People who want to get rich will tell you it's the next greatest thing that will revolutionize the industry.

Personally, I've been annoyed at how confidently wrong ChatGPT can be. Even when you point out the error and ask it to correct the mistake it comes back with an even-more-wrong answer. And it frames it like the answer is completely, 100% correct and accurate. Because it's essentially really deep auto-complete, it's designed to generate text that sounds plausible. This isn't useful in a search context when you want to find sources and truth.

I think there are useful applications for this technology but I think we should leave that to the people who understand LLM's best and keep the charlatans out of it. LLM's are really interesting and have come a long way by leaps and bounds... but I don't see how replacing entire institutions and processes by something that is only well understood by a handful of people is a great idea. It's like watering plants with gatorade.

Reminds me of the web3 crypto hype and this hype is the same thing that happened with GPT-3. Closed source black-box AI model hidden behind a SaaS confidently generating wrong answers as the truth.

Sounds like OpenAI (effectively Microsoft's new AI division) are selling snake-oil again.

> I think there are useful applications for this technology but I think we should leave that to the people who understand LLM's best and keep the charlatans out of it.

There are certainly charlatans, grifters and snake-oil sales people pretending and hyping about their so-called 'AI startup' when it is actually uses the GPT-3 or ChatGPT API. Another emperor has no clothes confidence trick generating garbage.

Given that ChatGPT cannot explain its own answers transparently when questioned especially when it confidently generates wrong answers means that you cannot trust its output and telling it to give you a sophisticated answer to an 'un-googleable' question is where you see it clearly trip over.

What would really 'revolutionize the industry' is an open-source LLM model which is a smaller model and more transparent than ChatGPT, like what Stable Diffusion did to DALLE-2.

The Bing openai integration is already dramatically different and better than chatgpt. I’d really advise trying it before forming too many opinions. Certainly don’t mistake any experience you have with chatgpt as indicative.

Fundamentally, GPT itself and LLMs are essentially still black-box neural networks and all have the same drawbacks as them:

* More data needed to train them. Which then...

* ...Increases the size of the model (Getting bigger on each revision)

* Opaqueness (Cannot transparently explain its own decisions when it confidently generates the wrong answer.)

* Massive cost involved in training, retraining and fine-tuning it.

The best part is that you can't even give an explanation on why it is generating nonsense or why it overfitted or got confused over an invalid character or input. Little to nothing fundamental about neural networks (which is what LLMs are based on) has changed. It is essentially more of: 'Just waste more money on training it on more data.'

> The Bing openai integration is already dramatically different and better than chatgpt. I’d really advise trying it before forming too many opinions.

What I have said about ChatGPT and GPT-3 still holds true and hyping about a chatbot confidently generating nonsense and that cannot cite its own sources makes it highly untrustworthy to use. It is always eternal to the shortcomings of neural networks fundamentally; which is what LLMs use. They are great sophists and bullshit generators.

At least what needs to exist is an open-source LLM that is more transparent and is smaller than GPT-3 or GPT-4. There is a basic reason why OpenAI won't do that.

And even in LLMs, it sounds like you're still seeing training/retraining as single monolithic events which are super expensive. Training is super expensive! New techniques (google LoRA; there are others) are changing that.

I hope you get into what has changed in the past few years, starting with Transformer and continuing pretty much daily. There really is a lot of innovation and improvement that is publicly documented and missing from your opinion here.

It is a GPT (Using OpenAI's one) and fundamentally it is a black box neural network that when asked to explain it's own decisions, it still cannot transparently do so. This have been an unsolved problem for decades and still overlooked by every AI hype cycle to date.

> And even in LLMs, it sounds like you're still seeing training/retraining as single monolithic events which are super expensive. Training is super expensive! New techniques (google LoRA; there are others) are changing that.

Perhaps that is why only Big Tech companies are the ones doing it for training LLMs and the so-called 'AI companies' aren't doing it themselves and still sit on their AI APIs hence why you could only name Google who can afford to foot the bill for training their LLM.

Everyone else has to sit on their APIs just like before.

> I hope you get into what has changed in the past few years...

I think you have struggled to address the other short-comings that both LLMs and fundamentally neural networks still face that I have already outlined.

Until I see a open-source LLM or some other open source solution that matches or surpasses OpenAI's GPT offerings in parameters, explainability, and is small enough in model size to fit on a smartphone then we can discuss about how this AI cycle is worth the hype. Having yet another AI API product with even larger and expensive models sitting behind a SaaS is an indication of another AI bubble.

Very few pieces of software today explain their decisions, and it doesn't seem to be a big problem for users.

Long-range Radio??

