bother to read carefully their claimed bibliography.

When you get an answer from an AI, the chances greatly increase that the answer regurgitates some errors present in the publications used for training. At least when you get the answer from a real book or research paper, it lists its sources and you can search them to find whether they have been reproduced rightly or wrongly. With an AI-generated answer it becomes much more difficult to check it for truthfulness.

I will give an example of what I mean, on which I happened to stumble today. I have read a chemistry article published in 2022 in a Springer journal. While the article also contained various useful information, it happened to contain a claim that seemed suspicious.

In 1782, the French chemist Guyton de Morveau has invented the word "alumine" (French) = "alumina" (Latin and English), to name what is now called oxide of aluminum, which was called earth of alum at that time ("terra aluminis" in Latin).

The article from 2022 claimed that the word "alumina" had already been used earlier with the same sense, by Andreas Libavius in 1597, who has been thus the creator of this word.

I have found this hard to believe, because the necessity for such a word has appeared only during the 18th century, when the European chemists, starting with the Swedish chemists, have finally gone beyond the level of chemical classification inherited from the Arabs and they have begun to classify all known chemical substances as combinations of a restricted set of primitive substances.

Fortunately, the 2022 article had a detailed bibliography, and using it I was able to find the original work from 1597 and the exact paragraph in it that was referred to. The claim of the 2022 article was entirely false. While the paragraph contained a word "alumina", that was not a singular feminine adjective (i.e. agreeing with "terra") referring to the "earth of alum". Instead of this, it was not a new word, but just the plural of the neuter word "alumen" (= English alum), in the sentence "alums or salts or other similar sour substances can be mixed in", where "alums" meant "various kinds of alum", like "salts" meant "various kinds of salt". Nowhere in the work of Libavius there was any mention of an earth that is a component of alum and that could be extracted from alum (in older chemistry, "earth" was the term for any non-metallic solid substance that neither dissolves in water nor burns in air).

I have given in detail this example, in order to illustrate the kinds of errors that I very frequently encounter whenever some authors claim to quote other works. While this was an ancient quotation, lots of similar errors appear when quoting more recent publications, e.g. when quoting Einstein, Dirac or the like.

I am pretty sure that if I would ask an AI assistant something, the number of errors in the answers will not be less than when reading publications written by humans, but the answers will be more difficult to verify.

Whoever thinks that they can get a quick answer to any important question in a few seconds and be done with it, are naive because the answer to any serious question must be verified thoroughly, otherwise there are great chances that those who trust such answers will just spread more disinformation, like the sources on which the AI has been trained.

Despite a lot of effort, I'm just not a highly skilled developer and I don't have any friends / colleagues I can turn to for assistance (I don't know a single software developer or even another person who enjoys video games). While resources like StackOverflow are certainly useful, having answers tailored to my specific situation really accelerates progress.

I'm not trying to cure cancer here and much of what would be considered the "best approach" for a small game architecture is unique to the developer. As such, AI is an incredible resource to lean on and get information tailored to my unique use case (here is my code...how does {topic} apply to my situation?")

And yes, I find errors from time to time, but that is good. It keeps me on my toes and forces me to really understand the response / perspective.

Google 55% as GPT is not a local search engine

GPT 45% but use it for more intelligent learning/conversations/knowledgebase.

If I had a GPT phone ... sorta like H.E.R. the movie I would rarely leave my phone's lockscreen. My AI device / super AI human friend would do everything for me including get me to the best lighting to take the best selfies...

For example: Take the ingredient list of a cosmetic or other product that could be 30-40 different molecules and ask ChatGPT to list out what each of them is and if any have potential issues.

You can then verify what it returns via search.

You can criticize LLMs all you want, but the fact is that they provide value to people in ways that alternatives simply don’t. The energy consumption is a concern, but don’t pretend there are viable alternatives when there aren’t.

