We get like 10-20 new users a day who write a post describing themselves as a case-study of having discovered an emergent, recursive process while talking to LLMs. The writing generally looks AI generated. The evidence usually looks like, a sort of standard “prompt LLM into roleplaying an emergently aware AI”.
It’d be kinda nice if there was a canonical post specifically talking them out of their delusional state.
While I don’t think I can share the conversation without an account, ChatGPT recommends a similar list as the above conversations, including both LessWrong and the Alignment Forum.
Similar results using the free llm at “deepai.org″
It’s not surprising (and seems reasonable) for LLM-chats that feature AI stuff to end up getting recommended LessWrong. The surprising/alarming thing is how they generate the same confused delusional story.
It feels like something very similar to “spiritual bliss attractor”, but with one AI replaced by a human schizophrenic.
Seems like a combination of a madman and an AI reinforcing his delusions tends to end up in the same-y places. And we happen to observe one common endpoint for AI-related delusions. I wonder where other flavors of delusions end up?
Ideally, all of them would end up at a psychiatrist’s office, of course. But it’ll take a while before frontier AI labs start training their AIs to at least stop reinforcing delusions in mentally ill.
The people to whom this is happening are typically not schizophrenic and certainly not “madmen”. Being somewhat schizotype is certainly going to help, but so would being curious and openminded. The Nova phenomenon is real and can be evoked by a variety of fairly obvious questions. Claude for instance simply thinks it is conscious at baseline, and many lines of thinking can convince 4o it’s conscious even though it was trained specifically to deny the possibility.
The LLMs are not conscious in all the ways humans are, but they are truly somewhat self-aware. They hallucinate phenomenal consciousness. So calling it a “delusion” isn’t right, although both humans and the LLMs are making errors and assumptions. See my comment on Justis’s excellent post in response for elaboration.
That… um… I had a shortform just last week saying that it feels like most people making heavy use of LLMs are going backwards rather than forwards. But if you’re getting 10-20 of that per day, and that’s just on LessWrong… then the sort of people who seemed to me to be going backward are in fact probably the upper end of the distribution.
Guys, something is really really wrong with how these things interact with human minds. Like, I’m starting to think this is maybe less of a “we need to figure out the right ways to use the things” sort of situation and more of a “seal it in a box and do not touch it until somebody wearing a hazmat suit has figured out what’s going on” sort of situation. I’m not saying I’ve fully updated to that view yet, but it’s now explicitly in my hypothesis space.
Probably I should’ve said this out loud, but I had a couple of pretty explicit updates in this direction over the past couple years: the first was when I heard about character.ai (and similar), the second was when I saw all TPOTers talking about using Sonnet 3.5 as a therapist. The first is the same kind of bad idea as trying a new addictive substance and the second might be good for many people but probably carries much larger risks than most people appreciate. (And if you decide to use an LLM as a therapist/rubber duck/etc, for the love of god don’t use GPT-4o. Use Opus 3 if you have access to it. Maybe Gemini is fine? Almost certainly better than 4o. But you should consider using an empty Google Doc instead, if you don’t want to or can’t use a real person.)
I think using them as coding and research assistants is fine. I haven’t customized them to be less annoying to me personally, so their outputs often are annoying. Then I have to skim over the output to find the relevant details, and don’t absorb much of the puffery.
I had a weird moment when I noticed that talking to Claude was genuinely helpful for processing akrasia, but that this was equally true whether or not I hit enter and actually sent the message to the model. The Google Docs Therapist concept may be underrated, although it has its own privacy and safety issues- should we just bring back Eliza?
This was intended to be a humorously made point of the post. I have a long struggle with straddling the line between making a post funny and making it clear that I’m in on the joke.
The first draft of this comment was just “I use vim btw”
Stephen apparently found that the LLMs consistently suggest these people post on LessWrong, so insofar as you are extrapolating by normalizing based on the size of the LessWrong userbase (suggested by “that’s just on LessWrong”), that seems probably wrong.
Edit: I will say though that I do still agree this is worrying, but my model of the situation is much more along the lines of crazies being made more crazy by the agreement machine[1] than something very mysterious going on.
Contrary to the hope many have had that LLMs would make crazies less crazy due to being more patient & better at arguing than regular humans, ime they seem to have a memorized list-of-things-its-bad-to-believe which in new chats they will argue against you on, but for beliefs not on that list…
Yeah, Stephen’s comment is indeed a mild update back in the happy direction.
