We tend to think that markets reflect consumer choice. But the artificial-intelligence boom has been powered by the investment decisions of concentrated capital rather than by consumer demand.
In order for AI to have a market there would need to be a large scale demonstration of the uses and benefits of AI in many various industries, most of which do not involve image generation or writing a single line of computer code.
I’m so sick of asking people for demonstrations of AI use and it being 100% programmers going “oh it’s really good at writing functions”, now I have no idea if it is, or isn’t, any good at writing functions, but since I don’t write functions, that example is of no use to me.
I want to know how good it is at writing technical documentation for proprietary hardware solutions that the AI will have no experience of in its training set, because I suspect it sucks at it. How good is it at taking poorly written guides in Chinese, translated into English via Klingon and ancient Babylonian, how does it handle it if it comes across the sentence “absolutely do not ever warning” with no additional information as to what I absolutely should never not do. Because that’s when thinking is involved, and AI doesn’t think.
Basically the general AI chat interaction gives a pretty good indication of capabilities and limitations as it pertains to various applications. It sounds right, but frequently is not right, and that has predictable implications.
Particularly it is really good at generating credible sounding content. It does not have a model for ‘facts’ so correctness is a persistent issue. For coding, the strategies involve having some sort of absolute validation to say whether it is “correct” or not, whether that’s some human reading it, or in the die-hard scenarios, someone writing a whole system to let it try, get pass/fail, try again on fail, and spin until it is done or you stop it because it just spins and spins. The former can be selectively useful, the latter I find to be more obnoxious than just doing the work.
So broadly, scenarios in which the “truth” is objective and/or in the input directly is where it can provide useful results. The examples of translation are an area where it is more helpful, since it isn’t tasked with adding any ‘facts’ merely reworking the input into another language. “Absolute do not ever warning” type stuff is the result of the human trying to write in non-native, so they would feed their native work into LLM and probably get better result.
The other super depressing thing is slop fiction and deep fakes. Here “facts” don’t matter, as long as it seems consistent. Even as it adds elements that may be ‘wrong’, well, it’s fiction so wrong doesn’t matter as much so long as it seems narratively consistent. Enabling people lacking in drive and creativity to pump out as much volume as people with a lot of drive and creativity with more interesting things to say.
Of course, the bubble is predicated on it working to replace most white collar work, which it is not poised to do. There’s a lot it can do, but it’s penchant for generating incorrect output is a liability that can’t be afforded in a lot of contexts. A lot of places just put the onus on the victim of the AI support bot by saying up front “you can’t trust anything this bot is going to say, but you are going to be getting the bot anyway”. Even as human support might be asked to intervene, the human is still likely to be able to review the chat quickly and maybe take a minute to see what the user wanted, click some button in the system the customer needed clicked, and move on, instead of 10 minutes of back and forth to figure out what the user needs. So they could still lay off 75% of those support people and, thanks to AI, still maintain apparent responsiveness, even if the users are annoyed. Lots of orgs have shifted to have sales be human because a customer gets pissed, but support from AI chatbot, because, well, you got their money, screw em, they can deal with some annoyance.
It’s quite terrible at writing technical documentation of even systems of pure code it’s coded itself , it’s been all down hill since opus 4.6 for writing in any kind of coherent English to convey information .
The thing is so full of slogans, irrelevant counting , and meta commentary announcing what it’s going to say moronically, and weird AF corporate sleazeball speak, now it can rattle on for pages saying mere sentences worth of information .
OK, here’s a real life use case, small time as it is, that has nothing to do with anything rarefied.
I have a TCL Android TV that, after some “unapproved measures”, runs without a Google account - FLauncher as the front end, telemetry blocked, home media streaming, YouTube ad-free. The yar har har works.
Two days ago the kids come in. TV’s not working.
Jellyfin: fine. Nova Player: fine. Stremio, YouTube, anything hitting the net: dead.
We’d had a brownout that day, so I figured the router was scrambled. Reset it, re-added the network… and Smarttube now launches, plays for 30 seconds, then hard-locks. Just weird JSON vomit and then dead.
At that point I’m thinking the TV is fucked. Options:
factory reset and reinstall the workarounds - 2-3hrs of pain-in-the-ass fuckery, but it wouldn’t be the first time.
crawl the net decoding cryptic error codes.
take it in for repair (good luck with that).
fall down a TV diagnostics rabbit hole.
Instead, I chose option 5)
I told my local LLM: "This TV is on this IP. This laptop has ADB access. I’ve given you MCP tools with ADB libraries. Here’s the issue, here’s what I tried. Check prior history of what I did for this project, figure out what’s wrong now and fix it, or tell me it’s broken.
If it looks like the brownout caused corruption, factory reset, stream-install the backup APKs directly, sign in my accounts, get it back up and running.
