What I read on social media about people and these resets gives off literal worst kind of addiction vibes. I've literally seen people talking about "Oh I had an existential crisis without Fable/GPT-5.6"
These people legitimately need help, or alternatively a social life.
Maybe its different on my end because I just use a sub outside of work for fun stuff. At work its not my money so I don't really care. I go to work, maybe use these subs at home every once in a while for a fun personal project and if I hit the limits (I rarely even do) I play video games or hang out with my wife/family.
People are borderline tying their identities to these models it seems, and yet most people aren't even building anything interesting.
Yes, 100%. The author seems to be describing, quite lucidly, how he is getting sucked into a gambling-like addiction to these LLMs… although he seems unaware of the implications of that.
1. They have left over capacity that will go unused otherwise. Most users have hit their limits before the end of the week. Might as well get some use out of the servers.
2. A backend change made migrating usage tracking hard, inaccurate, or impossible.
3. A usage tracking issue meant the usage wasn’t accurate anyway and they needed a reset to save face / avoid contract breach.
The addiction or marketing strategy seems too clever. Why explain something with cleverness that can be explained by laziness or stupidity.
My impression is that, particularly on Anthropic's side, their confusing approach to quota limits is an intentional attempt to make the actual limits or offerings more opaque. This will make it easier to limit quota without officially limiting quota, because the current real-world quota is a result of bonus offerings, limited-time access, and random resets.
So it's not that "limits went down"; it's that "the bonus offer ended as planned."
In reality their approach creates only friction and frustration and forces users to other competing products(I.e Kimi K3). Once they stop subsidising the prices the users will go away to cheaper models. I could bet my own max plans.
It's a weird thing to optimize for. If you care about availability for work you can just pay more where it feels limitless and never look at the limits. I think a lot of people have a tendency to want to minmax as a "game". I also occasionally have the urge to do it but then realize it's a means to an end, not the end itself. I would feel good having maximised usage even if it wasn't productive usage which is irrational.
I think the main problem is the usage limits. You would never feel like you are saving money by watching movies on Netflix. If they didn't have limits, many might find themselves using agents less actually.
> If you care about availability for work you can just pay more where it feels limitless and never look at the limits...
Depending on your work, it doesn't feel limitless - I've just upgraded to a $200 plan and I'm starting to see where I will even hit the edges of that, especially once resets and special offers start tapering off. But with a single AI question often costing $75 in API costs, I'm not at a point where I can just switch to API and not care about cost at all.
That means I need to carefully time the work to make sure I'm squeezing what I can into the 5 hour windows. Feels a bit like the days of mainframes and time slicing, when you had to book an allotted time for your workload to run.
putting aside the possibility of very badly implemented token limiting doing requiring our accidentally just doing a reset for whatever reason
we (industry) have reached the point where questions about productivity gain, use-cost analysis and similar get asked increasingly more often
by resetting tokens they are silently decreasing the cost and increasing the use (hit limit ==usefulness stops until end of period) they are effectively fudging any "simple" internal company studies in their favor leading to a potentially pretty bad surprise if this isn't noticed by the assessing personal
An alternative is that they bought to many GPU resources in relationship to actual demand and they still need to show that they are "highly used".
Either its is appreciated for anyone benefiting from it but fishy as which company which doesn't do something fishy gives out rabbats to already paying users without turning it into a advertisement/PR benefit???
I paid for a month of Claude and paradoxically I am not using it at all because of how quickly the 5 hr limit fills. Deepseek and Mimo do not give me that "Time's running out" anxiety.
If I had to guess uncharitably, I’d say they want you to feel like you keep getting something extra for “free” while you’re getting the same service you should have expected as normal 3 months ago.
I think your hedonic treadmill's motor is close to burning out. Might be time to hop off for a bit.
edit: Ok, I have been reflecting for a bit. This article really irked me and I realize now it's because I see a lot of a former version of myself in it.
Anyhow, to the author, I implore you to take a step back and look at what your mind is doing. Humans love prediction, and we love optimization, I get it. But, you're effectively rewriting history, and holding yourself to an impossible standard of perfectly optimizing past decisions with future information. As a result, you're sucking the joy out of pursuit that you seem to love.
I know it's easier said than done to not have your mind work against you like this. But, I think you might benefit from taking a step back and asking what your actual goals are in the work you're doing, and what sufficient value to justify the $100/mo subscription looks like in terms of output or work. Playing the token consumption optimization game is a losing battle that has turned a free gift into a source of angst.