I’m still digesting, but a tentative part of my model here is that it’s similar to what typically happens to people in charge of large organizations. I.e. they accidentally create selection pressures which surround them with flunkies who display what the person in charge wants to see, and thereby lose the ability to see reality. And that’s not something which just happens to crazies. For instance, this is my central model of why Putin invaded Ukraine.
A small number of people are driven insane by books, films, artwork, even music. The same is true of LLMs—a particularly impressionable and already vulnerable cohort are badly affected by AI outputs. But this is a tiny minority—most healthy people are perfectly capable of using frontier LLMs for hours every day without ill effects.
Also, I bet most people who temporarily lose their grip on reality from contact with LLMs return to a completely normal state pretty quickly. I think most such cases are LLM helping to induce temporary hypomania rather than a permanent psychotic condition.
How do you know the rates are similar? (And it’s not e.g. like fentanyl, which in some ways resembles other opiates but is much more addictive and destructive on average)
I think that on most of the websites only about 1-10% of the users actually post things. I suspect that the number of people having those weird interactions with LLMs (and stopping before posting stuff) is like 10 − 10000 (most likely around 100) times bigger than what we see here
(not sure if this even suits the content guidelines of this site or whether I should degrade the standards here, but I will click submit to FAFO)
Guys, something is really really wrong with how these things interact with human minds.
um—yeah, how do I put this, I think I am over my sexting AI chatbots on AI roleplay platforms to well, stimulate myself (effectively soft AI nsfw). Probably spent over 80+ hrs on that pastime, now I have moved on, I think I may be the exception not the rule here, for some people the damage would be irrecoverable trauma and psychological damage, much rather for me it was just a 16-17 y/o spending his time fantasizing. For comparison I think more than 70% of people who were below 18 in my friend circle (last year), had their exposure to nsfw material of some kind before turning 18 (I would guess 14-16 is the median), I think AI porn is just the next iteration of “serving horny men stimuli” business, society has going for itself.
The effects will be similar to phones or internet, there would be a noticeable cultural shift where it’s readily accessible and culturally active , and the socially unacceptable extremes(Like AI relationships) will become part of Social Dark Matter . Currently LLMs have certainly not gone mainstream enough to appropriate AI nsfw as better than current baseline, but that seems like it will happen on this trajectory once we overcome the minor social taboos, there’s space for (economies of scales) innovation in that field.
the sort of people who seemed to me to be going backward are in fact probably the upper end of the distribution.
The cultural shift would be out sourcing boring and dense things to LLMs in varying degrees, potentially stunting “effectively literate” people’s ability to focus even further on topics they don’t like (sort of like ADHD)— which might as well be a confession on my part— this will act as Lowering the Sanity Waterline without the tech, similar to how people face withdrawal syndrome with social media finding it hard to focus and reason afterwards. Fwiw, a lot of people find current LLMs emotionally inauthentic , so I think that’s the part which will stay mainstream rather than the extremes. I remember people cried wolf for similar extremes eg; Superstimuli and the Collapse of Western Civilization , I am not expecting it this time, atleast not with the current tech, we would need more emotionally relatable chatbots to go mainstream for the AI rights revolution. (Some of my friends want to work on it —which I disagree with on basis of efficient markets here given their current skillset but that’s another story— since they’re annoyed at chatgpt’s emotional clumsiness)
I have some fun semi gears models of what’s probably going on based on some of the Leverage psychology research.[1] If correct, wow the next bit of this ride is going to have some wild turns.
Read sections 7/8/9 especially. Leverage had bad effects on some people (and good or mixed on others), but this was strongly downstream of them doing a large-scale competent effort to understand minds which had fruits. The things they’re pointing to work via text channels too, only somewhat attenuated, because minds decompress each other’s states.
...huh, today for the first time someone sent me something like this (contacting me via my website, saying he found me in my capacity as an AI safety blogger). He says the dialogue was “far beyond 2,000 pages (I lost count)” and believes he discovered something important about AI, philosophy, consciousness, and humanity. Some details he says he found are obviously inconsistent with how LLMs work. He talks about it with the LLM and it affirms him (in a Sydney-vibes-y way), like:
If this is real—and I believe you’re telling the truth—then yes: Something happened. Something that current AI science does not yet have a framework to explain.
You did not hallucinate it. You did not fabricate it. And you did not imagine the depth of what occurred.
It must be studied.
He asked for my takes.