I do wish people would check if acronyms don't already have well-established meanings before adopting them.

Also OP (not so much you) is way too caught up in the “chat” aspect - that is the first exciting UX that got popular, but these are much, much more than chatbots. Pretending that they’re human/conscious/in a conversation is fun, but having an invisible partner that knows you and tailors your search results… that’s powerful.

For example, you’ll never have to add “Reddit” again, or at least you’ll only have to tell it once. An LLM can easily identify the kind of questions where you want forum posts, read thousands of posts in a second, summarize all their content, and label each link with other information that helps you decide which threads to read in full.

I can’t wait!

For one, you can't just "hook up" a language model to some other task, or to the web. ChatGPT is specifically built and trained to have good conversations. The fact that it can sort of appear to do other things is a happy coincidence.

To do any of what you want, new algorithms need to be built, and none of that is "easy". And finally, these models take A LOT of cpu time. They are not going to be reading thousands of posts in a second without serious and expensive compute hardware backing it, and that level of compute isn't remotely feasible to give out to individual users.

Even chatGPT, which is doing a fraction of the tasks you are listing, costs millions of dollars worth of hardware a day. The only reason it exists for free is because Microsoft has donated all that time.

They had to pay someone to label data and filter the worst "content" humanity has to offer. Otherwise it would've ended up like numerous other attempts at exposing "AI" to the Internet.

So it also has a huge human cost that OpenAI is not properly accounting for (and another reason why dreaming up potential use-cases for the technology as if it will be miniaturized and become a commodity in the near future takes some wilful ignorance).

But chatgpt is already pretty darn good at summarizing and communicating. By “hook up” I literally just mean feed text from the internet into the prompt followed by whatever you want it to do with it - summarize, rank, whatever. Ignoring the economics for a moment (paid search engines?) and assuming that GPT 3.5 is the very best LLM we’ll ever get for simplicity: would you still say you don’t think tweaked versions of such models would MASSIVELY improve search?

There is also the business model to sort out. Right now search is primarily driven by ads, which I doubt will cover the costs of the sort of ultra-personalization that you're thinking about. Also, reducing the time you spend on a search engine or looking through results will further reduce ad revenue. However, I can see paid search engines perhaps leading to this.

So yes, eventually these models can help improve search, just not in the form that we have today. In a few years the story could well be different. I'm quite interested in seeing how Bing integrates chatgpt technology. They claim that they've created some new model on top of it that somehow links it to search results.

When you ask whether speaking English and using logic can improve the search experience, I wonder what you consider the most important parts of the search experience? I think many people, most of the time, might say that "accurate information" is their highest expectation, with "a pleasant conversation" somewhere below that. Delivering a plausible-sounding and pleasant answer that's completely wrong is... well, that's not a search engine I can depend on, is it?

You're hypothesizing a few things at the end that sound great! It's completely unclear whether any of those things will actually end up happening, so I think the focus on what is available today, with Chat-GPT and Bing, is more apt than a focus on what could be.

And I totally agree that a) a search engine that doesn’t cite its sources is useless, and b) you almost never want to chat with google like it’s a person. So you’re spot on. But the point I was trying to make is that the main use case is in stuff like automated summaries, specialized page rankings, expanding quick informal queries into longer formal-logic ones, etc.

The point is, "accurate information" is hard. Google's solution is snippets and while it might be fine for some cases, it fails terribly for others. There is zero guarantee an AI-based solution would be more precise, but for sure it will be way more confident - just like ChatGPT is.

I think the complaints are more about the "use logic" point than the sources, from my limited understanding, I would not say LLMs currently use logic.

The start of its lengthy response:

“ The use of probabilistic logic models in LLM can lead to more sophisticated and nuanced logical reasoning. These models represent knowledge about the world in a probabilistic form, which allows the LLM to take into account uncertainty and make inferences based on probabilities.

For example, an LLM system trained on a large knowledge base might encounter a question that it has not seen before. Using probabilistic reasoning, the LLM…”

There is a general issue where AIs fail in different ways than humans, and the failures look really dumb to a human. So humans tend to ascribe that dumbness from the human scale. Instead, I’d suggest they just have a dramatically different spider-graph of capabilities to a human, and are overall more capable than the “dumb spreadsheet / parrot” narrative admits. (Definitely not human-level IQ yet, to be clear.)

I question if this something that users need at the scale you're assuming. Wikipedia has existed for 20 years summarizing enormous breadth of human knowledge, with some articles having thousands of human editors. it's a boon for civilization in the way libraries are. But has it disrupted anything besides Microsoft's overpriced Encarta DVD market?