Yeah I've been noticing for a while. Definitely didn't want to be one of the first to bring it up! Few weeks ago it was really unexpected resets the day before the actually reset. Now I'm noticing Claude keeps going for a while after the 7 day quota hits 100% (I setup my statusline to include quota as I got tired of checking the app), but also it could just be bad calibration and even at 100% I still actually have quota left.
The last few have corresponded to basically within the hour or two of the release of a competing model that is getting good reviews. Twice I've gone "well, I'm over my weekly limit for the next two days, may as well check this out" and almost just as I've finished that thought, a refresh landed and I went to blitz a week of usage in 48 hours instead because once it's gone, it's gone.
Hurray market competition and what capitalism was meant to be. Go above in providing a service or the customer leaves.
I think this is because they're having trouble with capacity planning. They plan conservatively, and they have extra capacity that's going unused. But they don't want to commit to anything, in case usage spikes.
This article reads like a description of gambling-addiction behavior; the author appears to be addicted to using LLMs.
He's eagerly awaiting his next hit of dopamine from his favorite model. He's setting timers to be ready for when his next hit comes available. He's spinning up unnecessary queries just to start the timer ticking on new models.
You could basically write the same article about some guy who sits in a casino all day eagerly awaiting double-your-winnings bonuses or similar.
> He's spinning up unnecessary queries just to start the timer ticking on new models.
This one just makes sense though. If you have a body of work to do that you know will exceed your 5 hour limit, then sending a message 3 hours before you start so that the reset happens in the middle enables you to do a task in one sitting.
Yeah, I subjectively definitely feel like using LLMs has aspects of gambling. It's 100% a Skinner box where I push a button with a high chance of a substantial reward. It's fun and addictive.
But setting an automatic prompt in the morning so limits are more sensible during the day and allow continuous work is just common sense.
Gambling addition implies dopamine hits from irregular and uncertain outcomes: as I note in the post, I don't get a dopamine hit from running agents. I know people say "LLMs are just gambling because they're next-token-maximizers that can't write real code" but with GPT 5.6 Sol (and a few new tricks I discovered) the outputs are much less irregular and uncertain. It's just typical engineering.
"Not wanting to waste money" is the polar opposite of gambling.
> Gambling addition implies dopamine hits from irregular and uncertain outcomes
Your post literally describes your fascination with trying to figure out the pattern of a "random" reward that you get, and trying to maximize the value you get out of it.
I put "random" in scare quotes because I strongly believe that—just as slot machine payouts are carefully structured to keep you playing—these LLM resets are structured to keep heavy users like you coming back to max out their usage, and to progressively upgrade it.
Several other commenters have also stated this same suspicion about the pattern of resets you're describing.
> "Not wanting to waste money" is the polar opposite of gambling.
From everything I've read about gambling addiction, particularly Jay Caspian Kang, that seems wrong.
The desire to "not waste money" and "get back to even" seems like a huge part of what motivates gamblers to keep gambling.
Logically, gambling is like going to the movies. You expect to pay x currency for y value of entertainment. If y falls short of expectations you might feel like you wasted your money, but who becomes addicted to going to the movies to try to get even? There is probably someone who has, but I’ve never heard of it and it doesn’t seem to be common; not like gambling addictions. For all intents and purposes it doesn’t happen.
But gambling addictions do happen, fairly regularly. Perhaps it is loss aversion coupled with the aforementioned dopamine hit associated with gambling that makes it so prevalent?
>
Logically, gambling is like going to the movies. You expect to pay x currency for y value of entertainment.
I don't think gambling is at all like paying a set price for a ticket and having a pretty good idea of how long the entertainment will last. If "gambling" means making a series of short-term bets for entertainment value, you don't have any clear idea how long you'll be entertained for or how much it will cost.
People will show up at a casino with let's say, $200 and a debit card, and expect that they'll be able to spend $50, be entertained for 2 hours, and then just leave… while secretly hoping that they'll actually leave with more than they came with.
Then they burn through their $50 in half an hour, and dip into their remaining stash in order to keep playing. OR they triple their money in half an hour and feel such a rush that they want to keep playing with more money. Then, repeat repeat repeat.
From what I've heard from gambling addicts (thankfully this is not something I'm prone to), there is something about the actual process of gambling that makes it addictive beyond the loss/reward system.
People who are really into roulette get a buzz off seeing the wheel spin, hearing the ball bouncing etc.
Probably that comes after the initial addiction to the reward function but it then strongly reinforces it.