The LLM people are heavily invested in ever bigger models to keep the research money flowing in, it wouldn't make sense to release a service that undercuts that.

that leaves independent actors - presumably building and maintaining an up to date database is difficult, so only the big search engines do.

LLMs store embeddings of individual tokens (usually parts of words), so a result of an actual search will be top-k embeddings and the corresponding tokens, similar to the output of a Google search. You could extract the initial matrix of embeddings from some open-weights model and find tokens closest to your query. However, it's not clear why do this. OP got coherent text, so that's not search.

It's _similar_, though, because attention in LLMs basically looks for most similar tokens. So to answer the question about the term, the LLM had to create a stream of tokens that's semantically closest to the given description. Well, this is somewhat like a search, but it's not exactly the same.

The reason is pretty simple. If the result you want is in the first few search hits, it's always better. Your query is shorter so there is less typing, the search engine is always faster, the results are far better because you side step the LLM hallucinating as it regurgitates the results it remembers on the page your would have read if you searched.

If you aren't confident of the search times, it can take 1/2 an hour of dicking around with different terms, clicking though a couple of pages of search results for each set of term, until you finally figure out the lingo to use. Figuring out what you are really after from that wordy description is the inner magic of LLM's.

Most often not true in the kind of searches I do. Say, I search for how to do something in the Linux terminal (not just the command, but the specific options to achieve a certain thing). Google will often take me to pages that do have the answer, but are full of ads and fluff, and I have to browse through several options until I find the ones I want. ChatGPT just gives me the answer.

And with any halfway decent model, hallucination only seems to be a problem in difficult or very specialized questions. Which I agree shouldn't be asked to LLMs (or not without verifying sources, at least). But over 90% of what I search aren't difficult or specialized questions, they're just things I have forgotten, or things that are easy but I don't know just because they're not in my area of expertise. For example as a learner of Chinese, I often ask it to explain sentences to me (translate the sentence, the individual words, and explain what a given word is doing in the sentence) and for that kind of thing it's basically flawless, there's no reason why it would hallucinate as such questions are trivial for a model having tons of Chinese text in its training set.

I asked Claude to give me a recipe that uses mushrooms and freezes well and it give me a decent looking soup recipe. It might not be the best soup ever, but it's soup, kinda hard to mess up. The alternative would be to get a recipe from the web with a couple dozen paragraphs about how this is the bestest soup ever and it comes from their grandma and reminds them of summer and whatnot.

Interesting, I just random words. LLM not care sentence.

But what I'm talking about is when I want to read the page for myself. Waste of time to have to wait for an LLM to chew on it.

Really, for many “page searches”, a good search engine should just be able to take you immediately to the page. When I search “Tom Hanks IMDB”, there’s no need to see a list of links - there’s obviously one specific page I want to visit.

https://notes.npilk.com/custom-search

Are you feeling lucky?

Unfortunately you can’t really show ads if you take someone directly to the destination without any interstitial content like a list of links…

Grok is great for finding details and background info about recent news, and of course it's great for deep-diving on general knowledge topics.

I also use Grok for quick coding help. I prefer to use AI for help with isolated coding aspects such as functions and methods, as a conversational reference manual. I'm not ready to sit there pretending to be the "pilot" while AI takes over my code!

For the record, I do not like Google's AI generated results it spams at me when I search for things. I want AI when I choose to use AI, not when I don't choose it. Google needs a way to switch that off on the web (without being logged in).

I know what I'm looking for. I just need exact URL.

Perplexity miserably fails at this.

Traditional search is dead, semantic search through AI is alive and well.

I can't yet count once AI misunderstood the meaning of my search while Google loves to make assumptions, rewrite my search query, and deliver the results that pay it the best which have the best ads (in my opinion as a lifetime user).

Lets not even mention how they willingly accept misleading ads atop the results which trick the majority of common users into downloading malware and adware on the regular.

The reason Google is still seeing growth (in revenue etc.) is that for a lot 'commercial' search still ends with this kind of action.

