4 ms·

I like Chiang’s stories, but he simply doesn’t know what he’s talking about when it comes to LLMs. Forget the philosophical points - he claims that LLMs are str

by emtel 2y ago

I like Chiang’s stories, but he simply doesn’t know what he’s talking about when it comes to LLMs. Forget the philosophical points - he claims that LLMs are strictly worse than search engines for information retrieval. This is just obviously false, I can give you dozens of conversations from my ChatGPT history over the past two months which would have been impossible to get answers to on Google.

benreesman 2y ago
LLMs are generally better search tools for certain topics today, but search engines have been in serious decline for at least 15 years. IMHO old-school Google remains the high water mark of generalized information retrieval, with advantages ranging from speed to semi-durable citation. I strongly suspect there is a cohort thing going on here, many HN users today weren’t involved in technology yet back when Google worked well. Much like beer for Homer Simpson, AI is the cause of and solution to all of the Internet’s problems.
- emtel 2y ago
  I’ve been using Google since about 2000/2001 or so. In fact I worked there for a time, and I still remember Craig Silverstein’s answer to “isn’t search a solved problem?” Which was: “in the future we want Google to be able to take any question and just give you the answer. So clearly we have a long way to go!” In any case, I do not believe there was ever a time it could answer all of the questions that LLMs can today. If the question had been asked and answered on the web, Google could (and can) find it, but many questions haven’t been asked!
  - kortilla 2y ago
    Search isn’t the same thing as answering a question
- nuancebydefault 2y ago
  Google search works at least as well as 15 years ago, the results are now just more interleaved with ads that are harder to distinguish from what you are really searching for. That's exactly where LLMs come in, the model inside the weights has more than answers, they can find sense in data.
bmitc 2y ago
What are the examples? I've had several LLM search result summaries contain flat out mistakes and incorrect statements.
- emtel 2y ago
  I’ll try to dig some up soon (I’m on my phone now). But of course the output contains errors sometimes. So do search engine results. The important thing for difficult questions is whether the right answer (or something pointing toward it) is available _at all_. Of course this assumes you can verify the answers somehow (usually easy with programming questions), but again, search engines have the same limitation.
  - bmitc 2y ago
    There isn't an expectation or claim that search engines answer anything. They just find things or don't find things.
  - vhantz 2y ago
    > But of course the output contains errors sometimes. So do search engine results. That's not true. Search engine results are links and (non-AI generated) summaries of existing resources on the web. No search engine returns links to resources it generated as the result of the search query. Those resources can have innacurate information, yes, but the search engine itself does not returns errors. LLMs output do not contain errors "sometimes". The output of an LLMs is never truthful nor false in itself. In the same way that the next word your keyboard suggests for you to type on a mobile device is never truthful nor false. It's simply the next suggestion based on the context. These two methods of accessing information very clearly do not have the same limitations. A search engine provide link to specific resources. A LLM generates some approximation of some average of some information. It's up to intelligent thinking people to decide whether a LLM or a search engine is currently the best way for them to parse through information in search for truth.
    - emtel 2y ago
      Obviously I meant that the content of the results can be inaccurate, and I assume you weren't actually confused about that.
OptionOfT 2y ago
I think those are 2 different issues. Searching for something, and finding it, is different from what ChatGPT / Claude does. Google (in the good old days) is like the library. You want to search 'how to plant potatoes on Mars'. No results. Well, you split it up, maybe a book on planting potatoes, and a book about missions to Mars that describes soil composition. Then, when you have those books you start reading, parsing, understanding, making connections, identifying what needs to be done etc. Maybe, if you're lucky, you find a book or a web page where somebody went through the thought exercise of finding out what would be needed to make it work. ChatGPT / Claude / ... are different in that they have the information in their corpus, and that the information they present you could actually be something that has never been written down in a book, or published on the web. That's why Google can't find it, but ChatGPT is able to present you with a satisfying answer. Now whether the answer is correct is a different issue. Do you have the knowledge to verify this?
- inglor_cz 2y ago
  OK, so I asked ChatGPT how to plant potatoes on Mars... ================================================= Planting potatoes on Mars would be a pretty fascinating (and challenging) task! While Mars has conditions that make growing traditional crops difficult—like low temperatures, low atmospheric pressure, and a lack of oxygen—scientists have been experimenting with ways to grow plants in Martian-like conditions. Here’s an overview of the process: 1. Create a Controlled Environment: Temperature: Mars’ average temperature is about -80°F (-60°C), far too cold for potatoes to grow. You’d need to create a greenhouse-like environment, potentially using materials like glass or transparent plastics, to keep the temperature warmer. Atmosphere: Mars has only about 1% of Earth's atmospheric pressure, which is too low to support plant life. A sealed greenhouse would be required to provide a breathable atmosphere with adequate pressure and oxygen levels. Light: Mars has less sunlight due to its distance from the Sun. You would need supplemental artificial light (perhaps LEDs) to mimic Earth’s daylight cycle for the plants. 2. Soil: Mars has soil, but it’s not exactly like Earth’s. It lacks organic material and has toxic elements like perchlorates that would harm plant growth. Scientists would need to either: Modify Martian soil by adding organic material (like compost) and removing or neutralizing toxic chemicals. Hydroponics: Grow potatoes without soil, using a nutrient-rich water solution instead.
BeetleB 2y ago
The comparison with search is faulty to begin with. Yes, you can search with an LLM, but that's a side effect of the tool. While I certainly also have found things via LLMs that I couldn't easily with a search engine, the number of false positives is huge. My heuristic is: If I ask an LLM something and it's easy to verify via Google because its answer narrows the search space - then I'll use it. Otherwise, Google is still king. Example: Asking an LLM the health benefits of supplement X is a waste of time. Verifying everything it tells me would be the same amount of work as asking a search engine. Example: Asking how to solve a given coding problem is great, because it drastically reduces the search space. I only have to look up the particular function/API calls it uses. Ditto for asking how to achieve a task in the command line - I can quickly verify the arguments are accurate via the man page. Most of the things I search for do not fall into this category, but in the category of "still need to do the same amount of work as just searching via Google."