I don't think there's an equivalent for paying to watch a movie because the time to payoff (or not) is too long and the sensory experience is too inconsistent to elicit a conditioned place preference.
LLMs on the other hand... the time to payoff is shorter and the experience is consistent every time. It's just lacking the tactile/sensory elements
LLMs tend to be a bit random, but are still more consistent and predictable than slot-machines. Also, gambling tends to exert lower effort and higher dopamine hits than vibe-coding, making it way more addictive.
But LLMs are still addictive to some extent. Maybe its around the same level as other behavioral addictions like food, social media, or gaming addictions.
For the problems I work on with GPT 5.6 Sol and the checks and balances I have in place, I estimate:
- 80% of prompts get everything correct and are confirmed correct with manual validation
- 19% of prompts make a minor mistake based on an ambiguity of the original prompt (user error not LLM error), but then reliably fixed in a followup prompt
- 1% of prompts causes more problems than it solves and is more pragmatic to just revert
For 99% good output, there isn't much of a dopamine rush when there is good output. The dopamine rushes are for the <1% odds.
From the other replies on this post, I suspect no one believes me, but I am offering these numbers in good faith.
I think those of us who are using AI consistently believe you and understand. I'd say roughly the same thing about Claude in terms of numbers.
I think many people who don't believe you just haven't built-up the kind of prompt history & MCP / CLI tooling etc that lets you get to the point where things work at that level of accuracy.
Hope it helps to know that at least some of us here understand and are seeing the same thing. And if it's anything like my experience with Fable, "always be more ambitious". The capabilities of the models are often limited only by what you're brave enough to ask for. I keep finding I'm not ambitious enough.
Respectfully, a lot of what you're saying in this thread sounds a lot like the lies that gamblers tell themselves. Saying this as someone with a strong tendency towards addictions.
Some of these things are only possible to really see in hindsight. Yes, you've been working on these things for a while, but these systems are notably different in their capacity and strings they pull on us.
Every single prompt worked without issue, and it got most of the way on the first try with the initial prompt (+ a couple visibility bugs due to the agent not having Computer Vision to see said menu bar app) such as:
> Create a SwiftUI menu bar app named `swiftmote` using theto create the most user friendly app following Apple's HID guidelines for creating a remote that can operate a Apple TV on a local network. Instead of reimplementing the protocols needs to interface with an Apple TV, use the Python package `pyatv` and host it within the SwiftUI app as a sidecar along with a Python installation.
I have my own Apple TV I can manually verify that it worked as expected, which is notable because the agent can't test or lie about this pipeline because it does not have access to the Apple TV.
That is not hallucination or psychosis. If you want, I can release all the prompts I used. (EDIT: Sure, why not, here are the prompts. If I don't complain about something in a followup prompt, assume it worked correctly: https://gist.github.com/minimaxir/30fa820daa1392da13026ec6aa... )
A lot of this really seems to be pattern matching on superficial similarities when the meat of the issue is probably the more important one. I think for LLM usage to be compared to gambling, it has to be a) universally negative-sum in the long run, and b) extremely addictive. a) seems very unlikely to be true, even if it is sometimes negative. b) seems to be the case for some subset of programmers, but again I would not assume this just because someone uses some statements that are sometimes uttered falsely by gamblers (the siren song of those statements being that if they were true, the gambling would be a good thing. That means the truth of those statements matters and should not be assumed to be false like they are more obviously so for gambling).
That 'triggers a surge of dopamine and creates highly addictive habits' [thanks Gemini!]
LLM use for code generation does exactly that, sometimes it works amazingly, sometimes it fails inexplicably. Whether it is negative sum or not doesn't really matter. Indeed it may well prove to be negative sum, especially if we step back a bit and consider the business benefit of the code produced, not just lines of code or even features produced.
I think whether it's negative sum or not matters quite a lot, really. The variability of the outcome is a thing that might create some addictive components, but it matters a lot whether it's something with positive effects with some potential negative effects that may need to be managed, or whether it's the same as gambling, and I think it's extremely unlikely that it is the same as gambling.
I think it is because the plans are crazy expensive. when you pay 200-300$/month for an online subscription, its normal to want to extract as much value as possible.
Oh this is a good idea actually, that reset checker site allows for setting up a scheduled agent job, so that GPT alerts me when there's been a reset :)
I have a subscription. If I don't have anything to do at the moment I don't use it. I don't think I need to set a timer every five hours and come up with tasks to do just to use my subscription. Seems like odd behavior?