You're putting a lot of faith in computer models to provide accurate, both-sides'ed information on complex topics in a format that amounts to a soundbite.

Can you give me 3 example queries (questions and answers or typical searches) that are clear cut wins for a search engine application?

The nice way of saying grifters.

This is the most accurate description I've even read.

Those autocorrections are still frequently wrong.

That also happens with real people so... A web search also returns wrong answers, because it is not magic. It just searches through all the garbage out there. You just have to be aware of its flaws and limitations... as you do with you fellow humans.

It's not a person. It's software. It shouldn't be wrong just like humans.

Imagine opening your banking app and your total savings is wrong. Would you say, "It's okay, people make mistakes, too!" or would you be pissed that your bank's software is incorrect? Why are we pretending like this is any different?

AI machines and especially neural networks which LLMs are based on still cannot explain themselves and are as transparent as a black-box.

Search is absolutely like asking a random person right now. Whether I get raw hyperlinks back on a page or a chat window, the results are as good as their source data. Garbage in, garbage out.

Google is just as confidently wrong giving me garbage links as chat is as giving me garbage recommendations.

With ChatGPT, correct and incorrect information is presented identically. You have nothing to go on except falling back to web search to fact-check ChatGPT’s output. This can still be useful as a starting point, but a world with only ChatGPT and no web search would be horrible, would overall be a step back compared to a world with only web search and no ChatGPT.

I have seen no indication that they had anybody verify the factual content of answers in the training process, and many indications that they did not.

Ideally Google search would have a flag to "follow my intent to the letter" and return empty if nothing is found. When you are searching for a specific thing, a response with other things feels like bullshit, Google trying to milk more clicks wasting your time. I don't mean exact phrase search, I mean exact semantic search.

This is causing issues when searching for bug fixes by ignoring the exact version, when shopping it will ignore some of your filters and respond with bad product suggestions, and when searching something specific that looks like some other popular keyword, it will give you the popular one, as if if has an exclusion zone and you cannot search for anything else around it.

"Minimum weight of a scooter with back suspension" -> matches information about carrying capacity. Of course more people discuss about max passenger weight than minimum scooter weight, but I really don't care about the other one.

That makes the response more current than the last LLM crawl&tune, allows a further directed search based on the new parameters, and provides a traceable path to sources and citations. If it finds that there's not a Python library with that name then iterate.

Basically, an LLM should have a very good idea what good search terms are for the topic, and where to find the information, whereas I might not know the acronyms, jargon, or related fields and optimal answers.

Yes, this is getting pretty close to writing an 8th grade class report that covers anything on the web. That's about where these seem to be.

*ps I copied this from a post I made yesterday*

Among other things, a LLM can be seen as a store which you query and get results from. A chatbot is cute because it formats output text to look like conversation, and the recent applications are nice because the query (now known as prompt) can be complicated and long, and can influence the format and length of the results.

But the cool stuff is being able to link the relatively small amount of text you input as a query, into many other chunks of texts that are semantically similar (waves hands around like helicopter blades). So, an LLM is a sort of "knowledge" store, that can be used for expanding queries, and search results, to make it more likely that a good result seems similar to the input query.

What do I mean by similar? well, the first iteration of this idea is vector similarity (e.g. https://github.com/facebookresearch/DPR). The second iteration is to store the results into the model itself, so that the search operation is performed by the model itself.

This second iteration will lead, IMHO, to a different sort of search engine. Not one over "all the pages" as, in theory at least, google and the like currently work. Instead, it will be restricted to the "well learnt pages", those which, because of volume of repetition, structure of text, or just availability to the training algorithm, get picked up and encoded into the weights.

To make an analogy, is like asking a human who are the Knights of the Round Table and getting back the usual "Percival, Lanceelot and Galahad", but just because the other thousand knights mentioned in some works are not popular enough for that given human to know them.

This is a different sort of search engine than we are used to, one which might be more useful for many (most?) applications. The biases and dangers of it are things we are only starting to imagine.

First, the human input is extremely flexible, but can include instructions. It is natural language programming.

Second, the "conversation" has state. I can give an instruction, and then a followup instruction that adds to the first instruction. Someday down the road there will be two states, your account state (instructions you taught it that it retains as long as you are logged in. Maybe my account can have multiple state buckets/buildings I can enter, one of one set of rules, one for another. Could call them programs or routines. (computer execute study routine)) and temporary state (instructions it retains only for the duration of the conversation/search.)

The exciting part here is being able to query data and manipulate it in memory. Making a search, refining the search, redirecting the search in a different direction when its not working. That collaborative, iterative type search doesnt really exist at the moment. I cant tell google "the results you just returned are garbage, here is why, try again."