Take purchasing a power drill for example, you might use an LLM for some research on what drills are best, but when you're actually looking to purchase you probably just want to find the product on Home Depot/Lowe's etc.

Ad-sponsored models are going to be dead as soon as people realize they can't trust output.

And because the entire benefit to LLM search is the convenience of removing a human-in-the-loop step (scanning the search results), there won't be a clear insertion/distinction point for ads without poisoning the entire output.

Those who can afford, buy unbiased.

Those who cannot, accept biased free services.

I suppose that's Google's hope for how future search turns out.

Over time subscription models will converge to subscription with advertisements. Like newspapers did.

Ads xor payment, or else you can fuck all the way off.

What you don't get with pay TV

  Originally for cable, but then you got ads with cable. 
  Then for streaming, but then you got ads with streaming. 
  Physical DVDs? Ads. 
  Paid Roku device? Ads. 
  Paid windows installation? Ads.

Example from my work. Many of our customers will search for our company name in order to find the login page. I've watched them do this over screen share.

When they do that, the top search result is an ad slot. The second search result is our login page.

We buy ads against out own company name so that we can be the ad slot as well as the first result. Otherwise a competitor who buys ads against our company name could be the top slot.

What? On Planet Earth, this is already a thing.

Kind of like a manual, with an index.

RTFM people.

Sounds trivial to integrate an LLM front end with a search engine backend (probably already done), and be able to type "frob language" and it gives you a curated clickable list of the top resources (language website, official tutorial, reference guide, etc) discarding spam and irrelevant search engine results in the process.

https://news.ycombinator.com/item?id=9224

The LLM could "intelligently" pick from the top several pages of results, discard search engine crap results and spam, summarize each link for you, and so on.

We don't have that now (or for 30 years - I should know, I was there, using Yahoo!, and Altavista, and Lycos and such back in the day).

Instead of the core of the answer coming from the LLM, it could piece together a few relevant contexts and just provide the glue.

How do you know the media isn't lying to you ? It's happened many times before (think pre-war propaganda)

Odds are pretty good that, at least for not very popular projects, the homepage's themselves would soon be produced by some LLM, and left at that, warts and all...

> In responding to user queries, Grok has a unique feature that allows it to decide whether or not to search X public posts and conduct a real-time web search on the Internet. Grok’s access to real-time public X posts allows Grok to respond to user queries with up-to-date information and insights on a wide range of topics.

Other considerations:

- Visiting the actual website, you’ll see the programming languages logo. That may be a useful memory aide when learning.

- The real website may have diagrams and other things that may not be available in your LLM tool of choice (grok).

- The ACT of browsing to a different web page may help some learners better “compartmentalize” their new knowledge. The human brain works in funny ways.

- i have 0 concerns of a hallucination when readings docs directly from the author/source. Unless they also jumped on the LLM bandwagon lol.

Just because you have a hammer in your hand doesn’t mean you should start trying to hammer everything around you friend. Every tool has its place.

For some cases I absolutely prefer an LLM, like discoverability of certain language features or toolkits. But for the details, I'll just google the documentation site (for the new terms that the LLM just taught me about) and then read the actual docs.

I'm hard pressed to construction an argument where, with widely-accessible LLM/LAM technology, that still looks like:

   1. User types in query
   2. Search returns hits
   3. User selects a hit
   4. User looks for information in hit
   5. User has information

F.ex. with the API example, why doesn't your future IDE directly surface the API (from its documentation)? Or your future search directly summarize exactly the part of the API spec you need?

