A few weeks ago, I asked ChatGPT to give me a cogent argument for making TEAC a Village BOARD rather than an advisory council, including pros and cons. I was amazed at the insight the resulting essay had into this somewhat obscure topic, and that it included bulleted lists for both pros and cons. It got a few assumptions wrong, but was at least as readable and logical as I could have written myself.
On the issue of detectability, I could envision a way for AI to be required to include a rhetorical watermark of some type, such as a sentence with a particular, unusual structure, that a verifier could detect and improve its accuracy. And, let's hope that AI DOES develop a way for itself to determine the veracity of source materials it incorporates, developing a metric to reduce the influence of low-scoring sources in its produced content.
Glad that that TEAC request yielded good results - it's always so helpful to have something to start with at least. Good idea on the "rhetorical watermark." And some kind of citing of sources needs to happen for sure! Thanks for reading! (Part 2 next week, so there's much on this subject..)
I am deeply skeptical that any sort of pause, noble as it might be, will happen. I think we have to adjust our strategy and confine this concept to the "would have been really nice if we could have done this instead" category. I'm not super pessimistic that this must end badly either, but the idea that we humans are going to be able to turn this tap off is probably fantasy.
"Trim the nails" is a great way to put that. I've concluded (not surprisingly, perhaps) that the best thing people like us can do about this is to raise attention. With that said, you're doing it! People like us need to keep thinking and talking about nuanced, complex topics like this (and especially this).
I don't know I still think the paranoia around AI is an over reaction. I have a hard time thinking it's going to have a bigger impact than the regular old internet. People already believe the first thing they see on Google and social media. And at the moment, these machine learning models cannot write their own code to make themselves better. Once the models are deployed they are static and have no feedback loop.
Great post. From my own experience, I share your concern/alarm, especially from the perspective of a writer. I find myself bouncing back and forth hating the fact that we are in this reality and feeling the pressure to incorporate/leverage AI tools into my own work. At the same time, I can also see the advantages.
My largest communications client recently started incorporating GPT-4 into the comms process, so I felt like I had no choice but to upgrade my personal free account to paid. So my first point: GPT-4 produces *much* better results than GPT-3.5. It still has problems -- a recent blog post draft we pumped out made up convincing-sounding references, claimed that other references were saying what they weren't, etc. Basically, it forced me to do a lot of fact-checking and editing (but this is more than half the job anyway?).
Much more useful has been pasting a blog post in and asking it to generate ten or fifteen insightful tweets or a bunch of headlines. Many of these are better than anything I would have come up with.
The key is telling it exactly what you want. In "default mode" it will generally produce bland results that you would expect if you tried to average all the content you find on the internet into mixing bowl. If you really want to stress-test it though, analyze what bothers you about a response that it spits out, and ask it to rewrite it with that in mind.
For your hoarding post for example, try "write a short blog post on how to hoard in a healthy way. Don't use bullet points. Add some personal anecdotes to support your point. Use appropriate metaphor." (Again, results for this will probably be better with GPT-4). Then for fun, ask it to rewrite in the style Joan Didion.
FWIW that ZD Net article is wrong, Chat can definitely produce metaphor, irony and sarcasm. I've prompted it to come with some really beautiful stuff, but it's quirky, and there's still a question of intent: this metaphor is beautiful, but is it what *I* want to say?
And if I'm crafting and recrafting careful prompts to get it to spit out something I could use, is it the AI doing the writing -- or is it just me? Corollary question: am I wasting time when I could have come up with this on my own?
At least in the short term I think the key question for most writers isn't going to be "will this replace me" but how/if I can/should use this tool to augment my own process.
oh so interesting Lakis. Now you make me want to do a second post where I go further on this into the "stress testing" - and make more specific demands, and also I'm very tempted to try out the 4 version for a month or something. That's a good idea about using it as a tweet/headline generator for what you already have. Really good points - if it's used properly, it's a good tool, not any replacement.
