3 ms·

I keep seeing this – it's a meme at this point – and I don't get it. Answering questions truthfully from a broad set of data while citing sources seems to be Ch

I keep seeing this – it's a meme at this point – and I don't get it. Answering questions truthfully from a broad set of data while citing sources seems to be ChatGPT's major weakness. It routinely "hallucinates," even with better prompting. It's much better at helping generate prose based on the prompt. I don't even buy the idea that "exponential growth" will improve things. One of the few things OpenAI is still open about is that they will keep adding more parameters at any cost. Techniques to reinforce truthful responses or even provide confidence based on whether the response involves out-of-sample & out-of-prompt guessing (i.e. BSing) seem not to be a priority.

brookst 4y ago
Gpt4 is much more accurate than gpt3.5, which was somewhat more accurate than gpt3, which was dramatically more accurate than gpt2. At some point you have to look at trends more than philosophical speculation.
- baryphonic 4y ago
  GPT4 is also much more expensive to use, even at the probably-running-at-a-loss pricing. Accuracy per dollar and/or joule doesn't seem to be increasing too much.
  - brookst 4y ago
    Sure, if we’re changing the complaint from “accuracy won’t get better” to “efficiency won’t get better.” But I don’t believe that either. Hardware design and manufacturing cycles are long and we’re just now starting to see transformer-optimized HW. Accuracy per dollar and per joule will improve by at least an order of magnitude, maybe two. But then we can move on to percent of human tasks LLMs are suitable replacements for, or the ethics of increasing automation.
    - baryphonic 4y ago
      I was critiquing accuracy of ChatGPT, which has a certain cost. "Just use GPT4, 5, ..." was the solution proposed in response, but that ain't free. In fact, it's one to two orders of magnitude more expensive.[0] I'm also pessimistic on custom hardware being that much of a help. I read about Cerberus's giant wafer chip in 2019[1] and even saw it at a conference in early 2020 pre-pandemic, and I have yet to see that make orders of magnitude worth of improvement. The attention mechanism simply requires huge amounts of multiply-adds, especially at GPT-scale. That costs energy and cash no matter how much custom hardware you throw at it. [0]https://openai.com/pricing https://openai.com/pricing [1]https://www.wired.com/story/cerebras-chip-cluster-neural-networks-ai/ https://www.wired.com/story/cerebras-chip-cluster-neural-net...

brookst 4y ago

Gpt4 is much more accurate than gpt3.5, which was somewhat more accurate than gpt3, which was dramatically more accurate than gpt2. At some point you have to look at trends more than philosophical speculation.

baryphonic 4y ago
GPT4 is also much more expensive to use, even at the probably-running-at-a-loss pricing. Accuracy per dollar and/or joule doesn't seem to be increasing too much.
- brookst 4y ago
  Sure, if we’re changing the complaint from “accuracy won’t get better” to “efficiency won’t get better.” But I don’t believe that either. Hardware design and manufacturing cycles are long and we’re just now starting to see transformer-optimized HW. Accuracy per dollar and per joule will improve by at least an order of magnitude, maybe two. But then we can move on to percent of human tasks LLMs are suitable replacements for, or the ethics of increasing automation.
  - baryphonic 4y ago
    I was critiquing accuracy of ChatGPT, which has a certain cost. "Just use GPT4, 5, ..." was the solution proposed in response, but that ain't free. In fact, it's one to two orders of magnitude more expensive.[0] I'm also pessimistic on custom hardware being that much of a help. I read about Cerberus's giant wafer chip in 2019[1] and even saw it at a conference in early 2020 pre-pandemic, and I have yet to see that make orders of magnitude worth of improvement. The attention mechanism simply requires huge amounts of multiply-adds, especially at GPT-scale. That costs energy and cash no matter how much custom hardware you throw at it. [0]https://openai.com/pricing https://openai.com/pricing [1]https://www.wired.com/story/cerebras-chip-cluster-neural-networks-ai/ https://www.wired.com/story/cerebras-chip-cluster-neural-net...