  Language Implementation Patterns (the book) |> Analyzing Languages (the part) |> Tracking and Identifying Program Symbols (the chapter) |> Resolving Symbols (the section)

  Unit Testing: Principles, Practices,and Patterns (the book) |> Making your tests work for you (the part) |> Mocks and test fragility (the chapter) |> The relationship between mocks and test fragility (the section) |> Intra-system vs. inter-system communications

  Python 3.13.3 Documentation (docs.python.org) |> The Python Standard Library |> Text Processing Services |> string

And if you are reading technical docs, especially good ones, each word is there for a reason. LLM throw some that information away, but they don't have your context to know if the stuff they throw away is useful or not. The text the summary omitted may likely contain an important caveat or detail you really should have known before starting to use that API.

I recently configured Chrome to only use google if I prefix my search with a "g ".

I don't like LLMs for two reasons:

* I can't really get a feel for the veracity of the information without double checking it. A lot of context I get from just reading results from a traditional search engine is lost when I get an answer from a LLM. I find it somewhat uncomfortable to just accept the answer, and if I have to double check it anyways, the LLM's answer is kind of meaningless and I might as well use a traditional search engine.

* I'm missing out on learning opertunities that I would usually get otherwise by reading or skimming through a larger document trying to find the answer. I appreciate that I skim through a lot of documentation on a regular basis and can recall things that I just happened to read when looking for a solution for another problem. I would hate it if an LLM would drop random tidbits of information when I was looking for concrete answers, but since its a side effect of my information gathering process, I like it.

If I were to use an AI assistant that could help me search and curate the results, instead of trying to answer my question directly. Hopefully in a more sleek way than Perplexity does with its sources feature.

At least that has been my experience. I admit I don't use LLMs very much.

1. Rules that get prefixed in front of your prompt as part of the real prompt ChatGPT gets. Like what they do with the system prompt.

And

2. Some content makes your prompt too big for the context windows where the rules get cut off.

Then, it might help to measure the tokens in the overall prompt, have a max number, and warn if it goes over it. I had a custom, chat app that used their API's with this feature built in.

Another possibility is, when this is detected, it asks you if you want to use one with a larger, context window. Those cost more. So, it would be presented as an option. My app let me select any of their models to do that manually.

I really like not being complimented on literally everything with a wall of text anymore.

Also, if that works, why doesn't copilot/cursor write lots of excessive code mixed with lots of prose only to distill it later?

The “thinking” models are really verbose output models that summarise the thinking at the end. These tend to outperform non-thinking models, but at a higher cost.

Anthropic lets you see some/all of the thinking so you can see how the model arrived at the answer.

This is my main reason for not using LLMs as a replacement for search. I want an accurate answer. I quote often search for legal or regulatory issues, health, scientific issues, specific facts about lots of things. i want authoritative sources.

You check the information you decide should be verified.

An LLM response without explicit mention of its provenance... There's no way to even guess whether it is authoritative.

Actually, it might be fully unbounded even for an n of 1.

What do you even use for double-check? Some random low-quality content farm? A glitchy LLM? An dodgy mirror of official docs full of ads? Or do you actually dig the source code for this?

And do you keep double-checking with all other information on the page... "A TOMLDecodeError will be raised on an invalid TOML document." - are you going to start an interactive session and check which error will be raised?

Just because you can find multiple independent sources saying the same thing doesn't mean it's correct.

In all honesty doing this for news and such brings me comfort. Because the truth is usually pretty vanilla.

"What I tell you three times is true"

Part of why I prefer to use a search engine is that I can see who is saying it, in what context. It might be Wikipedia, but also CIA world fact book. Or some blog but also python.org.

Or (lately) it might be AI SEO slop, reworded across 10 sites but nothing definitive. Which means I need to change my search strategy.

I find it easier (and quicker) to get to a believable result via a search engine than going via ChatGPT and then having to check what it claims.

And this is how LLMs perform when LLM-rot hasn't even become widely pervasive yet. As time goes on and LLMs regurgitate into themselves, they will become even less trustworthy. I really can't trust what an LLM says, especially when it matters, and the more it lies, the more I can't trust them.

