3 ms·

I simply don't understand this take. Yes, I think AI is incredibly over-hyped at the moment, and I've come to totally agree with Yann LeCun's assessment that LL

I simply don't understand this take. Yes, I think AI is incredibly over-hyped at the moment, and I've come to totally agree with Yann LeCun's assessment that LLMs have nothing to do with AGI, but I still think it's an incredibly useful tool if one is aware of its limitations. Things like: 1. ChatGPT is great at translations (at least for English-French), and I've found it too be a great language learning assistant. 2. It's incredibly useful for me when it comes to writing a first draft of complicated SQL queries. 3. It's great at essentially a faster document search and summarization tool if I don't know the keyword of what I'm looking for, e.g. when I want to ask "I want to code X, which libraries can help me do that?" 4. I use cursor.sh as my IDE and I find it incredibly useful, especially for finding answers about a large codebase when I'm a newbie to the repository. 5. I've found ChatGPT to be great for trip ideas and planning. In short, I think generative AI is a powerful if flawed tool, and I think the disillusionment comes when people think it's more than that.

cjk2 2y ago
I think the disillusion here is when you find out that these things are mostly smoke and mirrors that tend to generate very suboptimal solutions. For example I am forever having to shoot down ChatGPT generated code before it hits production which is naively and lazily generated and the engineer is not aware of this at all. Yes it can write SQL but does it understand the index strategy on the destination table or the consequences of running thousands of those queries a minute in a production database? No. Does it have a deep understanding of context rather than just syntax (ask a French language translator what they think of it). No. And with trip planning and ideas, which is a speciality of mine, it is always better to find someone local and ask them about it because a huge portion of the information out there isn't something a model can infer about because it isn't written down anywhere. Case in point, I'm off to the Azores islands shortly - it can't generate anything useful for me. I had someone suggest that already. On top of that it is compelled to over explain sometimes to the point of incorrectness, which it likes to do on travel advice. This becomes dangerous. I found someone local to talk to instead. So personally I think the LLM and question based uses are limited and quite ridiculous from an objectivity perspective. They barely scratch the surface of the problem and the hit rate is less than I'd get doing the legwork myself, which I do not find useful. And the information is hard to verify and judge if it's safe without deeper contextual knowledge. This is a complete mire. The other use cases, like classification and search are where I see it winning but those aren't sexy and don't play on our human emotions as well enough to part with cash. They will be left after the hype of course. Really I think people are buying into a romantic fantasy at the moment.
- aurareturn 2y ago
  >Yes it can write SQL but does it understand the index strategy on the destination table or the consequences of running thousands of those queries a minute in a production database? No. It does if you instruct it well enough. We're still really early and companies are figuring out how to best use it. It's been 1 year since GPT4 was released. We didn't invent the lightbulb, electric motor, telephone, radio, refrigerators etc. until many years after the invention of electricity. Right now, LLMs are very useful for many tasks. Over time, they'll continue to get more useful and inventions will be built on top of them.
  - cjk2 2y ago
    None of those things were commoditised in even a 20 year span. Many more things died. Effectively everyone is running on faith. That’s not good enough.
    - aurareturn 2y ago
      But electricity didn't die. Many products built on top of electricity died. I don't think AI is running on faith in 2024. I think it's running on very realistic projections and what is available now. I have the exact opposite view of you. I actually think the AI hype is not overblown. I think it's "underblown". I think we have a very straight forward path in the next few years for AI to be useful everywhere.
  - lomase 2y ago
    It does if you instruct it well enough with a maybe 100k in GPU time and a team of data scientist who have expertise in the domain. The other option you have is to hire somebody who can write a SELECT with a JOIN in SQL.
sumedh 2y ago
> I use cursor.sh as my IDE How does it compare to Co Pilot or the AI Assistant from JetBrains IDEs
- hn_throwaway_99 2y ago
  I can't answer that anymore. When I first used it it was much better than Copilot, especially for "ask my repo this question"-type queries, but I think Copilot may have added that functionality since I last used it.

cjk2 2y ago

I think the disillusion here is when you find out that these things are mostly smoke and mirrors that tend to generate very suboptimal solutions. For example I am forever having to shoot down ChatGPT generated code before it hits production which is naively and lazily generated and the engineer is not aware of this at all. Yes it can write SQL but does it understand the index strategy on the destination table or the consequences of running thousands of those queries a minute in a production database? No. Does it have a deep understanding of context rather than just syntax (ask a French language translator what they think of it). No. And with trip planning and ideas, which is a speciality of mine, it is always better to find someone local and ask them about it because a huge portion of the information out there isn't something a model can infer about because it isn't written down anywhere. Case in point, I'm off to the Azores islands shortly - it can't generate anything useful for me. I had someone suggest that already. On top of that it is compelled to over explain sometimes to the point of incorrectness, which it likes to do on travel advice. This becomes dangerous. I found someone local to talk to instead. So personally I think the LLM and question based uses are limited and quite ridiculous from an objectivity perspective. They barely scratch the surface of the problem and the hit rate is less than I'd get doing the legwork myself, which I do not find useful. And the information is hard to verify and judge if it's safe without deeper contextual knowledge. This is a complete mire. The other use cases, like classification and search are where I see it winning but those aren't sexy and don't play on our human emotions as well enough to part with cash. They will be left after the hype of course. Really I think people are buying into a romantic fantasy at the moment.

aurareturn 2y ago
>Yes it can write SQL but does it understand the index strategy on the destination table or the consequences of running thousands of those queries a minute in a production database? No. It does if you instruct it well enough. We're still really early and companies are figuring out how to best use it. It's been 1 year since GPT4 was released. We didn't invent the lightbulb, electric motor, telephone, radio, refrigerators etc. until many years after the invention of electricity. Right now, LLMs are very useful for many tasks. Over time, they'll continue to get more useful and inventions will be built on top of them.
- cjk2 2y ago
  None of those things were commoditised in even a 20 year span. Many more things died. Effectively everyone is running on faith. That’s not good enough.
  - aurareturn 2y ago
    But electricity didn't die. Many products built on top of electricity died. I don't think AI is running on faith in 2024. I think it's running on very realistic projections and what is available now. I have the exact opposite view of you. I actually think the AI hype is not overblown. I think it's "underblown". I think we have a very straight forward path in the next few years for AI to be useful everywhere.
- lomase 2y ago
  It does if you instruct it well enough with a maybe 100k in GPU time and a team of data scientist who have expertise in the domain. The other option you have is to hire somebody who can write a SELECT with a JOIN in SQL.

sumedh 2y ago

> I use cursor.sh as my IDE How does it compare to Co Pilot or the AI Assistant from JetBrains IDEs

hn_throwaway_99 2y ago
I can't answer that anymore. When I first used it it was much better than Copilot, especially for "ask my repo this question"-type queries, but I think Copilot may have added that functionality since I last used it.