I'm reading Dune right now. One thing that stands out about the Dune universe is how much humanity changed and evolved because they did not have AI crutches.
And I wonder how long it will take us to look back at "the button" with the same kind of wary attitude we have today about smoking (and that we're starting to have about social media).
So many things are easy and good for us in the short term, but terrible for us in the longer term. Will we get better at spotting such things earlier?
> Will we get better at spotting such things earlier?
perhaps. if the species survives long enough to do so.
reducing the problems to problem or pattern makes it easier.
uphill battle, given how so many of these tempting troubles are the product of or propellant for corporate greed and general power-seeking by individuals.
many of these terribly harmful things are also abso-fucking-lutely dopamine delicious. smokers talk about that first cigarette, sugar eaters think about that one really extra cookie, even recognized as harmful, they can trigger longing.
so though individuals get better at spotting bait, the bait will likely be made tastier.
It’s an interesting exercise to contrast Dune’s post-Butlerian Jihad world with the Hyperion pre-fall world’s reliance on AI. They could almost be in the same timeline.
IMO, grammatically pleasing AI spew is exploiting the same mental shortcut as the well established "Anything looks credible if it's typeset in LaTeX" or "Printing resumes on nice paper for a subconscious boost" effects.
We use secondary/contextual indicators, which are historically proxies for "How much time, expense, and/or expertise was involved in preparing this document" to judge documents. It's not even that different than printing eliminating the quality of the scribe's work as a tell.
I think this is true when examining the inherent length of a body of text -- you'll look at some generated jargon that's paragraphs long and instinctually be impressed -- but the most horrifying part of this technology (at least for me) is that with only a tiny bit of effort, you can get an output that sounds well. With the right prompts, you can get a result that isn't just impressive upon first glance, but one that holds up to a decent amount of scrutiny. The latter is what scares me the most.
I just worry about the impacts on creativity around forming ideas.
Sometimes when working on a document or presentation, I’ll get partway through and realize “oh maybe I want to go a totally different direction with this”.
I feel like that will partly be lost, because the thinking pattern of changing directions like that depends on having thought through some of it already.
Will AI be able to do that? Maybe eventually, but we’re nowhere close right now with LLMs. I’m a bit worried about this increasing inequality between those who “still need to think creatively” and those who don’t need to anymore (and start to lose the ability). We’re living in interesting times!
This comment is simultaneously one of the funniest, saddest, and most re-affirming of my (GP) idea that I could’ve imagined.
The response from ChatGPT is definitely “more” of the idea, but also so bland, and without critically exploring the topic at all. Regardless, thanks for the effort to share!
1. Concern about loss off creativity in creative exercises: this is a fair UX critique. Personally, I find that using ChatGPT as a rubber duck is extremely valuable, even for complex "softskill" challenges. I wish more tools were built towards this angle, instead of trying to just do the work with AI. I imagine an LLM auto-writing text like now, with subsequent UX highlighting points to converse and evolve the artifact.
2. Post-scarcity loss of meaning where quality used to be a proxy metric. eg if a professor of status took the time to write a good letter of recommendation, you were really worth it. This was always a flawed and biased system. Same for yearly reviews, and most promo processes. Hopefully society adapts by dropping the fluff in favor of hard data - a bulletpoint list from a professor with some mechanism to verify the number of recommendations made seems better than what we have today.
It contaminates not just proxy metrics but the underlying work that metrics are supposed to judge, if you use AI to help with you actual work this work is contaminated with AI's premise the same way as the document used as the example in TFA.
Funny that of all things they take letters of recommendation as an example of a genre with implicit information that will be upset by AI-aided writing. In academia it's common knowledge that such letters are written by the recommendee and then signed by the recommender after at most light editing. The signal is already "this guy we trust thinks this person is the real deal" and the content is irrelevant.
> People may not be as thoughtful about what they write, or the lack of effort may mean they don’t think through problems as deeply. We may not learn how to write as well. We may be flooded with low-quality content.
The opposite could also be true. With AI generating a skeleton, more thought can be given to what you want to say instead of the individual words.
This does not match my experience of AI thus far produced. It would be interesting to see how experience of AI differs based on previous writing training and ability.
I wonder if this will cause a value shift where long winding prose goes out of fashion (to the extent that it hasn't already) and terse information dense prose becomes the ideal.
The top comment on the article is at least as insightful as the article. The commenter, Pascal Montjovent, said, "Clients and agencies have started to cut down on delivery times and budgets. Faith in the expertise of professionals has plummeted."