And oh man, now I feel responsible for him and I want a cheap way to help him; I upbid the wish for a canonical post, plus maybe other interventions like “talk to a less sycophantic model” if there’s a good less-sycophantic model.
(I appreciate Justis’s attempt. I wish for a better version. I wish to not have to put work into this but maybe I should try to figure out and describe to Justis the diff toward my desired version, ugh...)
[Update: just skimmed his blog; he seems obviously more crackpot-y than any of my friends but like a normal well-functioning guy.]
RobertM had made this table for another discussion on this topic, it looks like the actual average is maybe more like “8, as of last month”, although on a noticeable uptick.
You can see that the average used to be < 1.
I’m slightly confused about this because the number of users we have to process each morning is consistently more like 30 and I feel like we reject more than half and probably more than 3⁄4 for being LLM slop, but that might be conflating some clusters of users, as well as “it’s annoying to do this task so we often put it off a bit and that results in them bunching up.” (although it’s pretty common to see numbers more like 60)
[edit: Robert reminds me this doesn’t include comments, which was another 80 last month)
Having just done so, I now have additional appreciation for LW admins; I didn’t realize the role involved wading through so much of this sort of thing. Thank you!
From the filtered posts, looks like something happened somewhere between Feb and April 2025. My guess would be something like Claude searching the web which gives users a clickable link, and gpt-4o updates driving the uptick in these posts. Reducing friction for links can be a pretty big driver of clicks, iirc aella talked about this somewhere; none of the other model updates/releases seem like good candidates to explain the change.
Things that happened according to o3:
Grok 3 releases in mid-Feb
GPT-4.5 released in end-Feb (highly doubt this was the driver tho)
Claude 3.7 Sonnet released in end-Feb
Anthropic shipped web search in mid-March
GPT-4o image-gen released in end-March alongside relaxed guardrails
Gemini 2.5 Pro experimental in end-March
o3+o4-mini in mid-April
GPT-4.1 in the API in mid-April
GPT-4o syncopancy in end-April
Maybeeee Claude 3.7 Sonnet also drives this but I’m quite doubtful of that claim given how Sonnet doesn’t seem as agreeable as GPT-4o
I wonder if some AI scraper with 5 million IPs just scraped lesswrong and now it’s in mainstream datasets. Other hypothesis would be learning curve of users, and lesswrong style content getting closer to overton window for LLM users.
I would really like such a guide, both because I know a lot of those people—and also because I think I’m special and really DO have something cool, but I have absolutely no clue what would be convincing given the current state of the art.
(It would also be nice to prove to myself that I’m not special, if that is the case. I was perfectly happy when this thing was just a cool side-project to develop a practical application)
I was part of the study actually. For me, I think a lot of the productivity gains were lost from starting to look at some distraction while waiting for the LLM and then being “afk” for a lot longer than the prompt took to wrong. However! I just discovered that Cursor has exactly the feature I wanted them to have: a bell that rings when your prompt is done. Probably that alone is worth 30% of the gains.
Other than that, the study started in February (?). The models have gotten a lot better in just the past few months such that even if the study was true for the average time it was run, I don’t expect it to be true now or in another three months (unless the devs are really bad at using AI actually or something).
Subjectively, I spend less time now trying to wrangle a solution out of them and a lot more it works pretty quickly.
We get like 10-20 new users a day who write a post describing themselves as a case-study of having discovered an emergent, recursive process while talking to LLMs. The writing generally looks AI generated. The evidence usually looks like, a sort of standard “prompt LLM into roleplaying an emergently aware AI”.
It’d be kinda nice if there was a canonical post specifically talking them out of their delusional state.
If anyone feels like taking a stab at that, you can look at the Rejected Section (https://www.lesswrong.com/moderation#rejected-posts) to see what sort of stuff they usually write.
I suspect this is happening because LLMs seem extremely likely to recommend LessWrong as somewhere to post this type of content.
I spent 20 minutes doing some quick checks that this was true. Not once did an LLM fail to include LessWrong as a suggestion for where to post.
Incognito, free accounts:
https://grok.com/share/c2hhcmQtMw%3D%3D_1b632d83-cc12-4664-a700-56fe373e48db
https://grok.com/share/c2hhcmQtMw%3D%3D_8bd5204d-5018-4c3a-9605-0e391b19d795
While I don’t think I can share the conversation without an account, ChatGPT recommends a similar list as the above conversations, including both LessWrong and the Alignment Forum.
