3 ms·

> company is shaping up to be to the future of society It's definitely a terrifyingly powerful company, but "future of society" seems a little strong. I think

by smt88 4y ago

> company is shaping up to be to the future of society It's definitely a terrifyingly powerful company, but "future of society" seems a little strong. I think it's the future for spam, writing rote emails/documents/code, disinformation, plagiarism, lying, cheating, and impersonation, but those are subsets of society -- not all of society.

whateveracct 4y ago
I've already had managers try to weigh in on technical discussions by posting ChatGPT's "thoughts" as if it had a seat at the table. They also quoted it to answer a question of when we need to worry about scale (a big part of the discussion we were having). The answer was useless btw. Just a coarse, high-level, unactionable summary of what we had already talked about. Dressed up in a few nicely-worded paragraphs. I wonder if this is gonna be a new pain in my ass. Wait until managers start using it to disagree with my estimates - I might blow my top lol.
brokencode 4y ago
What a bleak outlook you have. Here are some good things it can do: 1. Answer student questions and effectively act as a tutor. Duolingo is already using it like this, and ChatGPT can of course do it directly as well. 2. Make natural language interfaces to APIs easily. Look at the ChatGPT plugins announced yesterday for examples. 3. Provide basic customer support. Ideally, it could answer most common and basic questions and possibly even fix common problems via a plugin. Then the actual human customer support could step in for more complex problems. There are endless positive examples like these.
- pasabagi 4y ago
  Regarding 1., it appears to have about a 30% accuracy rate, and the other 60% is complete nonsense, often complete with fabricated citations. I dearly hope that nobody is ever encouraged to have this machine as their tutor.
  - motoxpro 4y ago
    You're right. They won't make any progress on this metric at all.
    - pasabagi 4y ago
      They might do, but what error rate do you think is acceptable? How do you actually measure and test the error rate? It's a use I can imagine in the future, but I think it's really premature to be using 'personal tutor' as a benefit (as openai do in their advertizing materials) when the program, as it stands, is essentially a fluent and convincing bullshitter, which is the single worst possible trait for a teacher.
      - motoxpro 4y ago
        Error is acceptable in all things that don't need to be deterministic. Which is most thing in life. How do I discover a good career path? Why do you structure a repository of a program into folders? What is the best place to vacation? What is the best way to learn math? What is a good way to articulate socialism? How do I increase my vocabulary? What is corporate strategy? Ask 10 different people "smart" people (define smart however you want), you'll get 10 different answers to these questions. These are all questions an LLM could answer amazingly. Probably a lot better than most humans. If you don't ask it what 2+2 is or who came to in America in 1875 then you get useful things. Asking a LLM deterministic questions right now is like asking a calculator what the meaning of life is. If you use the tool for something it's not good at you get unusable answers.
        pasabagi 4y ago
        If you ask an idiot what he thinks about something, and he gives you a totally wrong answer, you have still learned at least one fact: a person believes a thing. As a human, living in a democracy, that has some worth. ChatGPT's wrong answer has absolutely no value at all. Further, 10 different smart people will give 10 different answers because they have coherent worldviews and biases and proclivities, so by accounting for those, you can work out what the right answer is. Even if ChatGPT was anywhere close to a human expert when it comes to accuracy (what's the error rate in a peer reviewed journal article?) it would still have no coherent worldview or bias to contextualize its statements.
- JohnFen 4y ago
  Those seem underwhelming positives in the face of the obvious negatives.

by smt88 4y ago

whateveracct 4y ago

I've already had managers try to weigh in on technical discussions by posting ChatGPT's "thoughts" as if it had a seat at the table. They also quoted it to answer a question of when we need to worry about scale (a big part of the discussion we were having). The answer was useless btw. Just a coarse, high-level, unactionable summary of what we had already talked about. Dressed up in a few nicely-worded paragraphs. I wonder if this is gonna be a new pain in my ass. Wait until managers start using it to disagree with my estimates - I might blow my top lol.

brokencode 4y ago

What a bleak outlook you have. Here are some good things it can do: 1. Answer student questions and effectively act as a tutor. Duolingo is already using it like this, and ChatGPT can of course do it directly as well. 2. Make natural language interfaces to APIs easily. Look at the ChatGPT plugins announced yesterday for examples. 3. Provide basic customer support. Ideally, it could answer most common and basic questions and possibly even fix common problems via a plugin. Then the actual human customer support could step in for more complex problems. There are endless positive examples like these.

pasabagi 4y ago
Regarding 1., it appears to have about a 30% accuracy rate, and the other 60% is complete nonsense, often complete with fabricated citations. I dearly hope that nobody is ever encouraged to have this machine as their tutor.
- motoxpro 4y ago
  You're right. They won't make any progress on this metric at all.
  - pasabagi 4y ago
    They might do, but what error rate do you think is acceptable? How do you actually measure and test the error rate? It's a use I can imagine in the future, but I think it's really premature to be using 'personal tutor' as a benefit (as openai do in their advertizing materials) when the program, as it stands, is essentially a fluent and convincing bullshitter, which is the single worst possible trait for a teacher.
    - motoxpro 4y ago
      Error is acceptable in all things that don't need to be deterministic. Which is most thing in life. How do I discover a good career path? Why do you structure a repository of a program into folders? What is the best place to vacation? What is the best way to learn math? What is a good way to articulate socialism? How do I increase my vocabulary? What is corporate strategy? Ask 10 different people "smart" people (define smart however you want), you'll get 10 different answers to these questions. These are all questions an LLM could answer amazingly. Probably a lot better than most humans. If you don't ask it what 2+2 is or who came to in America in 1875 then you get useful things. Asking a LLM deterministic questions right now is like asking a calculator what the meaning of life is. If you use the tool for something it's not good at you get unusable answers.
      - pasabagi 4y ago
        If you ask an idiot what he thinks about something, and he gives you a totally wrong answer, you have still learned at least one fact: a person believes a thing. As a human, living in a democracy, that has some worth. ChatGPT's wrong answer has absolutely no value at all. Further, 10 different smart people will give 10 different answers because they have coherent worldviews and biases and proclivities, so by accounting for those, you can work out what the right answer is. Even if ChatGPT was anywhere close to a human expert when it comes to accuracy (what's the error rate in a peer reviewed journal article?) it would still have no coherent worldview or bias to contextualize its statements.
JohnFen 4y ago
Those seem underwhelming positives in the face of the obvious negatives.