Failing that, if it’s looking like a hardware issue, push a diagnostic tool, test the TV and tell me which parts I might need, as well as likely costs using eBay or Amazon. I’m happy to spend $50 on this shitbox and not one cent more. If the fix looks like it’s more than $50, find me a TV of this size under $200. Prefer local shop X or Y.
Present everything as markdown in the usual folder, using house style. Ping me on my phone if you need me to authorise anything or when you’re done with the TL;DR."
The agent ADBs in, pokes around, tests, checks the WiFi, and figures out the brownout had corrupted the date and time.
Turns out this p.o.s doesn’t have a RTC back up battery.
Wrong clock = certs flagged as invalid = no net.
It syncs the date and time with my location. Boom, everything comes back.
Now, could I have twigged it myself?
Yes - eventually. But the agent already had the tools, knew the TV’s prior state, and had a record of what I’d already tried.
It could’ve reset the TV, sideloaded all my apps, re-signed into my accounts - the whole lot - if needed.
This is not the only time I’ve had real world use of my local, and it had very little to do with “writing functions”.
I had a project - jailbreak my tcl. Over time, I’d stored enough context of what I had done in the LLM and its tools, so that it could do what I’d do (and more), faster, based on the constraints I had put in place.
People that call these things “spicy autocorrect” are demonstrating either a skill issue or a lack of imagination.
I get doing things because you can, that doesn’t mean it’s not wholly unsuited to the task and massive over kill that you only use because either 1. It’s self hosted so you are already paying for it or 2. It’s super cheap. You likely wouldn’t entertain such a set up if you were paying the true price for training and running the models.
Given the surplus of power from solar I have, my set up (discussed elsewhere), I can run the training process end to end over about 50 days. It could be fun to make something truly bespoke from scratch.
Returning to my prior example - it was only overkill for this task (and only because the problem turned out to be relatively simple), but I wanted to answer against the claim that people were using LLMs and agents for just programming.
The example I cited IS silly, but it happened - IRL - 3 days ago. I can cite several much more “serious” use cases if you’d like.
On the broader topic of costs: none of us can really comprehend the embedded costs in the things we use everyday.
However unlike (say) the recurring embedded costs of buying bottled water, a cup of coffee or streaming a movie, the costs of my set up are paid once and pay off multiple times.
Ultimately, I get why there’s a FuckAI sentiment… but really we should expand that to “Fuck Corpo AI”. My feeling is that it’s not AI that people are against per se, it’s the corpos and what they’re doing with them. Too much of that gets flattened and tribal.
The “true cost” argument only holds if you assume the only valid infrastructure is someone else’s. I don’t make that assumption, and I’d say that neither do others in the enthusiast community.
In order for AI to have a market there would need to be a large scale demonstration of the uses and benefits of AI in many various industries, most of which do not involve image generation or writing a single line of computer code.
I’m so sick of asking people for demonstrations of AI use and it being 100% programmers going “oh it’s really good at writing functions”, now I have no idea if it is, or isn’t, any good at writing functions, but since I don’t write functions, that example is of no use to me.
I want to know how good it is at writing technical documentation for proprietary hardware solutions that the AI will have no experience of in its training set, because I suspect it sucks at it. How good is it at taking poorly written guides in Chinese, translated into English via Klingon and ancient Babylonian, how does it handle it if it comes across the sentence “absolutely do not ever warning” with no additional information as to what I absolutely should never not do. Because that’s when thinking is involved, and AI doesn’t think.
Basically the general AI chat interaction gives a pretty good indication of capabilities and limitations as it pertains to various applications. It sounds right, but frequently is not right, and that has predictable implications.
Particularly it is really good at generating credible sounding content. It does not have a model for ‘facts’ so correctness is a persistent issue. For coding, the strategies involve having some sort of absolute validation to say whether it is “correct” or not, whether that’s some human reading it, or in the die-hard scenarios, someone writing a whole system to let it try, get pass/fail, try again on fail, and spin until it is done or you stop it because it just spins and spins. The former can be selectively useful, the latter I find to be more obnoxious than just doing the work.
So broadly, scenarios in which the “truth” is objective and/or in the input directly is where it can provide useful results. The examples of translation are an area where it is more helpful, since it isn’t tasked with adding any ‘facts’ merely reworking the input into another language. “Absolute do not ever warning” type stuff is the result of the human trying to write in non-native, so they would feed their native work into LLM and probably get better result.
The other super depressing thing is slop fiction and deep fakes. Here “facts” don’t matter, as long as it seems consistent. Even as it adds elements that may be ‘wrong’, well, it’s fiction so wrong doesn’t matter as much so long as it seems narratively consistent. Enabling people lacking in drive and creativity to pump out as much volume as people with a lot of drive and creativity with more interesting things to say.
