3 ms·

The Wolfram plugin also has extremely impressive examples [1]. If I were OpenAI, I would use the usage data to further train the model. They can probably use C

The Wolfram plugin also has extremely impressive examples [1]. If I were OpenAI, I would use the usage data to further train the model. They can probably use ChatGPT itself to determine when an answer it produced pleased the user. Then they can use that to train the next model. The internet is growing a brain. 1: https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers/ https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its...

bsaul 4y ago
In a way this is a bit less impressive than having chatgpt being able to generate those answers on its own only using learned information. This kinds of admits defeat regarding its ability to solve mathematical problems correctly or answer factual information.
- unit_circle 4y ago
  But doesn't this seem like the correct approach long term? Basically shelling out to what essentially amounts to a fact table when it gets a factual question
  - sp332 4y ago
    Yes, there's no point spending all that time and compute training and inferring ways to do math problems when you could just use a calculator.
  - arlcode 4y ago
    It is but it also shows current limitation of the reasoning skills of this models. A super human AGI would easily be able to use tools like Wolfram Alpha and also derive most of it on the fly or from memory just by thinking about it. If you set your expectations to AGI (right now) you will be disappointed but that doesn't mean it's not immensely useful. Unfounded hype and real technological progress seem to go hand in hand.
    - flakeoil 4y ago
      Should Einstein not be allowed to use a calculator and instead waste his time on performing the calculations by hand?
      - arlcode 4y ago
        Obviously yes. And how they use these tools is important. But it's quite interesting to see what these models can do internally and what they can't do yet. It possibly outlines future research areas (beyond " more scale") and opportunities for competitors to enter the market (which is usually better for consumers and society as a whole). Maybe that's the difference between productization and academic reasearch.

bsaul 4y ago

In a way this is a bit less impressive than having chatgpt being able to generate those answers on its own only using learned information. This kinds of admits defeat regarding its ability to solve mathematical problems correctly or answer factual information.

unit_circle 4y ago
But doesn't this seem like the correct approach long term? Basically shelling out to what essentially amounts to a fact table when it gets a factual question
- sp332 4y ago
  Yes, there's no point spending all that time and compute training and inferring ways to do math problems when you could just use a calculator.
- arlcode 4y ago
  It is but it also shows current limitation of the reasoning skills of this models. A super human AGI would easily be able to use tools like Wolfram Alpha and also derive most of it on the fly or from memory just by thinking about it. If you set your expectations to AGI (right now) you will be disappointed but that doesn't mean it's not immensely useful. Unfounded hype and real technological progress seem to go hand in hand.
  - flakeoil 4y ago
    Should Einstein not be allowed to use a calculator and instead waste his time on performing the calculations by hand?
    - arlcode 4y ago
      Obviously yes. And how they use these tools is important. But it's quite interesting to see what these models can do internally and what they can't do yet. It possibly outlines future research areas (beyond " more scale") and opportunities for competitors to enter the market (which is usually better for consumers and society as a whole). Maybe that's the difference between productization and academic reasearch.