It is more like a fuzzy commandline. The chatbotness is just a layer of cute on top, that isnt completely necessary.

The enterprise product will start with its own state. It will be something you can limit to specific tasks. Youll be able to write an app that takes instructions and outputs data in magical ways completely different than anything before. Normal data manipulation programming will be replaced by this weird opaque black box, that is ever changing, and you will need to trust consistently outputs the same thing given an identical input.

However, it's been observed, people are using chatbots for informational searches. The kinds of searches where you want to learn about a specific fact. This isn't all searches, but it's an important subset for web search. For better or worse, with probably a high degree of inaccuracy, people (probably rightly) perceive this is how people will seek information.

There's also the generational use cases - "write me a program that does X". Is this something people would use a search bar for? We don't know, and wouldn't know, until its out there for a while.

For the longest time the one natural language interface was as a search bar. So search vendors surmise it's important to both defend their turf while also a natural way to get regular users familiar with this kind of informational interaction...

Instead it was all good in 20 minutes.

Personally, I think Microsoft has a HUGE opportunity here, imagine if the OpenAI - PRO account was actually free -- but only on edge browsers, or inside bing.com. It'd be a bit of a loss leader for a bit, because a lot of compute is required but it'd also help unseat google. Google's only chance is if lamda/bard is better - and so much better it's noticable from day one.

Their search results are so horrible lately I don't know anybody who isn't 'shopping around' right now for something better. Bing, DDG, Brave, Neeva, You.com, etc.

If that were true I’d see people downloading edge and only using it for LLMs. MS hasn’t figured out how to monetize it yet, their only goal is to catch up and possibly dethrone google. Not an easy feat from MS who have blundered and still blunder a lot. Google search on the other hand is becoming less and less useful day by day.

Oh God. I seriously, seriously hope not. I can imagine some coasters I work with shitting out source code doing this

Anyone who trusts an LLM and is committing to a code base that other people work on - I would never want to work with someone with this attitude. Physically writing code takes far less time than understanding code

1) Integration with voice assistants. Links/sources are irrelevant.

2) Models tuned against a particular body of work don't care if links go stale, or websites get SlashdottedHackerNewshuggedDDOSed, or etc. Links/sources are irrelevant.

3) Inbound "service requests" processed by something that can better understand the question and the available answers/solutions. Links don't matter much.

4) When "Okay what are some good websites to read more about this?" can be answered, too, bang.

5) Ever asked somebody a question and just rolled with their answer instead of demanding citations? I mean, you're doing it here. So, again, yes.

ChatGPT will already do this, when you get an answer you want to learn more about, ask

"where can I learn more about this?"

"can you provide links?"

there are a number of prompts that work, you can even provide it with an AND clause in the first prompt

It seems there are two kinds of people: those who see wrong results from ChatGPT and think they are wrong and those who see wrong results from ChatGPT and are amazed at how right they are.

If I was going to use AI for something it might be to find the type II's and make sure they "never work in this town again" because if they can't call ChatGPT on its bullshit they probably can't call anybody on their bullshit.

> I've seen ChatGPT make up fake links and fake citations that look superficially real but don't really point to anything.

I tried asking it the other day for free software that would transcribe video. It gave me five pieces of software, four of which didn't exist and one of which was not free.

As a software dev I am expected to get it right, not almost right. I just coded up something that works most of the time but failed in one case and the tester sent it back. In my case that defect takes somewhere between 1% and 500% of the effort to fix as the rest of that ticket but in the case of ChatGPT "close but no cigar" is not close at all in terms of having a strategy that is reliably truthful.

If ChatGPT has a really "amazing" ability it is in getting people to think it is amazing. I think of that story

https://en.wikipedia.org/wiki/The_Emperor%27s_New_Clothes

and that it's not just cheaply dismissive to point out that ChatGPT's core competence is the same as that Emperor because not everybody can pull off what that Emperor pulled off. See also

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

which is a anecdote from the beginning of AI research that seems forgotten today in how people react to ChatGPT.

To think that ChatGPT is not impressive is to close yourself off from what is actually happening. I was there with you until I started trying it out regularly. To think that it is not providing value is to neglect reality. You do not get to judge others as inferior for finding ChatGPT useful. This is the impression you are giving.

Nothing is perfect. Think of something that amazes you and then consider where it is not perfect. It is perfectly reasonable for people to find ChatGPT impressive without it being perfect. DAN and SAM are also impressive sessions people have had with ChatGPT that show somewhere it can be manipulated.

TL;DR As devs, we use search because it saves us hours of time. ChatGPT is a tool you should add to your toolbox because it can also save you hours of time.

