Andrew Chen on the entertainment industry; Nat Friedman on the lack of applications; Mark McNeilly on working applications; Ethan Mollick on the value of experience
LLMs' input is the Internet. What is on the Internet will increasingly be generated by LLMs. What happens when LLM input is largely LLM output? Will that create a feedback loop where inaccuracies and biases are reinforced and amplified? Can LLMs be trained to reduce the feedback effect by, for example, identifying and rejecting bad or biased information?
It really depends on the type of model, and exactly *how* the output is fed back as input, but as a general rule, when you do that in machine learning, the model just overfits and loses all generalizability (i.e. becomes useless).
Saw a tweet about Edison’s response to the critique that he failed over 1000 times before getting the lightbulb right. Successfully excluding alternatives.
When, not if but also not yet, AI can test multiple possibilities to get the best one, it will be helpful. This requires a problem which has a clear way to evaluate the best of alternative answer responses.
All digital games are like this, but most of life isn’t. Yet a lot of trial & error testing might be.
Accuracy remains key for business usefulness, except maybe in predicting the future.
"There's still a massive capability overhang where we are still learning how to make these models useful to people." Huge +1. I've been working in data science/machine learning for ~15 years, and when these deep learning models started gaining attention ~5-8 years ago, that was, and remains, my biggest question.
These models are super cool from a technical perspective, but so what? Almost nobody needs to have their computer recognize (e.g.) whether a given digital image is a picture of a yellowjacket vs. a wasp or whatever, but that's what these models are good at.
I'm not really up on this, but is there a danger of the AI equivalent of "Group Think"? Will they all converge on the equivalent of a single 'answer' to the detriment of consodering alternatives?
LLMs' input is the Internet. What is on the Internet will increasingly be generated by LLMs. What happens when LLM input is largely LLM output? Will that create a feedback loop where inaccuracies and biases are reinforced and amplified? Can LLMs be trained to reduce the feedback effect by, for example, identifying and rejecting bad or biased information?
It really depends on the type of model, and exactly *how* the output is fed back as input, but as a general rule, when you do that in machine learning, the model just overfits and loses all generalizability (i.e. becomes useless).
Saw a tweet about Edison’s response to the critique that he failed over 1000 times before getting the lightbulb right. Successfully excluding alternatives.
When, not if but also not yet, AI can test multiple possibilities to get the best one, it will be helpful. This requires a problem which has a clear way to evaluate the best of alternative answer responses.
All digital games are like this, but most of life isn’t. Yet a lot of trial & error testing might be.
Accuracy remains key for business usefulness, except maybe in predicting the future.
"There's still a massive capability overhang where we are still learning how to make these models useful to people." Huge +1. I've been working in data science/machine learning for ~15 years, and when these deep learning models started gaining attention ~5-8 years ago, that was, and remains, my biggest question.
These models are super cool from a technical perspective, but so what? Almost nobody needs to have their computer recognize (e.g.) whether a given digital image is a picture of a yellowjacket vs. a wasp or whatever, but that's what these models are good at.
I'm not really up on this, but is there a danger of the AI equivalent of "Group Think"? Will they all converge on the equivalent of a single 'answer' to the detriment of consodering alternatives?