A few weeks ago, I asked ChatGPT to give me a cogent argument for making TEAC a Village BOARD rather than an advisory council, including pros and cons. I was amazed at the insight the resulting essay had into this somewhat obscure topic, and that it included bulleted lists for both pros and cons. It got a few assumptions wrong, but was at least as readable and logical as I could have written myself.
On the issue of detectability, I could envision a way for AI to be required to include a rhetorical watermark of some type, such as a sentence with a particular, unusual structure, that a verifier could detect and improve its accuracy. And, let's hope that AI DOES develop a way for itself to determine the veracity of source materials it incorporates, developing a metric to reduce the influence of low-scoring sources in its produced content.
Glad that that TEAC request yielded good results - it's always so helpful to have something to start with at least. Good idea on the "rhetorical watermark." And some kind of citing of sources needs to happen for sure! Thanks for reading! (Part 2 next week, so there's much on this subject..)
I am deeply skeptical that any sort of pause, noble as it might be, will happen. I think we have to adjust our strategy and confine this concept to the "would have been really nice if we could have done this instead" category. I'm not super pessimistic that this must end badly either, but the idea that we humans are going to be able to turn this tap off is probably fantasy.
I agree, it's unrealistic to think we can put the cat back in the bag. I hope we might at least trim its nails though.
"Trim the nails" is a great way to put that. I've concluded (not surprisingly, perhaps) that the best thing people like us can do about this is to raise attention. With that said, you're doing it! People like us need to keep thinking and talking about nuanced, complex topics like this (and especially this).
yes!
I don't know I still think the paranoia around AI is an over reaction. I have a hard time thinking it's going to have a bigger impact than the regular old internet. People already believe the first thing they see on Google and social media. And at the moment, these machine learning models cannot write their own code to make themselves better. Once the models are deployed they are static and have no feedback loop.
I hope you're right!
Great post. From my own experience, I share your concern/alarm, especially from the perspective of a writer. I find myself bouncing back and forth hating the fact that we are in this reality and feeling the pressure to incorporate/leverage AI tools into my own work. At the same time, I can also see the advantages.
My largest communications client recently started incorporating GPT-4 into the comms process, so I felt like I had no choice but to upgrade my personal free account to paid. So my first point: GPT-4 produces *much* better results than GPT-3.5. It still has problems -- a recent blog post draft we pumped out made up convincing-sounding references, claimed that other references were saying what they weren't, etc. Basically, it forced me to do a lot of fact-checking and editing (but this is more than half the job anyway?).
Much more useful has been pasting a blog post in and asking it to generate ten or fifteen insightful tweets or a bunch of headlines. Many of these are better than anything I would have come up with.
The key is telling it exactly what you want. In "default mode" it will generally produce bland results that you would expect if you tried to average all the content you find on the internet into mixing bowl. If you really want to stress-test it though, analyze what bothers you about a response that it spits out, and ask it to rewrite it with that in mind.
For your hoarding post for example, try "write a short blog post on how to hoard in a healthy way. Don't use bullet points. Add some personal anecdotes to support your point. Use appropriate metaphor." (Again, results for this will probably be better with GPT-4). Then for fun, ask it to rewrite in the style Joan Didion.
FWIW that ZD Net article is wrong, Chat can definitely produce metaphor, irony and sarcasm. I've prompted it to come with some really beautiful stuff, but it's quirky, and there's still a question of intent: this metaphor is beautiful, but is it what *I* want to say?
And if I'm crafting and recrafting careful prompts to get it to spit out something I could use, is it the AI doing the writing -- or is it just me? Corollary question: am I wasting time when I could have come up with this on my own?
At least in the short term I think the key question for most writers isn't going to be "will this replace me" but how/if I can/should use this tool to augment my own process.
oh so interesting Lakis. Now you make me want to do a second post where I go further on this into the "stress testing" - and make more specific demands, and also I'm very tempted to try out the 4 version for a month or something. That's a good idea about using it as a tweet/headline generator for what you already have. Really good points - if it's used properly, it's a good tool, not any replacement.
I have warm fuzzies.
no competition for YOU!
<3