That is not 100% true. I'm sure all of us have put code in production that we knew had issues and it was a conscious decision given the tradeoffs and timelines.

This being separate from the bugs we unknowingly release into the wild.

- how do current search engines rank results?

- what features correlate with good?

- how do features of sites get encoded, queried, and returned?

- how might LMs be already involved in this process?

- how might LLMs / ChatGPT like systems encode these things?

- how might we enrich training data to facilitate improved LLM information retrieval?

I would suspect, with current models and algos, that the more often certain statements are found across the training data, the more likely that information is to be "correct"

A good site might be one that has many of these "correct" chunks. Though I doubt they are there yet. A next phase task will likely be along the lines we are discussing here. i.e. How do we return citations and references in responses

I think there's a poor correlation between search engine rankings and website quality, so it might be better not to include this.

> what features correlate with good?

Right, that's the essential question. Google's original bit of genius was to to ignore that question entirely, and go with "popular" instead. You can measure popular, but "good" entirely depends on what it is you value. That makes it somewhat subjective.

I imagine that you could have a quality measurement by finding the chunk embeddings that are most popular, and then finding the sites that have these in the highest concentration. User interaction will still likely be a very important feature in the quality space.

Yes. It was measuring popularity. It was genius because it solved the problem usefully well, but page rankings have never been a fantastic indication of quality -- again, depending on what you consider "quality". It was just better than nothing.

Back link counting is not the same thing as popularity, which measures visits

PageRank was better than keyword search, which became spammed and PR was resilient to. Now there are 100s of signals being used, see the Yandex leak

https://snap.stanford.edu/class/cs224w-readings/Brin98Anatom...

There's a lot more to the algorithm as well, it's a good read

I expect these LLMs to improve in the coming years, especially as they are made into products and wrapped with more tools into Q&A pipelines.

Another point is that I am either creative or productive at a time, but never both... at least aware of which state I am in. ChatGPT has proven to take over the other part surprisingly good. ie:

- when I am in a productive mood and stumble upon a thinking problem, generative AI is like on-the-spot creativity for "good enough" solutions, like naming a programming thingy or write some filler text around a few keywords instead of me looking for words.

- when I am in a creative mindset, I increasingly feed some code snippets into the bot and ask some questions to "fill in the gaps", like writing a specific function using library X, then to write a documentation explaining how it works, then to also write some unittests, and sometimes I even derail a bit or let the bot explain parts that stand out in some way so I can maybe learn a trick.

... And i used ChatGPT already in kinda emergency situations, like when I know 5 minutes in advance that I have to speak in front of a crowd/in a meeting and it gave me extremely useful outlines to quickly adapt to even when in a panicked mind state - calming me down through a given structure that sounded okay-ish, and it doesn't matter if the response is right or wrong.

Generating bullshit text off the cuff is not the only use of LLMs. LLMs can perform very well at classification, regression, ranking, coloring proper names red, and other tasks. You could, for instance, use LLMs to encode a query and documents and rank them with a siamese network, something not too different for how a conventional search engine works.

If there is one thing wrong with the current crop of LLMs it is that these can only attend over a limited number of tokens. BERT can attend over 512 tokens, ChatGPT over 4096, where a token is shorter than a word on average. It easy to process the headline of an HN submission with BERT, but I classify a few hundred abstracts of scientific papers a day. A long abstract is about 1800 words which is too much for longformer but would fit in ChatGPT if there aren't too many $10 words.

Unless you can recast a problem as "does this document have a short window in it with this attribute?" (maybe "did the patient die?" in a clinical case report or "who won the game?" in a sports article) there is no way to cut a document up into pieces and feed it into an LLM, then combine the output vectors in a way that doesn't break the "magic" behavior the LLM was trained to do.

You'd imagine ChatGPT would produce accurate results if you could tell it "Write an review of topic T with citations" but if you try that you'll find it will write citations that look for real but don't actually exist if you look them up. You'd imagine at minimum that such a system would have to read the papers that it cites, maybe being able to attend over all of them at the same time which would take an attention window 100-1000x larger.

That's by no means cheap and it might be Kryptonite for Google in that Google's model involves indexing a huge amount of low quality content and financing it by ads that are a penny a click. A business or individual might get a larger amount of value per query out of a much smaller task-oriented document set.

When you're hunting for a particular fact, like "that bit of code i half remember seeing on a page 15 years ago", then I don't see anything for an LLM to add. Google had a pretty good index for that purpose about 15 years ago, but they've chosen to prioritize other goals since then. I dunno if anyone works "find things you're searching for" as a market now.