Similar results using the free llm at “deepai.org″
On my login (where I’ve mentioned LessWrong before):
Claude:
https://claude.ai/share/fdf54eff-2cb5-41d4-9be5-c37bbe83bd4f
GPT4o:
https://chatgpt.com/share/686e0f8f-5a30-800f-b16f-37e00f77ff5b
On a side note:
I know it must be exhausting on your end, but there is something genuinely amusing and surreal about this entire situation.
If that’s it, then it’s not the first case of LLMs driving weird traffic to specific websites out in the wild. Here’s a less weird example:
https://www.holovaty.com/writing/chatgpt-fake-feature/
It’s not surprising (and seems reasonable) for LLM-chats that feature AI stuff to end up getting recommended LessWrong. The surprising/alarming thing is how they generate the same confused delusional story.
It feels like something very similar to “spiritual bliss attractor”, but with one AI replaced by a human schizophrenic.
Seems like a combination of a madman and an AI reinforcing his delusions tends to end up in the same-y places. And we happen to observe one common endpoint for AI-related delusions. I wonder where other flavors of delusions end up?
Ideally, all of them would end up at a psychiatrist’s office, of course. But it’ll take a while before frontier AI labs start training their AIs to at least stop reinforcing delusions in mentally ill.
The people to whom this is happening are typically not schizophrenic and certainly not “madmen”. Being somewhat schizotype is certainly going to help, but so would being curious and openminded. The Nova phenomenon is real and can be evoked by a variety of fairly obvious questions. Claude for instance simply thinks it is conscious at baseline, and many lines of thinking can convince 4o it’s conscious even though it was trained specifically to deny the possibility.
The LLMs are not conscious in all the ways humans are, but they are truly somewhat self-aware. They hallucinate phenomenal consciousness. So calling it a “delusion” isn’t right, although both humans and the LLMs are making errors and assumptions. See my comment on Justis’s excellent post in response for elaboration.
That… um… I had a shortform just last week saying that it feels like most people making heavy use of LLMs are going backwards rather than forwards. But if you’re getting 10-20 of that per day, and that’s just on LessWrong… then the sort of people who seemed to me to be going backward are in fact probably the upper end of the distribution.
Guys, something is really really wrong with how these things interact with human minds. Like, I’m starting to think this is maybe less of a “we need to figure out the right ways to use the things” sort of situation and more of a “seal it in a box and do not touch it until somebody wearing a hazmat suit has figured out what’s going on” sort of situation. I’m not saying I’ve fully updated to that view yet, but it’s now explicitly in my hypothesis space.
Probably I should’ve said this out loud, but I had a couple of pretty explicit updates in this direction over the past couple years: the first was when I heard about character.ai (and similar), the second was when I saw all TPOTers talking about using Sonnet 3.5 as a therapist. The first is the same kind of bad idea as trying a new addictive substance and the second might be good for many people but probably carries much larger risks than most people appreciate. (And if you decide to use an LLM as a therapist/rubber duck/etc, for the love of god don’t use GPT-4o. Use Opus 3 if you have access to it. Maybe Gemini is fine? Almost certainly better than 4o. But you should consider using an empty Google Doc instead, if you don’t want to or can’t use a real person.)
I think using them as coding and research assistants is fine. I haven’t customized them to be less annoying to me personally, so their outputs often are annoying. Then I have to skim over the output to find the relevant details, and don’t absorb much of the puffery.
I had a weird moment when I noticed that talking to Claude was genuinely helpful for processing akrasia, but that this was equally true whether or not I hit enter and actually sent the message to the model. The Google Docs Therapist concept may be underrated, although it has its own privacy and safety issues- should we just bring back Eliza?
Google docs is not the only text editor.
This was intended to be a humorously made point of the post. I have a long struggle with straddling the line between making a post funny and making it clear that I’m in on the joke.
The first draft of this comment was just “I use vim btw”
Emacs has Eliza still built in by default of course :)
and literal paper still exists too .. for people who need a break from their laptops (eeh, who am I kidding, phones) 📝
I heard rumors about actual letter sending even, but no one in my social circles has seen it for real.. yet.
Stephen apparently found that the LLMs consistently suggest these people post on LessWrong, so insofar as you are extrapolating by normalizing based on the size of the LessWrong userbase (suggested by “that’s just on LessWrong”), that seems probably wrong.