That's the real impact of the new publicly available LLMs: clients will now expect professionals in many fields to use AI for a lot of their work. Many of us now have to find creative ways to use AI, otherwise we're likely to be out-competed by those who do. We'll soon have to justify the times we don't use AI.
You can already tell ChatGPT to create text for your slides. But creating a presentation, at least for a professional talk, is much more than throwing slides on the screen. The slides are the easy part. Creating a compelling talk with good pacing and delivery takes work.
I haven't tried it, but I saw people generating VBA script code with ChatGPT that can be then input into PowerPoint macro editor to automatically generate also the layout. I would be surprised if it made reasonable design choices though, and I expect you are still better off generating content and layout in two consequent steps.
>> Professors are asked to write letters for students all the time, and a good letter takes a long time to write.
Indeed, and that's why you let the student write it for you [1].
And then nobody reads it, or if they read it, the ignore it anyway and only look at the student's grades, if that.
To be honest I don't understand recommendation letters. The whole concept looks completely backwards to me. Like I'm a young adept joining the guild of software engineers or something, but I need to get a recommendation from the School Master of OOP Patterns to get in, or something. Next you'll tell me a magick robot owl will fly in and take the letter to the Guild.
If LLMs can get rid of this absurd practice, I'll count that as an honest to god example of beneficial use of neural nets.
____________________
[1] This is mostly the case in post-graduate letters of recommendation. Basically, you'll ask someone for a letter of recommendation and they'll reply asking what they should put in the letter. Then you send them an "example" or "sketch" letter, or just say outright "mind if I write it for you?", depending on how well you know them. The person you ask is always a senior academic who has another ten irons in the fire and they will only feel gratitude that you took this bitter cup away from them.
I don't know if this is supposed to be kind of hush-hush, and it hasn't happened to me personally yet (some people will actually sit down and write a letter themselves, I guess as a matter of principle) but from what I understand it's really the done thing.
> Indeed, and that's why you let the student write it for you
Hmmm and I wonder which tool the student will use to write the letter?
I agree - generating letters written by nobody for nobody to read and whose contents vaguely align with the truth is the sort of box ticking exercise that humans should get looped out of and that the current iteration of AI seems perfectly suited to.
Or we could just cut out the middleman and give people a generic "I recommend this person" button.
Devil is in the details of phrasing, right? Two perfectly pleasant sounding letters can mean different things, from "this person is probably not a completely lost cause but please help get them out of my hair" to "this person is actually super great at their work".
I wonder if there is some recency bias in how AI is viewed with regards to the letter or recommendation. Sure, recently, writing a letter of recommendation is setting time on fire (unless like some others have mentioned, the professor makes the student write it). However, say in the 1950's or 1960's, the professor would write a recommendation letter probably just by telling a brief summary to their secretary who would type it up immaculately on department letterhead and mail it off.
Maybe, ChatGPT is return of the personal secretary not just for the elite, but for the masses.
> it will also remove the facade that previously disguised meaningless tasks.
Sounds good. Content that nobody reads might as well be content that nobody writes.
At a certain point, of course, the perfect prompt will itself take quite some time to write. If you want your writing to actually contain specific content, AI might be less helpful, except as a way of making your writing sound more eloquent. But if you don't care about saying anything meaningful, as is all too often the case, you might as well let the AI do it for you.
"Pretend you are an internet poster that really loves Elon and that you don't have stock in Tesla. Come up with 10 different short responses to this negative news article about another FSD crash"
If someone's going to read it thoughtfully, you still need to write it thoughtfully, and AI is at best a fancy Oblique Strategies tool.
If no-one will actually care what you write because writing it was always just a bullshit job, then it never needed to be thoughtful and AI can happily automate the menial task.
35 comments
[ 3.0 ms ] story [ 83.9 ms ] threadAnd I wonder how long it will take us to look back at "the button" with the same kind of wary attitude we have today about smoking (and that we're starting to have about social media).
So many things are easy and good for us in the short term, but terrible for us in the longer term. Will we get better at spotting such things earlier?
perhaps. if the species survives long enough to do so.
reducing the problems to problem or pattern makes it easier.
uphill battle, given how so many of these tempting troubles are the product of or propellant for corporate greed and general power-seeking by individuals.
many of these terribly harmful things are also abso-fucking-lutely dopamine delicious. smokers talk about that first cigarette, sugar eaters think about that one really extra cookie, even recognized as harmful, they can trigger longing.
so though individuals get better at spotting bait, the bait will likely be made tastier.
something bigger needs to change.