I also have a Netflix subscription. I watch a couple of things on it and stop. I don't think to myself I need to maximize my subscription so let me watch movies all the time and wake up at 2 AM to make sure the next movie starts.
> it's just insanely irritating to work with a tool that 1) limits its' own use 2) with a random interval.
Do you understand how the psychological response to the "random" disappearance of an annoyance is pretty much exactly the same as the psychological response to the "random" appearance of a reward?
I put "random" in square quotes because neither are in fact totally random, but both are clearly quite carefully engineered to provoke the desired response.
> you're projecting
I am not projecting. My total lifetime gambling consists of maybe 10 or 15 cash poker games with high school friends, ten minutes at a casino in Montréal which I found a revolting experience, and receiving a few $1 scratch lottery tickets as party favors.
How does that mean you're not projecting? Yes if you oversimplify an LLM's responses to be random, and you squint real hard, it kinda resembles gambling, but an essay's worth of words is more meaningful to a human reading them than a human pulling a lever on a slot machine hoping for three sets of three cherries.
Addiction is about negative impact on your private and professional life and the life of those around you and doing things about your will or things you know are bad for you.
Might be the case here, but just setting a timer and being generally hyped about something is not enough for that.
Very weird how convinced some here seem to be that addiction is involved while they apparently don't know anything about the diagnosis criteria.
Competition. If Anthropic had followed through on their plan to put Fable under API pricing, users would have jumped ship en masse to GPT 5.6 (or perhaps to K3 if it turns out to be good enough.)
My understanding is that they are no longer threatening to do it, it's part of the Max plan now (which I use.) The $20/month subscribers might still have to pay as they go, though.
I hope so! In the model selector in claude code in the browser it still says "Available until July 19" next to Fable. Also on Friday I did get an error while using Fable that said something like "This model requires API credits to use" but that disappeared after about 20 minutes....
As one of them, I'm just gonna buy another $20 subscription to OpenAI and use Sol while that's available. Why on earth would I do anything else? Fable is not magical enough to pay for credits for it with the current competition.
It feels like Anthropic was affected by resource limitations upon launch or big adoption curves in the past, and now they're tyring to be mindful of having less of those, in several ways, one of which is trying to keep capacity to keep core services running and let the high demand not impact the rest of the system.
WHen it doesn't turn out that way, it opens up options.
At the same time if other model providers were anticipating Anthropic or someone else to have problems at launch, and were waiting in the wings with their own models to launch competitively, it sets off a capacity release competition.
One might lower prices, one might give it away, or just make it available, and then there are users on other platforms with limits, or free until a certain day, etc.
They're giving everyone their next hit.
What I read on social media about people and these resets gives off literal worst kind of addiction vibes. I've literally seen people talking about "Oh I had an existential crisis without Fable/GPT-5.6"
These people legitimately need help, or alternatively a social life.
Maybe its different on my end because I just use a sub outside of work for fun stuff. At work its not my money so I don't really care. I go to work, maybe use these subs at home every once in a while for a fun personal project and if I hit the limits (I rarely even do) I play video games or hang out with my wife/family.
People are borderline tying their identities to these models it seems, and yet most people aren't even building anything interesting.
Definitely, although there is a massive sampling bias: the minority of people who are in that state are also much more likely to post about it online.
Most likely (hopefully) 99% of users aren't like that and they'll just log off if quotas get slashed
> They're giving everyone their next hit.
Yes, 100%. The author seems to be describing, quite lucidly, how he is getting sucked into a gambling-like addiction to these LLMs… although he seems unaware of the implications of that.
https://news.ycombinator.com/item?id=48961596
I can see three reasons:
1. They have left over capacity that will go unused otherwise. Most users have hit their limits before the end of the week. Might as well get some use out of the servers.
2. A backend change made migrating usage tracking hard, inaccurate, or impossible.
3. A usage tracking issue meant the usage wasn’t accurate anyway and they needed a reset to save face / avoid contract breach.
The addiction or marketing strategy seems too clever. Why explain something with cleverness that can be explained by laziness or stupidity.
My guess is they are getting folks used to the idea of:
1. The end of unlimited token subsidization and a new status quo of metered usage limits.
2. Adding to the above, self-initiated usage limit resets give you some control and again help lessen the sting of the end of subsidization.
Right now though, I think Anthropic and OpenAI are in open MAU war ahead of their respective IPOs and that might be a bigger factor.