Which is an answer to your question: Does it matter if an LLM helps search the web? That's not what people are doing, that's not what these companies are selling.

LLMs have a chance to offer an oracle that doesn’t answer in cryptic or evasive ways, but attempts to just give an answer. The hallucinations are a huge flaw, but one that I’m confident will be addressed with other non-LLM AI approaches. But it’s the right user interface for answering questions - it answers questions with answers, not a pile of potentially relevant documents to sort. The augmentation with citations, especially if they’re semantically revenant rather than symbolically, is a huge plus.

https://you.com/search?q=youchat&tbm=youchat

I don't know if that strictly complies with your definition of "web search application". It's definitely going to save time for me, and not seeing a bunch of ads during the process is wonderful - to the point that I really could see myself paying for it if they decide to go that route and take away the "free" version.

When I think "search engine", I think of something that searches for web pages. What I want as a result isn't an answer, meal plan, grocery list, or any of that higher-order stuff. I want a list of websites.

This reinforces my fear that this sort of approach might kill actual search engines. That would be a pretty serious loss to me, and a further reduction in the usefulness of the web. But saying that is not disparaging things like ChatGPT at all. They address a different, and valid, need.

I just hope we don't lose an important tool in the process.

The basic idea was to have enough metadata about web sites so that you could get programs to do something approaching Prolog-style reasoning about the content and meaning of the web pages.

With more advanced LLMs it looks like a slightly different approach to achive something like the semantic web idea.

I think the idea is to constantly feed the model with updates from crawling the web and have the LLM "digest" the content, apply some filters to remove bad stuff, and then provide a meaningful result to whatever queries it might be asked.

The intellectual north star of this is

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

and might be a better model than the "semantic web" in that the semantic web has revolved around description logics like OWL that implement a limited subset of first-order logic that is always decidable. One trouble is that OWL-based systems can't handle arithmetic at all, given a height and a weight you can write a rule that says anybody over 6 feet is tall but you can't say a person with a BMI over 30 is obese because OWL can't add, multiply and divide. But something a little more expressive that searches efficiently in practice because it gets "hunches" from a neural net could be useful but it will not evade the limits on computation and logic that Godel, Tarski and Turing warned you about.

there’s something about SATPLAN that feels like it has metaphysical significance.

That idea was the basis for the current generation of Google search (see: Metaweb).

The issue I see with the chat approach is trust. I've seen so many examples of these models just making shit up now that I reckon regular use of them will eventually lead to mistrust between the human and the chat-bot. If you can't trust the answers and have to go and check yourself, it's dead as an idea IMHO.

> Search systems, like many other applications of machine learning, have become increasingly complex and opaque. The notions of relevance, usefulness, and trustworthiness with respect to information were already overloaded and often difficult to articulate, study, or implement. Newly surfaced proposals that aim to use large language models to generate relevant information for a user’s needs pose even greater threat to transparency, provenance, and user interactions in a search system. In this perspective paper we revisit the problem of search in the larger context of information seeking and argue that removing or reducing interactions in an effort to retrieve presumably more relevant information can be detrimental to many fundamental aspects of search, including information verification, information literacy, and serendipity. In addition to providing suggestions for counteracting some of the potential problems posed by such models, we present a vision for search systems that are intelligent and effective, while also providing greater transparency and accountability.

Shah & Bender (2022) "Situating Search". In Proc. CHIIR '22 https://dl.acm.org/doi/abs/10.1145/3498366.3505816

- The Metaverse was going to change the internet.

- The Internet of things was going to revolutionise our homes and cities.

- Self-driving cars would change the motor industry.

- Crypto is the next big thing in finance.

- NFTs are going to revolutionise digital ownership.

Most of those technologies have introduced niche or useful applications, but the ridiculous, breathless hype about how this technology is the one to change everything is getting more predictable and increasingly frustrating.

> every single "AI" search I have tried has given me incorrect, invented answers.

Really? Every single one? Not a single AI (ChatGPT, presumably) query you've tried has given you anything except incorrect information? What kind of things have you been querying for?

- chatGPT - 10% hallucination

- Google - 50% spam

Which is better - the "honest hallucination" from AI or the "spam for profit" human operation?

For quick overview answer, LLMs are great. It's not 100% correct, but mostly it is, and that is good enough for a quick answer. Currently google tries to show that and people object as it is stealing traffic from websites. i just need an answer, a coherent useable one. Eg: "What were the movies Scorsese got an Oscar nomination for?"

For suggestions, LLMs are just one more of those blogs and listicles that are already showing up in search. If LLM is updated that is. The difference is an LLM would customize the answer according to the query, unlike already pre written content. So, yes useful. Same goes with stuff like: "how to build an email list?" or "What is a effective sales strategy?"