We use secondary/contextual indicators, which are historically proxies for "How much time, expense, and/or expertise was involved in preparing this document" to judge documents. It's not even that different than printing eliminating the quality of the scribe's work as a tell.
I just worry about the impacts on creativity around forming ideas.
Sometimes when working on a document or presentation, I’ll get partway through and realize “oh maybe I want to go a totally different direction with this”.
I feel like that will partly be lost, because the thinking pattern of changing directions like that depends on having thought through some of it already.
Will AI be able to do that? Maybe eventually, but we’re nowhere close right now with LLMs. I’m a bit worried about this increasing inequality between those who “still need to think creatively” and those who don’t need to anymore (and start to lose the ability). We’re living in interesting times!
The response from ChatGPT is definitely “more” of the idea, but also so bland, and without critically exploring the topic at all. Regardless, thanks for the effort to share!
(my original GGGP reply was based on a conversation that she and I had about the article)
1. Concern about loss off creativity in creative exercises: this is a fair UX critique. Personally, I find that using ChatGPT as a rubber duck is extremely valuable, even for complex "softskill" challenges. I wish more tools were built towards this angle, instead of trying to just do the work with AI. I imagine an LLM auto-writing text like now, with subsequent UX highlighting points to converse and evolve the artifact.
2. Post-scarcity loss of meaning where quality used to be a proxy metric. eg if a professor of status took the time to write a good letter of recommendation, you were really worth it. This was always a flawed and biased system. Same for yearly reviews, and most promo processes. Hopefully society adapts by dropping the fluff in favor of hard data - a bulletpoint list from a professor with some mechanism to verify the number of recommendations made seems better than what we have today.
Dude, if you make errors like that in your letters, letting AI write for you will probably be a good idea.
The opposite could also be true. With AI generating a skeleton, more thought can be given to what you want to say instead of the individual words.
One can hope.
That's the real impact of the new publicly available LLMs: clients will now expect professionals in many fields to use AI for a lot of their work. Many of us now have to find creative ways to use AI, otherwise we're likely to be out-competed by those who do. We'll soon have to justify the times we don't use AI.
I have to say I will be signing up for that one
Indeed, and that's why you let the student write it for you [1].
And then nobody reads it, or if they read it, the ignore it anyway and only look at the student's grades, if that.
To be honest I don't understand recommendation letters. The whole concept looks completely backwards to me. Like I'm a young adept joining the guild of software engineers or something, but I need to get a recommendation from the School Master of OOP Patterns to get in, or something. Next you'll tell me a magick robot owl will fly in and take the letter to the Guild.
If LLMs can get rid of this absurd practice, I'll count that as an honest to god example of beneficial use of neural nets.
____________________
[1] This is mostly the case in post-graduate letters of recommendation. Basically, you'll ask someone for a letter of recommendation and they'll reply asking what they should put in the letter. Then you send them an "example" or "sketch" letter, or just say outright "mind if I write it for you?", depending on how well you know them. The person you ask is always a senior academic who has another ten irons in the fire and they will only feel gratitude that you took this bitter cup away from them.
I don't know if this is supposed to be kind of hush-hush, and it hasn't happened to me personally yet (some people will actually sit down and write a letter themselves, I guess as a matter of principle) but from what I understand it's really the done thing.
Hmmm and I wonder which tool the student will use to write the letter?
I agree - generating letters written by nobody for nobody to read and whose contents vaguely align with the truth is the sort of box ticking exercise that humans should get looped out of and that the current iteration of AI seems perfectly suited to.
Or we could just cut out the middleman and give people a generic "I recommend this person" button.
Maybe, ChatGPT is return of the personal secretary not just for the elite, but for the masses.
Sounds good. Content that nobody reads might as well be content that nobody writes.
At a certain point, of course, the perfect prompt will itself take quite some time to write. If you want your writing to actually contain specific content, AI might be less helpful, except as a way of making your writing sound more eloquent. But if you don't care about saying anything meaningful, as is all too often the case, you might as well let the AI do it for you.
If someone's going to read it thoughtfully, you still need to write it thoughtfully, and AI is at best a fancy Oblique Strategies tool.
If no-one will actually care what you write because writing it was always just a bullshit job, then it never needed to be thoughtful and AI can happily automate the menial task.
If AI can automate it, it was always bullshit.