by emtel 2y ago

benreesman 2y ago

LLMs are generally better search tools for certain topics today, but search engines have been in serious decline for at least 15 years. IMHO old-school Google remains the high water mark of generalized information retrieval, with advantages ranging from speed to semi-durable citation. I strongly suspect there is a cohort thing going on here, many HN users today weren’t involved in technology yet back when Google worked well. Much like beer for Homer Simpson, AI is the cause of and solution to all of the Internet’s problems.

emtel 2y ago
I’ve been using Google since about 2000/2001 or so. In fact I worked there for a time, and I still remember Craig Silverstein’s answer to “isn’t search a solved problem?” Which was: “in the future we want Google to be able to take any question and just give you the answer. So clearly we have a long way to go!” In any case, I do not believe there was ever a time it could answer all of the questions that LLMs can today. If the question had been asked and answered on the web, Google could (and can) find it, but many questions haven’t been asked!
- kortilla 2y ago
  Search isn’t the same thing as answering a question
nuancebydefault 2y ago
Google search works at least as well as 15 years ago, the results are now just more interleaved with ads that are harder to distinguish from what you are really searching for. That's exactly where LLMs come in, the model inside the weights has more than answers, they can find sense in data.

bmitc 2y ago

What are the examples? I've had several LLM search result summaries contain flat out mistakes and incorrect statements.