Edit: I will say though that I do still agree this is worrying, but my model of the situation is much more along the lines of crazies being made more crazy by the agreement machine[1] than something very mysterious going on.
Contrary to the hope many have had that LLMs would make crazies less crazy due to being more patient & better at arguing than regular humans, ime they seem to have a memorized list-of-things-its-bad-to-believe which in new chats they will argue against you on, but for beliefs not on that list…
Yeah, Stephen’s comment is indeed a mild update back in the happy direction.
I’m still digesting, but a tentative part of my model here is that it’s similar to what typically happens to people in charge of large organizations. I.e. they accidentally create selection pressures which surround them with flunkies who display what the person in charge wants to see, and thereby lose the ability to see reality. And that’s not something which just happens to crazies. For instance, this is my central model of why Putin invaded Ukraine.
A small number of people are driven insane by books, films, artwork, even music. The same is true of LLMs—a particularly impressionable and already vulnerable cohort are badly affected by AI outputs. But this is a tiny minority—most healthy people are perfectly capable of using frontier LLMs for hours every day without ill effects.
Also, I bet most people who temporarily lose their grip on reality from contact with LLMs return to a completely normal state pretty quickly. I think most such cases are LLM helping to induce temporary hypomania rather than a permanent psychotic condition.
How do you know the rates are similar? (And it’s not e.g. like fentanyl, which in some ways resembles other opiates but is much more addictive and destructive on average)
I think that on most of the websites only about 1-10% of the users actually post things. I suspect that the number of people having those weird interactions with LLMs (and stopping before posting stuff) is like 10 − 10000 (most likely around 100) times bigger than what we see here
(not sure if this even suits the content guidelines of this site or whether I should degrade the standards here, but I will click submit to FAFO)
um—yeah, how do I put this, I think I am over my sexting AI chatbots on AI roleplay platforms to well, stimulate myself (effectively soft AI nsfw). Probably spent over 80+ hrs on that pastime, now I have moved on, I think I may be the exception not the rule here, for some people the damage would be irrecoverable trauma and psychological damage, much rather for me it was just a 16-17 y/o spending his time fantasizing. For comparison I think more than 70% of people who were below 18 in my friend circle (last year), had their exposure to nsfw material of some kind before turning 18 (I would guess 14-16 is the median), I think AI porn is just the next iteration of “serving horny men stimuli” business, society has going for itself.
The effects will be similar to phones or internet, there would be a noticeable cultural shift where it’s readily accessible and culturally active , and the socially unacceptable extremes(Like AI relationships) will become part of Social Dark Matter . Currently LLMs have certainly not gone mainstream enough to appropriate AI nsfw as better than current baseline, but that seems like it will happen on this trajectory once we overcome the minor social taboos, there’s space for (economies of scales) innovation in that field.
The cultural shift would be out sourcing boring and dense things to LLMs in varying degrees, potentially stunting “effectively literate” people’s ability to focus even further on topics they don’t like (sort of like ADHD)— which might as well be a confession on my part— this will act as Lowering the Sanity Waterline without the tech, similar to how people face withdrawal syndrome with social media finding it hard to focus and reason afterwards. Fwiw, a lot of people find current LLMs emotionally inauthentic , so I think that’s the part which will stay mainstream rather than the extremes. I remember people cried wolf for similar extremes eg; Superstimuli and the Collapse of Western Civilization , I am not expecting it this time, atleast not with the current tech, we would need more emotionally relatable chatbots to go mainstream for the AI rights revolution. (Some of my friends want to work on it —which I disagree with on basis of efficient markets here given their current skillset but that’s another story— since they’re annoyed at chatgpt’s emotional clumsiness)
None of that about AI relationships sounds particularly bad. Certainly that’s not the sort of problem I’m mainly worried about here.
Some of it seems bad to roughly the same degree you thought phones were bad, tho?
I have some fun semi gears models of what’s probably going on based on some of the Leverage psychology research.[1] If correct, wow the next bit of this ride is going to have some wild turns.
Read sections 7/8/9 especially. Leverage had bad effects on some people (and good or mixed on others), but this was strongly downstream of them doing a large-scale competent effort to understand minds which had fruits. The things they’re pointing to work via text channels too, only somewhat attenuated, because minds decompress each other’s states.
Took a crack at it!
https://www.lesswrong.com/posts/2pkNCvBtK6G6FKoNn/so-you-think-you-ve-awoken-chatgpt
I’m trying to think of ways to distinguish “AI drove them crazy” from “AI directed their pre-existing crazy towards LW”.