Really, these days, either I know some resource exists and I want to find it, in which case a search engine makes much more sense than an LLM which might hallucinate, or I want to know if something is possible / how to do it, and the LLM will again hallucinate an incorrect way to do it.

I've only found LLMs useful for translation, transcription, natural language interface, etc.

LLMs have mostly been useful for three things: single line code completion (in GoLand), quickly translating JSON, and generating/optimizing marketing texts.

I use LLMs as a sounding board. Often if I'm trying to tease out the shape of a concept in my head, it's best to write it out. I now do this in the form of a question or request for information and dump it into the LLM.

"Search" can mean a lot of things. Sometimes I just want a website but can't remember the URL (traditional); other times I want an answer (LLMs); and other times, I want a bunch of resources to learn more (search+LLMs).

Bad: summarizing scientific research or technical data

Great: finding travel ideas or clarifying aspects of a franchise's fictional universe.

  - Long reads when I just want to know the conclusion
  - Reading terms of services that I would otherwise have blindly accepted
  - Summarizing key requirements in large tender documents prior to bidding on work

Especially if I'm looking for a small fact buried in the first results.

Instead I use a search engine and do my own reading and filtering. This way I learn what I'm researching, too, so I don't fall into the vicious cycle of drug abu ^H^H^H^H^H laziness. Otherwise I'll inevitably rely more on more on that thing, and be a prisoner of my own doing by increasingly offloading my tasks to a black box and be dependent on it.

Google recently (unrequested) provided me with very detailed AI generated instructions for server config - instructions that would have completely blown away the server. There will be someone out there who just follows the bouncing ball, I hope they've got good friends, understanding colleagues, and good backups!

What a weird sentence. What accuracy guarantees does Kagi have? Or, if you're not "offloading your brain to it", can't you do the same with an LLM?

Moreover, Kagi is a paid service. It has no ads, no hidden ranking, nothing to earn money by manipulating you. On the contrary you, the user, can add filters and ranking modifiers to promote the sites you find to be useful/truthful and demote others which push slop and SEO optimized content to your eyeballs. This is per user, and is not meddled with.

This makes Kagi very deterministic (unlike LLMs), very controllable (unlike LLMs), and very personalized (unlike LLMs). Moreover, Kagi gives you ~20 results or so per search, and no fillers (again, unlike LLMs).

I don't use Kagi's AI assistance features, and I don't pay for the "assistant" part of it, either.

I don't offload my brain to Kagi, because I don't prompt it until it gives me something I like. Instead, I get the results, read them, learn what I'm looking for, and possibly document what I got out from that research. This usage pattern, is again very different than prompting an LLM until it gives you something somewhat works or sounds plausible.

I do the hard work of synthesizing and understanding the answer. Not reading some slop and accepting it at face value.

Similarly, I don't offload my brain to LLMs.

> I do the hard work of synthesizing and understanding the answer. Not reading some slop and accepting it at face value.

Again, it's not necessary to accept LLM output at face value.

Use tools, think for yourself, sure. This applies to various tools: Kagi, LLMs, and others. None of these give you "accuracy guarantees". You usually have to think for yourself.

My favourite example of a situation where you don't have to think for yourself is asking an LLM to implement a function in a very strongly typed language. There only is one implementation of `a -> a`. For `(a -> b) -> List a -> List b`, you could return an empty list instead of performing map. There aren't that many implementations of `(a -> b -> b) -> b -> List a -> b` (three as far as I can see: left/right fold and a function which just returns the accumulator). It's easier to verify the LLM solution than to implement it yourself!

As for AI search, I do find it extremely useful when I don't know the right words to search for. The LLM will instantly figure out what I'm trying to say.

And the ratio between using search engine and Kagi’s LLM agent with search is still 70% search. Sometimes, searching is faster, sometimes asking AI is faster.

I use LLM-s for what they are good at, generative stuff. I know some task take me a long time and I can shortcut with LLM-s easily.