My impression is that, particularly on Anthropic's side, their confusing approach to quota limits is an intentional attempt to make the actual limits or offerings more opaque. This will make it easier to limit quota without officially limiting quota, because the current real-world quota is a result of bonus offerings, limited-time access, and random resets.
So it's not that "limits went down"; it's that "the bonus offer ended as planned."
In reality their approach creates only friction and frustration and forces users to other competing products(I.e Kimi K3). Once they stop subsidising the prices the users will go away to cheaper models. I could bet my own max plans.
Are you saying that because they are ending subsidies, they are giving more subsidies?
It's a weird thing to optimize for. If you care about availability for work you can just pay more where it feels limitless and never look at the limits. I think a lot of people have a tendency to want to minmax as a "game". I also occasionally have the urge to do it but then realize it's a means to an end, not the end itself. I would feel good having maximised usage even if it wasn't productive usage which is irrational.
I think the main problem is the usage limits. You would never feel like you are saving money by watching movies on Netflix. If they didn't have limits, many might find themselves using agents less actually.
> If you care about availability for work you can just pay more where it feels limitless and never look at the limits...
Depending on your work, it doesn't feel limitless - I've just upgraded to a $200 plan and I'm starting to see where I will even hit the edges of that, especially once resets and special offers start tapering off. But with a single AI question often costing $75 in API costs, I'm not at a point where I can just switch to API and not care about cost at all.
That means I need to carefully time the work to make sure I'm squeezing what I can into the 5 hour windows. Feels a bit like the days of mainframes and time slicing, when you had to book an allotted time for your workload to run.
https://www.cs.cornell.edu/wya/AcademicComputing/text/earlyt...
Heroin dealer running a loss leader operation.
I'll tell you what, that crack is really moreish.
putting aside the possibility of very badly implemented token limiting doing requiring our accidentally just doing a reset for whatever reason
we (industry) have reached the point where questions about productivity gain, use-cost analysis and similar get asked increasingly more often
by resetting tokens they are silently decreasing the cost and increasing the use (hit limit ==usefulness stops until end of period) they are effectively fudging any "simple" internal company studies in their favor leading to a potentially pretty bad surprise if this isn't noticed by the assessing personal
An alternative is that they bought to many GPU resources in relationship to actual demand and they still need to show that they are "highly used".
Either its is appreciated for anyone benefiting from it but fishy as which company which doesn't do something fishy gives out rabbats to already paying users without turning it into a advertisement/PR benefit???
I paid for a month of Claude and paradoxically I am not using it at all because of how quickly the 5 hr limit fills. Deepseek and Mimo do not give me that "Time's running out" anxiety.
If I had to guess uncharitably, I’d say they want you to feel like you keep getting something extra for “free” while you’re getting the same service you should have expected as normal 3 months ago.
I think your hedonic treadmill's motor is close to burning out. Might be time to hop off for a bit.
edit: Ok, I have been reflecting for a bit. This article really irked me and I realize now it's because I see a lot of a former version of myself in it.
Anyhow, to the author, I implore you to take a step back and look at what your mind is doing. Humans love prediction, and we love optimization, I get it. But, you're effectively rewriting history, and holding yourself to an impossible standard of perfectly optimizing past decisions with future information. As a result, you're sucking the joy out of pursuit that you seem to love.
I know it's easier said than done to not have your mind work against you like this. But, I think you might benefit from taking a step back and asking what your actual goals are in the work you're doing, and what sufficient value to justify the $100/mo subscription looks like in terms of output or work. Playing the token consumption optimization game is a losing battle that has turned a free gift into a source of angst.
Yeah I've been noticing for a while. Definitely didn't want to be one of the first to bring it up! Few weeks ago it was really unexpected resets the day before the actually reset. Now I'm noticing Claude keeps going for a while after the 7 day quota hits 100% (I setup my statusline to include quota as I got tired of checking the app), but also it could just be bad calibration and even at 100% I still actually have quota left.
Read: AI companies found a way to make AI chatbot to look less expensive driving people to become more addicted to it.
This is the result of a forgotten phenomenon called "competition".
The last few have corresponded to basically within the hour or two of the release of a competing model that is getting good reviews. Twice I've gone "well, I'm over my weekly limit for the next two days, may as well check this out" and almost just as I've finished that thought, a refresh landed and I went to blitz a week of usage in 48 hours instead because once it's gone, it's gone.
Hurray market competition and what capitalism was meant to be. Go above in providing a service or the customer leaves.
