4 ms·

Feel like you could interchange that particular title and the rest of it will work for 99% of HN stories.

hn_throwaway_99 3y ago
I really feel like that's underselling it. Yes, many stories are responded to with "a mix of curiosity, concern, and reminiscence". But it still got a lot of the specific details right: 1. First, it identified the top story correctly, "Google has a secret browser hidden inside the settings". 2. "Others express concerns about the potential security risks and the implications for parental or school controls" and "Some users reminisce about similar workarounds in the past or share their own experiences with bypassing restrictions" are really specific to that story and not generally applicable. I know this may come across as being pedantic, but as much as AI can be over-hyped, I also see a lot of cases where it is underhyped (e.g. "it's just doing autocomplete!") that also aren't really a fair assessment of its capabilities.
- dns_snek 3y ago
  > 1. First, it identified the top story correctly, "Google has a secret browser hidden inside the settings". How is parroting the top title back to the user supposed to be impressive?
  - hn_throwaway_99 3y ago
    That it speaks English? Being human, I think a lot of us tend to forget what an incredibly difficult problem natural language processing is. I'm always a little baffled when people discount the abilities of LLMs, when meanwhile I'm blown away that I'm having a perfectly valid, seemingly-human-level conversation with a machine. Yes, all of the problems with things like inaccuracy, hallucinations and bias need to be called out and highlighted, but if you had showed me a ChatGPT transcript a year ago I would have thought you were bluffing, that it was really a "human behind the curtain". Point being, the person didn't ask for the system to output "console.log((retrieveHNFrontPage()).stories[0].title". The person simply asked "What is the general sentiment of the comments on the current top hacker news story?" And you might argue that there is some parsing logic you could implement to pull out "current top hacker news story" and translate that to code, but the fact is that currently nothing is better than an LLM at taking practically any natural language request and turning it into an actionable piece of code.
    - grugagag 3y ago
      It is very impressive indeed and something that was not possible before. Im blown away by it, played a bit with it and still do from time to time and yet I move on with my life not caring too much, I wish I did. At work I got signed up for copilot trial and am dreading it a bit as I don’t really have time or a real need for it currently. On the other hand when done with work I want to spend as much time away from modern technology as I can. Returning to the story, again Im blown away by this project and what’s possible with LLMs, will give it a go a few times but will probably continue to take notes manually and not rely too much on tools like these except for when I explocitly need it.
- Lockal 3y ago
  1. Solved by feeding text version of front page into response - https://github.com/steven-tey/chathn/blob/main/app/api/chat/functions.ts#L69 https://github.com/steven-tey/chathn/blob/main/app/api/chat/... . It can not even read linked articles. What you see can be replaced with 4 buttons: get_top_stories + get_story + get_story_with_comments + summarize_top_story. 2. "secret browser" automatically triggers generally applicable conspiracy & common journalistic statements. If you try to ask more specific questions, it immediately fallbacks to: "As an AI language model, I don't have access to real-time data or the ability to browse external websites like Hacker News".
  - kuhewa 3y ago
    Still amazes me how readily people are to assume GPT is doing something noteworthy without a second thought to simpler ways it could be faking an intelligent and specific analysis
- lvncelot 3y ago
  While I totally agree with you, I've read the comment you're replying to as more of a (lighthearted) jab at HN comments than an actual evaluation of the results.
- boredumb 3y ago
  It is cool, sorry, I was more picking on the HN comments being predictable than the AI itself. If anything it's actually a bit of an ode to the AI itself because it managed to (hilariously) hit the nail on the head of hacker news discussions with a single article where as it took me a few hundred articles to see that same pattern :D

I really feel like that's underselling it. Yes, many stories are responded to with "a mix of curiosity, concern, and reminiscence". But it still got a lot of the specific details right: 1. First, it identified the top story correctly, "Google has a secret browser hidden inside the settings". 2. "Others express concerns about the potential security risks and the implications for parental or school controls" and "Some users reminisce about similar workarounds in the past or share their own experiences with bypassing restrictions" are really specific to that story and not generally applicable. I know this may come across as being pedantic, but as much as AI can be over-hyped, I also see a lot of cases where it is underhyped (e.g. "it's just doing autocomplete!") that also aren't really a fair assessment of its capabilities.

dns_snek 3y ago
> 1. First, it identified the top story correctly, "Google has a secret browser hidden inside the settings". How is parroting the top title back to the user supposed to be impressive?
- hn_throwaway_99 3y ago
  That it speaks English? Being human, I think a lot of us tend to forget what an incredibly difficult problem natural language processing is. I'm always a little baffled when people discount the abilities of LLMs, when meanwhile I'm blown away that I'm having a perfectly valid, seemingly-human-level conversation with a machine. Yes, all of the problems with things like inaccuracy, hallucinations and bias need to be called out and highlighted, but if you had showed me a ChatGPT transcript a year ago I would have thought you were bluffing, that it was really a "human behind the curtain". Point being, the person didn't ask for the system to output "console.log((retrieveHNFrontPage()).stories[0].title". The person simply asked "What is the general sentiment of the comments on the current top hacker news story?" And you might argue that there is some parsing logic you could implement to pull out "current top hacker news story" and translate that to code, but the fact is that currently nothing is better than an LLM at taking practically any natural language request and turning it into an actionable piece of code.
  - grugagag 3y ago
    It is very impressive indeed and something that was not possible before. Im blown away by it, played a bit with it and still do from time to time and yet I move on with my life not caring too much, I wish I did. At work I got signed up for copilot trial and am dreading it a bit as I don’t really have time or a real need for it currently. On the other hand when done with work I want to spend as much time away from modern technology as I can. Returning to the story, again Im blown away by this project and what’s possible with LLMs, will give it a go a few times but will probably continue to take notes manually and not rely too much on tools like these except for when I explocitly need it.
Lockal 3y ago
1. Solved by feeding text version of front page into response - https://github.com/steven-tey/chathn/blob/main/app/api/chat/functions.ts#L69 https://github.com/steven-tey/chathn/blob/main/app/api/chat/... . It can not even read linked articles. What you see can be replaced with 4 buttons: get_top_stories + get_story + get_story_with_comments + summarize_top_story. 2. "secret browser" automatically triggers generally applicable conspiracy & common journalistic statements. If you try to ask more specific questions, it immediately fallbacks to: "As an AI language model, I don't have access to real-time data or the ability to browse external websites like Hacker News".
- kuhewa 3y ago
  Still amazes me how readily people are to assume GPT is doing something noteworthy without a second thought to simpler ways it could be faking an intelligent and specific analysis
lvncelot 3y ago
While I totally agree with you, I've read the comment you're replying to as more of a (lighthearted) jab at HN comments than an actual evaluation of the results.
boredumb 3y ago
It is cool, sorry, I was more picking on the HN comments being predictable than the AI itself. If anything it's actually a bit of an ode to the AI itself because it managed to (hilariously) hit the nail on the head of hacker news discussions with a single article where as it took me a few hundred articles to see that same pattern :D