So here's a ChatGPT example query* which is completely off:

https://chatgpt.com/share/67f5a071-53bc-8013-9c32-25cc2857e5...

* It's intentionally bad be able to compare with Google.

And here's the web result, which is spot on:

https://imgur.com/a/6ELOeS1

LLM's are great when you want AN answer, and not get side tracked.

Search is great when you want to know what answers are out there. The best example is Recipes... From what spices go into chai to the spice mix in any given version of chili (let's not start on beans).

The former is filling in missing knowledge the latter is learning.

https://imgur.com/a/boNS2YZ

https://chatgpt.com/share/67f5a9f9-f0a8-800d-9101-aafb88e455...

which I think is way better than google.

Google offered me a few hits with existing businesses, with ChatGPT I need to do another query.

Out of curiosity I tried it and it did take me to a wholesale company (single result), but the Google results are better with cheaper options (multiple good results), I can also parse the list faster with my eye.

Sure, I can just write a better prompt:

https://chatgpt.com/share/67f5b09b-c154-8013-840f-934af8302f...

This is my third attempt to get it right, but it found me one which I haven't seen before. However I would still do a Google search to be thorough and get the best deal.

So yeah, I do still use search engines, specifically Kagi and (as a fallback) DuckDuckGo. From either of them I might tack on a !g if I'm dissatisfied with the results, but it's pretty rare for Google's results to be any better.

When I do use an LLM, it's specifically for churning through some unstructured text for specific answers about it, with the understanding that I'll want to verify those answers myself. An LLM's great for taking queries like "What parts of this document talk about $FOO?" and spitting out a list of excerpts that discuss $FOO that I can then go back and spot-check myself for accuracy.

Ex. https://www.google.com/search?udm=14&q=your%20query

DuckDuckGo also has an option where you can turn off the AI search so that you don't have to specify every time. I've found DDG sometimes gives me better results than Google and sometimes doesn't.

For example Jeep consistently lands at the bottom of the reliability ratings. Try asking GPT if Jeeps are reliable. The response reads like Jeep advertising.

My impression is that different llms are more or less people pleasing. I found grok is more willing to tell me something is a bad idea.

Looking at the reasoning traces for for the new reasoning models you can actually see how fine tuning is moving toward having models list the assumptions around data sources, which should be trusted, list multiple perspectives and then summarize, resulting in better answers. You can do that today with non-reasoning models, but you need to prompt engineer it to ask for that explicitly. This process of identifying not just extant content, but teaching systems how to approach problem analysis (instruction tuning, reasoning traces, etc ...) will be key to influencing how the models work and increasingly how they are differentiated.

In general, the models lean towards being Yes-Men on just about every topic, including things without official sources. I think this is a byproduct of them being trained to be friendly and agreeable. Nobody wants a product that's rude or contrarian, and this puts a huge finger on the scale. I imagine an a model unfiltered for safety and attitude and political correctness would have less of this bias (but perhaps more of other biases)

I very much prefer more disagreeable, critical models. GPT 4o and o3-mini will sometimes not tell you that you, e.g., didn't attach a file you asked to be analyzed and will instead hallucinate its contents, presumably not to upset you. Of course, their hallucinations are way more annoying.

https://chatgpt.com/share/67f57459-2744-8009-a94e-3b67dce8fd...

“[Jeeps] often score below average in reliability rankings from sources like Consumer Reports and J.D. Power.”

https://g.co/gemini/share/b5e5ea80548b

Seems entirely reasonable to me. Didn't have to trick it into providing citations.

If you want to know how modern Jeep models stack up against their peers in terms of reliability, try asking GPT that question!

Our current LLM are kneecapped because they are very reluctant to be negative.

For me, searches fall into one of three categories, none of which are a good fit for LLMs:

1. A single business, location, object, or concept (I really just want the Google Maps or Wikipedia page, and I'm too lazy to go straight to the site). For these queries, LLMs are either overkill or outdated.