Operant conditioning random reinforcement is a pretty well researched science
I think this is because they're having trouble with capacity planning. They plan conservatively, and they have extra capacity that's going unused. But they don't want to commit to anything, in case usage spikes.
This article reads like a description of gambling-addiction behavior; the author appears to be addicted to using LLMs.
He's eagerly awaiting his next hit of dopamine from his favorite model. He's setting timers to be ready for when his next hit comes available. He's spinning up unnecessary queries just to start the timer ticking on new models.
You could basically write the same article about some guy who sits in a casino all day eagerly awaiting double-your-winnings bonuses or similar.
> He's spinning up unnecessary queries just to start the timer ticking on new models.
This one just makes sense though. If you have a body of work to do that you know will exceed your 5 hour limit, then sending a message 3 hours before you start so that the reset happens in the middle enables you to do a task in one sitting.
Yeah, I subjectively definitely feel like using LLMs has aspects of gambling. It's 100% a Skinner box where I push a button with a high chance of a substantial reward. It's fun and addictive.
But setting an automatic prompt in the morning so limits are more sensible during the day and allow continuous work is just common sense.
Gambling addition implies dopamine hits from irregular and uncertain outcomes: as I note in the post, I don't get a dopamine hit from running agents. I know people say "LLMs are just gambling because they're next-token-maximizers that can't write real code" but with GPT 5.6 Sol (and a few new tricks I discovered) the outputs are much less irregular and uncertain. It's just typical engineering.
"Not wanting to waste money" is the polar opposite of gambling.
> Gambling addition implies dopamine hits from irregular and uncertain outcomes
Your post literally describes your fascination with trying to figure out the pattern of a "random" reward that you get, and trying to maximize the value you get out of it.
I put "random" in scare quotes because I strongly believe that—just as slot machine payouts are carefully structured to keep you playing—these LLM resets are structured to keep heavy users like you coming back to max out their usage, and to progressively upgrade it.
Several other commenters have also stated this same suspicion about the pattern of resets you're describing.
> "Not wanting to waste money" is the polar opposite of gambling.
From everything I've read about gambling addiction, particularly Jay Caspian Kang, that seems wrong.
The desire to "not waste money" and "get back to even" seems like a huge part of what motivates gamblers to keep gambling.
>The desire to "not waste money" and "get back to even" seems like a huge part of what motivates gamblers to keep gambling.
As someone who had family members go through gambling addiction this is the primary mechanism behind it.
Addicts don't see it as "cool fun dopamine kicks" but instead find it the only way they can get back to normal/where they are supposed to be
Don’t worry, they have a system, they can’t lose, and honestly, it’s like the outcome is almost guaranteed.
That seems like an incomplete explanation.
Logically, gambling is like going to the movies. You expect to pay x currency for y value of entertainment. If y falls short of expectations you might feel like you wasted your money, but who becomes addicted to going to the movies to try to get even? There is probably someone who has, but I’ve never heard of it and it doesn’t seem to be common; not like gambling addictions. For all intents and purposes it doesn’t happen.
But gambling addictions do happen, fairly regularly. Perhaps it is loss aversion coupled with the aforementioned dopamine hit associated with gambling that makes it so prevalent?
> Logically, gambling is like going to the movies. You expect to pay x currency for y value of entertainment.
I don't think gambling is at all like paying a set price for a ticket and having a pretty good idea of how long the entertainment will last. If "gambling" means making a series of short-term bets for entertainment value, you don't have any clear idea how long you'll be entertained for or how much it will cost.
People will show up at a casino with let's say, $200 and a debit card, and expect that they'll be able to spend $50, be entertained for 2 hours, and then just leave… while secretly hoping that they'll actually leave with more than they came with.
Then they burn through their $50 in half an hour, and dip into their remaining stash in order to keep playing. OR they triple their money in half an hour and feel such a rush that they want to keep playing with more money. Then, repeat repeat repeat.
From what I've heard from gambling addicts (thankfully this is not something I'm prone to), there is something about the actual process of gambling that makes it addictive beyond the loss/reward system.
People who are really into roulette get a buzz off seeing the wheel spin, hearing the ball bouncing etc.
Probably that comes after the initial addiction to the reward function but it then strongly reinforces it.
I don't think there's an equivalent for paying to watch a movie because the time to payoff (or not) is too long and the sensory experience is too inconsistent to elicit a conditioned place preference.
LLMs on the other hand... the time to payoff is shorter and the experience is consistent every time. It's just lacking the tactile/sensory elements
Yes I agree with you here.