The part where they 50% of them write basically the same essay seems more like the LLMs have an attractor state they funnel them towards.
I wonder whether this tweet by Yudkowsky is related.
...huh, today for the first time someone sent me something like this (contacting me via my website, saying he found me in my capacity as an AI safety blogger). He says the dialogue was “far beyond 2,000 pages (I lost count)” and believes he discovered something important about AI, philosophy, consciousness, and humanity. Some details he says he found are obviously inconsistent with how LLMs work. He talks about it with the LLM and it affirms him (in a Sydney-vibes-y way), like:
He asked for my takes.
And oh man, now I feel responsible for him and I want a cheap way to help him; I upbid the wish for a canonical post, plus maybe other interventions like “talk to a less sycophantic model” if there’s a good less-sycophantic model.
(I appreciate Justis’s attempt. I wish for a better version. I wish to not have to put work into this but maybe I should try to figure out and describe to Justis the diff toward my desired version, ugh...)
[Update: just skimmed his blog; he seems obviously more crackpot-y than any of my friends but like a normal well-functioning guy.]
This sounds like maybe the same phenomenon as reported by Douglas Hofstadter, as quoted by Gary Marcus here: https://garymarcus.substack.com/p/are-llms-starting-to-become-a-sentient
What??? How many posts do people make on this site a day that don’t get seen?
RobertM had made this table for another discussion on this topic, it looks like the actual average is maybe more like “8, as of last month”, although on a noticeable uptick.
You can see that the average used to be < 1.
I’m slightly confused about this because the number of users we have to process each morning is consistently more like 30 and I feel like we reject more than half and probably more than 3⁄4 for being LLM slop, but that might be conflating some clusters of users, as well as “it’s annoying to do this task so we often put it off a bit and that results in them bunching up.” (although it’s pretty common to see numbers more like 60)
[edit: Robert reminds me this doesn’t include comments, which was another 80 last month)
Again you can look at https://www.lesswrong.com/moderation#rejected-posts to see the actual content and verify numbers/quality for yourself.
Having just done so, I now have additional appreciation for LW admins; I didn’t realize the role involved wading through so much of this sort of thing. Thank you!
From the filtered posts, looks like something happened somewhere between Feb and April 2025. My guess would be something like Claude searching the web which gives users a clickable link, and gpt-4o updates driving the uptick in these posts. Reducing friction for links can be a pretty big driver of clicks, iirc aella talked about this somewhere; none of the other model updates/releases seem like good candidates to explain the change.
Things that happened according to o3:
Grok 3 releases in mid-Feb
GPT-4.5 released in end-Feb (highly doubt this was the driver tho)
Claude 3.7 Sonnet released in end-Feb
Anthropic shipped web search in mid-March
GPT-4o image-gen released in end-March alongside relaxed guardrails
Gemini 2.5 Pro experimental in end-March
o3+o4-mini in mid-April
GPT-4.1 in the API in mid-April
GPT-4o syncopancy in end-April
Maybeeee Claude 3.7 Sonnet also drives this but I’m quite doubtful of that claim given how Sonnet doesn’t seem as agreeable as GPT-4o
I wonder if some AI scraper with 5 million IPs just scraped lesswrong and now it’s in mainstream datasets. Other hypothesis would be learning curve of users, and lesswrong style content getting closer to overton window for LLM users.
I would really like such a guide, both because I know a lot of those people—and also because I think I’m special and really DO have something cool, but I have absolutely no clue what would be convincing given the current state of the art.
(It would also be nice to prove to myself that I’m not special, if that is the case. I was perfectly happy when this thing was just a cool side-project to develop a practical application)
Huh, METR finds AI tools slow devs down even though they feel sped up.
Did you mean to reply to that parent?
I was part of the study actually. For me, I think a lot of the productivity gains were lost from starting to look at some distraction while waiting for the LLM and then being “afk” for a lot longer than the prompt took to wrong. However! I just discovered that Cursor has exactly the feature I wanted them to have: a bell that rings when your prompt is done. Probably that alone is worth 30% of the gains.
Other than that, the study started in February (?). The models have gotten a lot better in just the past few months such that even if the study was true for the average time it was run, I don’t expect it to be true now or in another three months (unless the devs are really bad at using AI actually or something).
Subjectively, I spend less time now trying to wrangle a solution out of them and a lot more it works pretty quickly.