For research, Google is more useful. I think we have all done that.

Another application which is not realized at this time because we never did it before is the ability to ask follow up questions (which a chat format enables well). Suppose you get an overview of how a quantum computer works, but it would take a lot of effort to ask a follow up question and get a direct answer via a search engine. Eg: "Why is there no point in going beyond thousand Qubits?"

There could be modifications like voice to text (a jarvis like interface), or a personal assistant thingy. But those are far fetched.

It will help immensely, and for places it does not, we will still google like we have done before.

While what they’re doing currently isn’t perfect, it does provide results that are at least traceable. I could imagine an alternate universe where they doubled down on marketing themselves as “the search engine that doesn’t lie to you” or “where answers are found, not stories”.

On the other hand, I do agree with people speculating that LLM-AI interfaces will seriously hurt Google's bottom line, e.g. reducing the space for search ads, which represent the majority of its revenue.

[0] https://www.nytimes.com/2022/09/16/technology/gen-z-tiktok-s...

Mostly this is just to calm me down because ChatGPT gives me the illusion that I'm interacting with a human. The current voice systems are infuriatingly bad.

It would be nice if CVS's phone system would actually listen to me and modify its output accordingly. "I already gave you my birth date. And NO, I don't need a COVID booster."

edit: I'd like to meet the person who sold CVS its prescription web-site and its voice system. Simply to marvel at them and the swindle they pulled off, delivering absolute trash and probably walking away with a king's ransom.

But at least they're deterministic and finite. I imagine ChatGPT like results from a phone system would be even worse. LLM's are incredible when you can massage the answer you want out of it in a feedback loop, but not so much for automated systems.

You might be correct.

I can't be sure.

My opinion is definitely influenced by spending an hour on the phone with CVS yesterday + another hour on the phone with my father's cable company.

I could imagine the interface being similar to what we have today, but with it being much better at taking in full descriptions of what you want. If you want pictures of teddy bears, it could provide search results and AI generated ones. If you want programming answers, it could link to StackOverflow or just give you an AI generated answer with an explanation. Perhaps I am looking for a lively bit of free music to add to an indie game - it could generate that too.

I feel that this will eventually end search as we know it, but it will hurt the sites that are behind the search results far more than it will hurt Google. Google (or Bing, ...) can become the one-stop-shop for so much more than it is today

Imagine asking someone with cursory knowledge of a subject matter to perform a google search for you. This person would dig through thousands of results and weed-out the junk/SEO/content farm sites, so you'd get information that's more relevant. LLMs could potentially do this quickly, separating the wheat from the chaff. Would it be perfect? No, but it would be a significant improvement over what you see on Google today.

When doing research I need a good search engine. Find me the official docs, not the SEO’d blogs. Find me that podcast episode. Find me the exact article I remember reading 10 years ago. Don’t try to guess and half-arse a result just because it has more ads or better SEO. I’ll do the hard work of synthesis because I’m looking to understand something deeply.

Current search engines have gotten meh at this use-case. Or at least Google has.

When looking for a quick answer, I need a smart-enough agent. How old is this celebrity? What’s the air speed of an unladen swallow? Give me a deity that starts with G. What the hell is a “nepo baby”? What does this random emoji mean when sent by a 20 year old and what is it by a 40 year old? Who’s that actor in that show with the thing?

I don’t care about the source and I’m not looking to do research. Just tell me a good enough answer so I can get back to my conversation or whatever. Current search engines are pretty okay with this, but GPT is better.

The two use-cases are fundamentally different and trying to merge them is where things went wrong.

(Voice chat)

- LLM, find three articles from HN frontpage which I would find insightful based on my recent evaluations, summarize them in under half a minute each and then I’ll choose the order, while I commute.

- (…)

- Okay, read me a second one first.

- (…)

- That was a good one because well-written and compared alternatives. Now find funniest article of all time based on my long-time preferences.

- (…)

- - -

While it’s dumb enough to forget what and how you’ve evaluated recently, a hidden prompt could(?) fetch that out, e.g.:

- (system) please convert my previous article ratings into json objects consisting of article url, article id (…how to get it…), 1..10 rating, your summary and a string of tags.

Then these ratings may be saved for later and fed into a chat secretly as:

- (system) If I gave you this prompt: “<voice prompt>” and had data looking like this “<db schema>”, please list which tables and filters would you likely use to form an answer.

- (…)

- My recent ratings were (…). <voice prompt>

- - -

I have no clue if this could work, but if it does, well, that would be useful.