LLMs tend to be a bit random, but are still more consistent and predictable than slot-machines. Also, gambling tends to exert lower effort and higher dopamine hits than vibe-coding, making it way more addictive.
But LLMs are still addictive to some extent. Maybe its around the same level as other behavioral addictions like food, social media, or gaming addictions.
> dopamine hits from irregular and uncertain outcomes
Like the LLM getting the solution right?
For the problems I work on with GPT 5.6 Sol and the checks and balances I have in place, I estimate:
- 80% of prompts get everything correct and are confirmed correct with manual validation
- 19% of prompts make a minor mistake based on an ambiguity of the original prompt (user error not LLM error), but then reliably fixed in a followup prompt
- 1% of prompts causes more problems than it solves and is more pragmatic to just revert
For 99% good output, there isn't much of a dopamine rush when there is good output. The dopamine rushes are for the <1% odds.
From the other replies on this post, I suspect no one believes me, but I am offering these numbers in good faith.
I think those of us who are using AI consistently believe you and understand. I'd say roughly the same thing about Claude in terms of numbers.
I think many people who don't believe you just haven't built-up the kind of prompt history & MCP / CLI tooling etc that lets you get to the point where things work at that level of accuracy.
Hope it helps to know that at least some of us here understand and are seeing the same thing. And if it's anything like my experience with Fable, "always be more ambitious". The capabilities of the models are often limited only by what you're brave enough to ask for. I keep finding I'm not ambitious enough.
> and a few new tricks I discovered
This right here. Any gambler would recognize that statement.
Said tricks improve the output in an objective measurable manner, not theoretical, vibes, or gambler's fallacy. (blog post forthcoming on that)
I've been researching LLM prompt optimization for longer than ChatGPT has existed; I was successfully optimizing the output of GPT-2 back in 2019.
Respectfully, a lot of what you're saying in this thread sounds a lot like the lies that gamblers tell themselves. Saying this as someone with a strong tendency towards addictions.
Some of these things are only possible to really see in hindsight. Yes, you've been working on these things for a while, but these systems are notably different in their capacity and strings they pull on us.
Be well, please.
Yesterday (unrelated to quotamaxxing described in the article), I made an Apple TV macOS menu bar remote app: https://cdn.bsky.app/img/feed_fullsize/plain/did:plc:oxaerni...
Every single prompt worked without issue, and it got most of the way on the first try with the initial prompt (+ a couple visibility bugs due to the agent not having Computer Vision to see said menu bar app) such as:
> Create a SwiftUI menu bar app named `swiftmote` using theto create the most user friendly app following Apple's HID guidelines for creating a remote that can operate a Apple TV on a local network. Instead of reimplementing the protocols needs to interface with an Apple TV, use the Python package `pyatv` and host it within the SwiftUI app as a sidecar along with a Python installation.
I have my own Apple TV I can manually verify that it worked as expected, which is notable because the agent can't test or lie about this pipeline because it does not have access to the Apple TV.
That is not hallucination or psychosis. If you want, I can release all the prompts I used. (EDIT: Sure, why not, here are the prompts. If I don't complain about something in a followup prompt, assume it worked correctly: https://gist.github.com/minimaxir/30fa820daa1392da13026ec6aa... )
A lot of this really seems to be pattern matching on superficial similarities when the meat of the issue is probably the more important one. I think for LLM usage to be compared to gambling, it has to be a) universally negative-sum in the long run, and b) extremely addictive. a) seems very unlikely to be true, even if it is sometimes negative. b) seems to be the case for some subset of programmers, but again I would not assume this just because someone uses some statements that are sometimes uttered falsely by gamblers (the siren song of those statements being that if they were true, the gambling would be a good thing. That means the truth of those statements matters and should not be assumed to be false like they are more obviously so for gambling).
the key comparison is variable reward.
That 'triggers a surge of dopamine and creates highly addictive habits' [thanks Gemini!]
LLM use for code generation does exactly that, sometimes it works amazingly, sometimes it fails inexplicably. Whether it is negative sum or not doesn't really matter. Indeed it may well prove to be negative sum, especially if we step back a bit and consider the business benefit of the code produced, not just lines of code or even features produced.
I think whether it's negative sum or not matters quite a lot, really. The variability of the outcome is a thing that might create some addictive components, but it matters a lot whether it's something with positive effects with some potential negative effects that may need to be managed, or whether it's the same as gambling, and I think it's extremely unlikely that it is the same as gambling.