2. Product reviews, setup instructions, and other real-world blog posts. LLMs want to summarize these, and I don't want that.

3. Really specific knowledge in a limited domain ("2017 Kia Sedona automatic sliding door motor replacement steps," "Can I exit a Queue-Triggered Azure Function without removing it from the queue?"). In these cases, the LLMs are so prone to hallucination that I can't trust them.

The answer I'm seeking is not always on reddit itself, but google limited to reddit is far more likely to give me quality starting links than google unbound is.

The AI word salad summaries for each individual page have no toggle (unless you count !g).

Searching 'octatrack' has the wikipedia page on the top right as a summary box thing. no ai word salad for me :shrug:

I've mostly switched to using Claude these days, with MCPs for websearch and fetching specific remote or local files. It answers questions generally very accurately (from the source documents it identifies) and includes citations.

I've found that people that haven't really tried the latest models, and just rely on whatever knowledge is in the model training are really missing out on the potential power. GPT4o+ and equivalent models really changed the game. And using tools to do a search, or pull in your code, or run a db query or whatever enables them to either synthesize information or generate context relevant material. Not perfect for everything, but much better than a year ago, or what people are doing with the free systems.

If it can't manage one small fact on something that was covered quite a bit in the previous month, then it is worse than useless. At the very least it should say I don't know. It reminds me of that one guy we all know that does nothing but make stuff up when talking about stuff outside of their wheel house. Never backs down, never learns anything, and ultimately dumped from the relationship.

Even without much customization (lenses, scoring, etc) it's so much better (for my use cases) I happily pay for it.

Recently I have also started to use Perplexity more for "research for a few minutes and get back to me" type of things.

Queries like "what was that Python package for X" I usually ask an AI right from my editor, or ChatGPT if I'm in the browser already.

2 recent success stories:

I was toying around with an esp32 - i was experimenting to turn it into a bluetooth remote control device. The online guides help to an extent, setting up and running sample projects, but the segue into deploying my own code was less clear. LLMs are "expert beginners" so this was a perfect request for it. I was able to jump from demos to live deploying my own code very quickly.

Another time I was tinkering with opnsense and setting up VLANs. The router config is easy enough but what I didnt realize before diving in was that the switch and access point require configuration too. What's difficult about searching this kind of problem is that most of the info is buried in old blog posts and forum threads and requires a lot of digging and piecing together disparate details. I wasn't lucky enough to find someone who did a writeup with my exact setup, but since LLMs are trained on all these old message boards, this was again a perfect prompt playing to its strengths.

The results from LLMs are still too slow, vary too much in quality and still frequently hallucinate.

My typical use-case is that when I'm looking for an answer I make a search query, sometimes a few. Then scan through the list of results and open tabs for the most promising of them - often recognising trusted, or at least familiar, sites. I then scan through those tabs for the best results. It turns out I can scan rapidly - that whole process only takes a few seconds, maybe a minute for the more complex queries.

I've found LLMs are good when you have open-ended questions, when you're not really sure what you're looking for. They can help narrow the search space.

At most I use AI now to speed up my research phase dramatically. AI is also pretty good at showing what is in the ballpark for more popular tools.

However I am missing forum style communities more and more, sometimes I don't want the correct answer, I want to know what someone that has been in the trenches for 10 years has to say, for my day job I can just make a phone call but for hobbies, side projects etc I don't have the contacts built up and I don't always have local interest groups that I can tap for knowledge.

LLMs can't be trusted, you have no way to tell between a correct answer and a hallucination. Which means I often end up searching what the LLM told me just to check, and it is often wrong.

Search engines can also lead you to false information, but you have a lot more context. For example, a StackOverflow answer has comments, and often, they point out important nuances and inaccuracies. You can also cross-reference different websites, and gauge how reliable the information is (ex: primary source vs Reddit post). A well trained LLM can do that implicitly, but you have no idea how it did for your particular case.