emtel 2y ago
I’ll try to dig some up soon (I’m on my phone now). But of course the output contains errors sometimes. So do search engine results. The important thing for difficult questions is whether the right answer (or something pointing toward it) is available _at all_. Of course this assumes you can verify the answers somehow (usually easy with programming questions), but again, search engines have the same limitation.
- bmitc 2y ago
  There isn't an expectation or claim that search engines answer anything. They just find things or don't find things.
- vhantz 2y ago
  > But of course the output contains errors sometimes. So do search engine results. That's not true. Search engine results are links and (non-AI generated) summaries of existing resources on the web. No search engine returns links to resources it generated as the result of the search query. Those resources can have innacurate information, yes, but the search engine itself does not returns errors. LLMs output do not contain errors "sometimes". The output of an LLMs is never truthful nor false in itself. In the same way that the next word your keyboard suggests for you to type on a mobile device is never truthful nor false. It's simply the next suggestion based on the context. These two methods of accessing information very clearly do not have the same limitations. A search engine provide link to specific resources. A LLM generates some approximation of some average of some information. It's up to intelligent thinking people to decide whether a LLM or a search engine is currently the best way for them to parse through information in search for truth.
  - emtel 2y ago
    Obviously I meant that the content of the results can be inaccurate, and I assume you weren't actually confused about that.

OptionOfT 2y ago

I think those are 2 different issues. Searching for something, and finding it, is different from what ChatGPT / Claude does. Google (in the good old days) is like the library. You want to search 'how to plant potatoes on Mars'. No results. Well, you split it up, maybe a book on planting potatoes, and a book about missions to Mars that describes soil composition. Then, when you have those books you start reading, parsing, understanding, making connections, identifying what needs to be done etc. Maybe, if you're lucky, you find a book or a web page where somebody went through the thought exercise of finding out what would be needed to make it work. ChatGPT / Claude / ... are different in that they have the information in their corpus, and that the information they present you could actually be something that has never been written down in a book, or published on the web. That's why Google can't find it, but ChatGPT is able to present you with a satisfying answer. Now whether the answer is correct is a different issue. Do you have the knowledge to verify this?

inglor_cz 2y ago
OK, so I asked ChatGPT how to plant potatoes on Mars... ================================================= Planting potatoes on Mars would be a pretty fascinating (and challenging) task! While Mars has conditions that make growing traditional crops difficult—like low temperatures, low atmospheric pressure, and a lack of oxygen—scientists have been experimenting with ways to grow plants in Martian-like conditions. Here’s an overview of the process: 1. Create a Controlled Environment: Temperature: Mars’ average temperature is about -80°F (-60°C), far too cold for potatoes to grow. You’d need to create a greenhouse-like environment, potentially using materials like glass or transparent plastics, to keep the temperature warmer. Atmosphere: Mars has only about 1% of Earth's atmospheric pressure, which is too low to support plant life. A sealed greenhouse would be required to provide a breathable atmosphere with adequate pressure and oxygen levels. Light: Mars has less sunlight due to its distance from the Sun. You would need supplemental artificial light (perhaps LEDs) to mimic Earth’s daylight cycle for the plants. 2. Soil: Mars has soil, but it’s not exactly like Earth’s. It lacks organic material and has toxic elements like perchlorates that would harm plant growth. Scientists would need to either: Modify Martian soil by adding organic material (like compost) and removing or neutralizing toxic chemicals. Hydroponics: Grow potatoes without soil, using a nutrient-rich water solution instead.

BeetleB 2y ago

The comparison with search is faulty to begin with. Yes, you can search with an LLM, but that's a side effect of the tool. While I certainly also have found things via LLMs that I couldn't easily with a search engine, the number of false positives is huge. My heuristic is: If I ask an LLM something and it's easy to verify via Google because its answer narrows the search space - then I'll use it. Otherwise, Google is still king. Example: Asking an LLM the health benefits of supplement X is a waste of time. Verifying everything it tells me would be the same amount of work as asking a search engine. Example: Asking how to solve a given coding problem is great, because it drastically reduces the search space. I only have to look up the particular function/API calls it uses. Ditto for asking how to achieve a task in the command line - I can quickly verify the arguments are accurate via the man page. Most of the things I search for do not fall into this category, but in the category of "still need to do the same amount of work as just searching via Google."