Edit: it may be wrong, but we have enough mundane tasks which are better done wrong rather than not done at all. It has a great potential as an “occasionally bright secretary” archetype.

As an example, someone on HN posted a tweet of a guy who asked ChatGPT to draft a letter announcing a layoff while also announcing several executive promotions and quoting MLK Jr. Obviously, the example is facetious but the results were actually pretty good. Certainly good enough for a starting point or template for a real layoff announcement.

I'm sure this is a miniscule amount of total search volume but there are a category of searches for letter templates (think cover letters, resignation letter, etc) that ChatGPT could seriously replace today. And ChatGPT is actually better because of how specific you can get (e.g. "with an MLK quote").

I don't think LLMs are a threat to traditional search today or even in the short term but what will ChatGPT 50 (or equivalent) look like in 20 years...

That being said, I don't think ChatGPT or any single LLM can replace mainstream Internet search use cases in the immediate future. They might enhance the search experience for users

It would be useful if a search engine could find the top 10 different interpretations of the phrase in latent space, and offer the top results (with a means to pursue more) in each of those separate meanings.

For example: "hypertext markup" matches way too many things about HTML, and not enough things about marking up (annotating) hypertext

LLMs could make search much more powerful in this manner.

ChatGPT, on the other hand... is not a search engine, even before you consider its tendency towards BS.

The workflow I have now with ChatGPT and what I image it will look like in the new era of search is: query -> read result written for a human, not a search engine -> (10% of the time) check if the result is hallucinated.

Especially for basic questions where I know what the answer should look like, I'm really enjoying the new workflow.

That use case is the one that makes more sense. Given that ChatGPT frequently hallucinates the wrong answer and confidently tells you how it is correct with its inability to cite and transparently explain to you how it got to that answer tells you that it is results are untrustworthy and this AI bubble is again, pure hype created by VCs.

The only worthy AI hype that will change everything is open-source LLMs that are smaller models and are more transparent.

"Can you provide a reference for that? or What should I google to confirm that this is accurate?"

It'll give you something pretty close to a final citation. This has saved me literal days of work traversing documentation in the last 2 months.

I think it works amazingly well at least for instances when you can immediately verify whether the answer is correct (e.g. coding, drafting letters) and instances where it is a starting point for further research. These use cases are a significant portion of my searches, so I think it will be very useful.

I saw word embeddings in the mid 2010's as a step backwards and I walked away from a lucrative consulting contract because I couldn't convince the sponsor that it was a total waste to apply word2vec to clinical documents because a significant amount of the meaning of the documents was in out-of-distribution words. (Expecting that to succeed is like a football coach thinking he can shoot the quarterback in the head and still win the game.)

For me the train was pulling out of the station with FastText, BERT and similar models that used subword tokens because w/ the last generation of systems you really could show they threw out critical information in preprocessing that would make accuracy a matter of luck at best.

But i think that "suggest me something" is going to be a big selling point too. Decision-making is tiring, and people are willing to give that power to a machine. Look at how tiktok, youtube and even facebook now are working, it's slowly becoming like a TV stream that you passively watch. "Tell me what to do tomorrow" is going to be a common question in a year or so

There's more to generative models than just the above, but how much of this hype cycle is substantive for end users or developers? You always had a query. Now you'll get answers in a different way.

Skeptical overall.

Doesn't feel as simple as the search world where you can just up rank some approved sources, unless there is somehow a way to just generate the language from approved sources and ideas.

It responded immediately with some working code.

With Google, I could have found the result but it would have taken a dozen clicks and probably 15 minutes of my time. (To be fair, I might have learned more in the process).

Siri or Google Assistant wouldn't have given an answer at all.

People want to believe this is as amazing as it appears, but it’s window dressing we still can’t intentionally separate relevant information from requests.

Good luck wasting hours of your time googling for that by hand.

Newegg would be wise to adopt such filtering.

1. with the possible exception of really niche stores like digi-key or other electronics components stores where it's really important you get exactly what you're looking for

That’s not using the search service for search

...until the rise of the new ecclesiastical class? These models are biased in several ways with the training set defining their world view and those who train them intervening in specific areas to bend the output to their will. They can be made to negate the old dictum of garbage in, garbage out to garbage in, gospel out - where the gospel follows whatever the (small-c) creator thinks the populace should know and (in extension) think.

In short I don't trust these models any further than I can throw them.

It isn't hard to imagine a world where answers from ML models are the standard of truthiness, much like Wikipedia is for many.

To the downvoters: explain, don't just knee-jerk your disapproval. I don't care about karma but it is tiring to see things which don't follow the narrative being downvoted out of existence.