What are the specs for new Goolge Pixel 9a? LLM can't answer this may after a year they can.

Not anymore

I've always found it curious how the level of technology in Children of Men seemed to have stalled at around the time of the outbreak. From where I see it, the maturation of LLMs is the zombie outbreak for human communication.

Last night, I asked Claude 3.7 Sonnet to obtain historical gold prices in AUD and the ASX200 TR index values and plot the ratio of them, it got all of the tickers wrong - I had to google (it then got a bunch of other stuff wrong in the code).

Also yesterday, I was preparing a brief summary of forecasting metrics/measures for a stakeholder and it incorrectly described the properties of SMAPE (easily validated by checking Wikipedia).

I constantly have issues with my direct reports writing code using LLM's. They constantly hallucinate things for some of the SDK's we use.

Was a bit more useful at questions like "Rank these stocks by exposure to the Chinese market", as you can prioritise your own research but in the end you just have to go through the individual company filings yourself.

Google seems to be better at bringing up a variety of stackoverflow and blogposts relevant to my search queries. Qwant seems to struggle exactly with that: it's great at giving exactly what I was searching for but that's sometimes not what I was looking for, if you get what I mean.

In a sense, LLM's are actually perfect for that. But like you say, the super confident hallucinations are just too frustrating. Literally every time I've asked it a serious programming question, it's hallucinated an API that doesn't exist. Everyone seems to be focusing on letting LLM's solve math and thinking problems. That's exactly what _I'm_ good at. I would much rather have an LLM that is good at combining sources and giving me facts (knowledge, rather than thinking) while most of all being able to say "I don't know".

But now, the veracity of most LLMs' responses is terrible. They often include “sources” unrelated to what they say and make up hallucinations when I search for what I'm an expert in. Even Gemini in Google Search told me yesterday that Ada Lovelace invented the first programming language in the 18th century. The trust is completely gone.

So, I'm back to the plain old search. At least it doesn't obscure its sources, and I can get a sense of the veracity of what I find.

I mean, for everyone else it was never there to begin with. Hallucinations are constantly raised as the biggest issues with AI. According to the tests, and my experience, newer AI models are objectively better, not worse than the ones from a few years ago. They still have a long way to go/may never be fully trustworthy though.

What I have lost trust in, and what I and many feel has become much worse over the few years, is Google search, and all the other search engines that are based on it.

Isn't that what ChatGPT/anthropic web access does?

I recently upgraded my video card, and I run a 4K display. Suddenly the display was randomly disconnecting until I restarted the monitor. I googled my brains out trying to figure out the issue, and got nowhere.

So I gave ChatGPT a shot. I told it exactly what I upgraded from/to, and which monitor I have, and it said "Oh, your HDMI 2.0 cable is specced to work, but AMD cards love HDMI2.1 especially ones that are grounded, so go get one of those even if it's overspecced for your setup."

So I did what it said, and it worked.

I just searched for "What is inherit_errexit?" at Perplexity. Eight sources were provided and none of them were the most authoritative source, which is this page in the Bash manual:

https://www.gnu.org/software/bash/manual/html_node/The-Shopt...

Whereas, when I searched for "inherit_errexit" using Google Search, the above page was the sixth result. And when I searched for "inherit_errexit" using DuckDuckGo, the above page was the third result.

I continue to believe that LLMs are favored by people who don't care about developing an understanding of subjects based on the most authoritative source material. These are people who don't read science journals, they don't read technical specifications, they don't read man pages, and they don't read a program's source code before installing the program. These are people who prioritize convenience above all else.

This makes a lot of sense to me. As a young guy in the 90's I was told that some day "everyone will be fluent in computers" and 25 years later it's just not true. 95% of my peers never developed their fluency, and my kids even less so. The same will hold try for AI, it will be what smartphones were to PCs: A dumbed down interface for people who want to USE tech not understand it.