3 ms·

I don’t think it’s good. ChatGPT is clearly better but that too is not that great. We are still a couple more iterations away. I also wonder if Google is at dis

by lvl102 4y ago

I don’t think it’s good. ChatGPT is clearly better but that too is not that great. We are still a couple more iterations away. I also wonder if Google is at disadvantage because they’re not utilizing Nvidia hardwares? Or not to the extent OpenAI does?

haolez 4y ago
The difference is probably related to RLHF. It seems to matter a lot more than the size of the model and the hardware's capabilities. And OpenAI probably excels at that.
- cubefox 4y ago
  Bing uses GPT-4 with their own fine-tuning, and they had to put some artificial restrictions in place in order for the model not "misbehave". Apparently they don't use RLHF but some combination of SL and prompting. OpenAI didn't have such problems.
  - saberience 4y ago
    What are SL and RLHF? I would also add it's best not assume others know acronyms like these.
    - cubefox 4y ago
      Sorry. SL = supervised learning. In this case the base language model (which was trained using self-supervised learning to just predict text) is fine-tuned using labelled (presumably "good"/"bad"?) dialogue examples. RL = reinforcement learning. RLHF = reinforcement learning from human feedback. Here human raters reward or punish the model for good/bad dialogue.
      - saberience 4y ago
        Thank you! I'm going through the steps now of learning the Maths needed for Data Science and this is useful to me :)
godshatter 4y ago
In my opinion, iterating ChatGPT and the like isn't going to get us anywhere, at least in terms of general AI. It's an interesting trick that might have some uses, but iterating it isn't going to lead to more understanding. It's effectively an extensively well-read and quick-thinking bullshitter. It's knows a lot of words and sentences, it can put them together in ways that sound creative or intelligent, but it doesn't understand anything at all about what it's actually saying. Right now, I see it as a dead end. I'd rather more effort was going into parsing language into the underlying concepts, making connections between them, and looking for new understanding in how those concepts might fit together in new ways or finding new ways of using that underlying understanding to solve problems. Basically what people, do, but better.
- quonn 4y ago
  We don‘t even have a path toward what you‘re suggesting. None at all. So ChatGPT is far closer to a solution.
- selfhoster11 4y ago
  ChatGPT (especially GPT-4) is the leanest, meanest Prolog engine known to mankind. It absolutely slays on inference tasks if you state some data and rules upfront. It knows how to translate between structured formats and natural language. That's not nothing. You could do this earlier, maybe, if you had a lot of specialised AI knowledge. Now, someone like me, with little AI training but plenty of programming skill, can "feel it out" and start to do useful tasks with a dozen hours practice. Even if it's a wasteful way to process/translate data, it's still huge. Electron somehow won out over native apps thanks to ease of use, for example.
mrbombastic 4y ago
Have you tried chatgpt with gpt 4 yet? my limited experience is that it is much better than 3.5. Trying to avoid the excessive hype but I was pretty impressed.

by lvl102 4y ago

haolez 4y ago

The difference is probably related to RLHF. It seems to matter a lot more than the size of the model and the hardware's capabilities. And OpenAI probably excels at that.

cubefox 4y ago
Bing uses GPT-4 with their own fine-tuning, and they had to put some artificial restrictions in place in order for the model not "misbehave". Apparently they don't use RLHF but some combination of SL and prompting. OpenAI didn't have such problems.
- saberience 4y ago
  What are SL and RLHF? I would also add it's best not assume others know acronyms like these.
  - cubefox 4y ago
    Sorry. SL = supervised learning. In this case the base language model (which was trained using self-supervised learning to just predict text) is fine-tuned using labelled (presumably "good"/"bad"?) dialogue examples. RL = reinforcement learning. RLHF = reinforcement learning from human feedback. Here human raters reward or punish the model for good/bad dialogue.
    - saberience 4y ago
      Thank you! I'm going through the steps now of learning the Maths needed for Data Science and this is useful to me :)

godshatter 4y ago

In my opinion, iterating ChatGPT and the like isn't going to get us anywhere, at least in terms of general AI. It's an interesting trick that might have some uses, but iterating it isn't going to lead to more understanding. It's effectively an extensively well-read and quick-thinking bullshitter. It's knows a lot of words and sentences, it can put them together in ways that sound creative or intelligent, but it doesn't understand anything at all about what it's actually saying. Right now, I see it as a dead end. I'd rather more effort was going into parsing language into the underlying concepts, making connections between them, and looking for new understanding in how those concepts might fit together in new ways or finding new ways of using that underlying understanding to solve problems. Basically what people, do, but better.

quonn 4y ago
We don‘t even have a path toward what you‘re suggesting. None at all. So ChatGPT is far closer to a solution.
selfhoster11 4y ago
ChatGPT (especially GPT-4) is the leanest, meanest Prolog engine known to mankind. It absolutely slays on inference tasks if you state some data and rules upfront. It knows how to translate between structured formats and natural language. That's not nothing. You could do this earlier, maybe, if you had a lot of specialised AI knowledge. Now, someone like me, with little AI training but plenty of programming skill, can "feel it out" and start to do useful tasks with a dozen hours practice. Even if it's a wasteful way to process/translate data, it's still huge. Electron somehow won out over native apps thanks to ease of use, for example.

mrbombastic 4y ago

Have you tried chatgpt with gpt 4 yet? my limited experience is that it is much better than 3.5. Trying to avoid the excessive hype but I was pretty impressed.