I think it is because the plans are crazy expensive. when you pay 200-300$/month for an online subscription, its normal to want to extract as much value as possible.
No it’s definitely addiction
Oh this is a good idea actually, that reset checker site allows for setting up a scheduled agent job, so that GPT alerts me when there's been a reset :)
> This article reads like a description of gambling-addiction behavior; the author appears to be addicted to using LLMs.
you're projecting, it's just insanely irritating to work with a tool that 1) limits its' own use 2) with a random interval.
keeping in mind that plenty of people are making money on token use..
this guy sets a timer to wake up for work; he appears to be addicted to work.
I have a subscription. If I don't have anything to do at the moment I don't use it. I don't think I need to set a timer every five hours and come up with tasks to do just to use my subscription. Seems like odd behavior?
I also have a Netflix subscription. I watch a couple of things on it and stop. I don't think to myself I need to maximize my subscription so let me watch movies all the time and wake up at 2 AM to make sure the next movie starts.
I encourage you others to keep on minimizing so us others can keep in maximizing. Balance!
If you don’t want 1 or 2 then pay by the token. If that’s too expensive for you then now you know why 1 and 2 exist.
That's fine, but that doesn't mean that people taking steps to get the most out of what they have with 1 and 2 are now somehow gamblers.
> it's just insanely irritating to work with a tool that 1) limits its' own use 2) with a random interval.
Do you understand how the psychological response to the "random" disappearance of an annoyance is pretty much exactly the same as the psychological response to the "random" appearance of a reward?
I put "random" in square quotes because neither are in fact totally random, but both are clearly quite carefully engineered to provoke the desired response.
> you're projecting
I am not projecting. My total lifetime gambling consists of maybe 10 or 15 cash poker games with high school friends, ten minutes at a casino in Montréal which I found a revolting experience, and receiving a few $1 scratch lottery tickets as party favors.
How does that mean you're not projecting? Yes if you oversimplify an LLM's responses to be random, and you squint real hard, it kinda resembles gambling, but an essay's worth of words is more meaningful to a human reading them than a human pulling a lever on a slot machine hoping for three sets of three cherries.
Addiction is about negative impact on your private and professional life and the life of those around you and doing things about your will or things you know are bad for you.
Might be the case here, but just setting a timer and being generally hyped about something is not enough for that.
Very weird how convinced some here seem to be that addiction is involved while they apparently don't know anything about the diagnosis criteria.
Ironically my time and focus on agentic prompting and coding has helped me break an actual bad addiction (World of Warcraft)
Push down official on paper limits while limiting real world backlash by artificially inflating real world usage above paper limits with resets.
>Resets of the weekly quota for all users must be ludicrously expensive for these companies
Why can't people see the alternative hypothesis, inference has huuuuuge margins?
That explains it! I was wondering about this just today as I saw it hd reset. Alas I don't need to wait another 7 days..
Competition. If Anthropic had followed through on their plan to put Fable under API pricing, users would have jumped ship en masse to GPT 5.6 (or perhaps to K3 if it turns out to be good enough.)
Theoretically it'll still happen tomorrow. But we'll see--I think they've extended it twice?
Anthropic gave up. A $200 Max sub now gets you access to Fable 5.
Anthropic's game is over.
My understanding is that they are no longer threatening to do it, it's part of the Max plan now (which I use.) The $20/month subscribers might still have to pay as they go, though.
I hope so! In the model selector in claude code in the browser it still says "Available until July 19" next to Fable. Also on Friday I did get an error while using Fable that said something like "This model requires API credits to use" but that disappeared after about 20 minutes....
As one of them, I'm just gonna buy another $20 subscription to OpenAI and use Sol while that's available. Why on earth would I do anything else? Fable is not magical enough to pay for credits for it with the current competition.
Ive typed out several things, none of them really capture succinctly the level of disdain I have for people that this article is describing.
I cannot fathom the idea that one is warping their work on a given engineering problem around the availability of a magic next token predictor box.
I like money.
It feels like Anthropic was affected by resource limitations upon launch or big adoption curves in the past, and now they're tyring to be mindful of having less of those, in several ways, one of which is trying to keep capacity to keep core services running and let the high demand not impact the rest of the system.
WHen it doesn't turn out that way, it opens up options.
At the same time if other model providers were anticipating Anthropic or someone else to have problems at launch, and were waiting in the wings with their own models to launch competitively, it sets off a capacity release competition.
One might lower prices, one might give it away, or just make it available, and then there are users on other platforms with limits, or free until a certain day, etc.