we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
-
@davidgerard cheap Chinese replacement AI will probably be a thing. Not sure if that's the stupidest and shortest-term thing to do, but if you can't afford /access the AI you built something with, you need a substitute.
@audunmb currently it's still heavily subsidised, and training is expensive so the local models will only go out of date.
I originally thought local LLMs for coding would be what LLMs were left with after the bubble pop, but @tante pointed out they'll all quickly go out of date.
We already see current maintained chatbot models have a habit of putting in unfavoured, obsolete and even deprecated code constructs cos that's overwhelmingly what the model was trained on.
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
@davidgerard local models. It's already where LLMs make "the most sense", and I'm sure there will be a wave of hype once "Gen AI is around the corner" stops convincing investors. It'll be "streamlined Lean AI that knows Your Business and doesn't talk about goblins!".
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
@davidgerard Consulting firms will step up and offer "AI Abatement" solutions. Companies will pay $1000/hour to Accenture for a fresh grad making minimum wage. The code will end up worse.
-
@audunmb currently it's still heavily subsidised, and training is expensive so the local models will only go out of date.
I originally thought local LLMs for coding would be what LLMs were left with after the bubble pop, but @tante pointed out they'll all quickly go out of date.
We already see current maintained chatbot models have a habit of putting in unfavoured, obsolete and even deprecated code constructs cos that's overwhelmingly what the model was trained on.
@davidgerard @audunmb @tante I wonder if there are any programming languages that are so obscure the ChatterBots don't know anything about them? *Breaks out the SNOBOL manual.. *
-
@audunmb currently it's still heavily subsidised, and training is expensive so the local models will only go out of date.
I originally thought local LLMs for coding would be what LLMs were left with after the bubble pop, but @tante pointed out they'll all quickly go out of date.
We already see current maintained chatbot models have a habit of putting in unfavoured, obsolete and even deprecated code constructs cos that's overwhelmingly what the model was trained on.
@davidgerard @audunmb @tante And since the available sources for training an LLM will contain more and more of their own productions, they'll soon be entering the signal degradation phase when then more recent the data an LLM is trained on, the worse its output will be.
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
@davidgerard Same thing they do when they need to clean up anything – hire apprentices and juniors and let them bodge everything together.
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
@davidgerard I have been telling folks that my naively optimistic take is the tech displaced will go back to work cleaning up the aftermath of AI slop. I think a massive tech collapse is more likely.
-
@davidgerard @audunmb @tante And since the available sources for training an LLM will contain more and more of their own productions, they'll soon be entering the signal degradation phase when then more recent the data an LLM is trained on, the worse its output will be.
@aaribaud @davidgerard @audunmb @tante : « eating your own shit is never healthy »
https://ploum.net/2022-12-05-drowning-in-ai-generated-garbage.html (written in 2022)
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
@davidgerard My pessimistic view is that that there won't be anything they're able to do except offer platitudes and excuses. Hiring takes time, building a knowledge base takes time, etc.
It will be yet another example of our systems slowly breaking now and not working consistently. Well accept that sometimes things just don't work anymore, and we'll all sagely nod our heads and remember the times when we thought about trite things like longevity and reliability. -
@davidgerard local models. It's already where LLMs make "the most sense", and I'm sure there will be a wave of hype once "Gen AI is around the corner" stops convincing investors. It'll be "streamlined Lean AI that knows Your Business and doesn't talk about goblins!".
@btuftin not sure the up to date training will happen https://circumstances.run/@davidgerard/116533291722823630
maybe we'll rebuild tech on python 2.7
-
@davidgerard cheap Chinese replacement AI will probably be a thing. Not sure if that's the stupidest and shortest-term thing to do, but if you can't afford /access the AI you built something with, you need a substitute.
@audunmb @davidgerard My understanding is probably crap, but don't coding models need to be not heavily quantised? Essentially, fairly memory heavy, while the cost to run is hardware plus electricity? So, unless china is able to innovate much harder, or get much cheaper electricity, I'm not sure how they'd cost less - wage arbitrage doesn't really factor into this one...
-
@aaribaud @davidgerard @audunmb @tante : « eating your own shit is never healthy »
https://ploum.net/2022-12-05-drowning-in-ai-generated-garbage.html (written in 2022)
@ploum @aaribaud @audunmb @tante see yesterday's Pivot, model collapse seems to be visibly hitting ChatGPT https://pivot-to-ai.com/2026/05/06/openai-chatgpt-goes-goblin-mode-let-none-say-model-collapse/
-
@davidgerard cheap Chinese replacement AI will probably be a thing. Not sure if that's the stupidest and shortest-term thing to do, but if you can't afford /access the AI you built something with, you need a substitute.
@audunmb @davidgerard Nope. I mean, people will use Deepseek and Qwen and stuff, but they already are.
The future is a co. fine tuning a small-ish CodeGemma or such on their codebase, and making that available to employees.
It's the "good enough" policy, followed by a "no, your request for an expensive Claude subscription is denied, it's too expensive and the internal model we've made available is good enough, now get out of my office."
That's the future, and should've been the present.
-
@audunmb @davidgerard My understanding is probably crap, but don't coding models need to be not heavily quantised? Essentially, fairly memory heavy, while the cost to run is hardware plus electricity? So, unless china is able to innovate much harder, or get much cheaper electricity, I'm not sure how they'd cost less - wage arbitrage doesn't really factor into this one...
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
Umm Anyone checked if any LLM Model was trained in FreePascal?
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
@davidgerard a combination of outsourcing, underpaid internships, and just pulling more work on remaining staff.
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
When the bubble pops, everyone will be left too poor to drive any rehiring of subject matter experts. The job losses are already locked in.
-
@btuftin not sure the up to date training will happen https://circumstances.run/@davidgerard/116533291722823630
maybe we'll rebuild tech on python 2.7
@davidgerard @btuftin python 3 is close to 20 years old, and it's been usable for 15, most python code written in the last 10 years, while python was serioulsy picking up, has been 3.6+, llms are very much up to date on *that*, and for js you say, well, i'm told the react they write is not the latest and greatest, but it's better than what i'd do.
I predict either a slowdown on adoption of new libraries/language features, or a move on from the current ai tech, but probably not a turn around.
-
we joke that when the AI bubble pops and the managers can't afford the chatbot any more, the surviving companies will hire the people who know how shit works to clean up
but this is of course optimistic. observed behaviour is that they will instead do the stupidest and shortest-term thing they can do instead of ever doing it properly.
so what do you envision this might be?
for clarity, i think when the AI bubble pops, which I place as some time next year at the latest - and you can hear the screeching noises in 2026 - the current recession signs will turn into a full Great Depression 2, so those surviving companies will also be doing not so great
@davidgerard "the surviving companies will hire the people who know how shit works to clean up", anyone thinking that clearly hasn't worked in tech. I've been in this shit for over 30 years, no one gets paid to clear up anything.
It's rare for a company with paying customers to care if their product is rubbish. This becomes truer the more the customer pays. (and prices charged are rarely proportional to the quality of the product. -
Umm Anyone checked if any LLM Model was trained in FreePascal?
@DBG3D @davidgerard i tried to ask it something simple yet weird enough it couldn't just regurgitate from examples, i can't tell if it's good or not, but feel free to have a look https://chatgpt.com/share/69fc916c-aee4-8333-9534-7f76f1a78687