Of course, the bubble is predicated on it working to replace most white collar work, which it is not poised to do. There’s a lot it can do, but it’s penchant for generating incorrect output is a liability that can’t be afforded in a lot of contexts. A lot of places just put the onus on the victim of the AI support bot by saying up front “you can’t trust anything this bot is going to say, but you are going to be getting the bot anyway”. Even as human support might be asked to intervene, the human is still likely to be able to review the chat quickly and maybe take a minute to see what the user wanted, click some button in the system the customer needed clicked, and move on, instead of 10 minutes of back and forth to figure out what the user needs. So they could still lay off 75% of those support people and, thanks to AI, still maintain apparent responsiveness, even if the users are annoyed. Lots of orgs have shifted to have sales be human because a customer gets pissed, but support from AI chatbot, because, well, you got their money, screw em, they can deal with some annoyance.
It’s quite terrible at writing technical documentation of even systems of pure code it’s coded itself , it’s been all down hill since opus 4.6 for writing in any kind of coherent English to convey information .
The thing is so full of slogans, irrelevant counting , and meta commentary announcing what it’s going to say moronically, and weird AF corporate sleazeball speak, now it can rattle on for pages saying mere sentences worth of information .
OK, here’s a real life use case, small time as it is, that has nothing to do with anything rarefied.
I have a TCL Android TV that, after some “unapproved measures”, runs without a Google account - FLauncher as the front end, telemetry blocked, home media streaming, YouTube ad-free. The yar har har works.
Two days ago the kids come in. TV’s not working.
Jellyfin: fine. Nova Player: fine. Stremio, YouTube, anything hitting the net: dead.
We’d had a brownout that day, so I figured the router was scrambled. Reset it, re-added the network… and Smarttube now launches, plays for 30 seconds, then hard-locks. Just weird JSON vomit and then dead.
At that point I’m thinking the TV is fucked. Options:
factory reset and reinstall the workarounds - 2-3hrs of pain-in-the-ass fuckery, but it wouldn’t be the first time.
crawl the net decoding cryptic error codes.
take it in for repair (good luck with that).
fall down a TV diagnostics rabbit hole.
Instead, I chose option 5)
I told my local LLM: "This TV is on this IP. This laptop has ADB access. I’ve given you MCP tools with ADB libraries. Here’s the issue, here’s what I tried. Check prior history of what I did for this project, figure out what’s wrong now and fix it, or tell me it’s broken.
If it looks like the brownout caused corruption, factory reset, stream-install the backup APKs directly, sign in my accounts, get it back up and running.
Failing that, if it’s looking like a hardware issue, push a diagnostic tool, test the TV and tell me which parts I might need, as well as likely costs using eBay or Amazon. I’m happy to spend $50 on this shitbox and not one cent more. If the fix looks like it’s more than $50, find me a TV of this size under $200. Prefer local shop X or Y.
Present everything as markdown in the usual folder, using house style. Ping me on my phone if you need me to authorise anything or when you’re done with the TL;DR."
The agent ADBs in, pokes around, tests, checks the WiFi, and figures out the brownout had corrupted the date and time.
Turns out this p.o.s doesn’t have a RTC back up battery.
Wrong clock = certs flagged as invalid = no net.
It syncs the date and time with my location. Boom, everything comes back.
Now, could I have twigged it myself?
Yes - eventually. But the agent already had the tools, knew the TV’s prior state, and had a record of what I’d already tried.
It could’ve reset the TV, sideloaded all my apps, re-signed into my accounts - the whole lot - if needed.
This is not the only time I’ve had real world use of my local, and it had very little to do with “writing functions”.
I had a project - jailbreak my tcl. Over time, I’d stored enough context of what I had done in the LLM and its tools, so that it could do what I’d do (and more), faster, based on the constraints I had put in place.
People that call these things “spicy autocorrect” are demonstrating either a skill issue or a lack of imagination.
You managed to set up that tech stack but needed an agent with unrestricted access to figure out the time was wrong 😂
There are skill issues and a lack of imagination here but it’s not from people dismissing the use of LLM…
Using a sledgehammer to crack a wallnut can be satisfying :)
Besides, the kids are always finding fun and inventive ways to break the TV, so I figured set it up once and then let the bot handle it.
Twelfth Doctor: This is Clara, not my assistant. She’s a… Some other word.
Clara: I’m his carer.
Twelfth Doctor: Yeah, my carer. She cares so I don’t have to.
I get doing things because you can, that doesn’t mean it’s not wholly unsuited to the task and massive over kill that you only use because either 1. It’s self hosted so you are already paying for it or 2. It’s super cheap. You likely wouldn’t entertain such a set up if you were paying the true price for training and running the models.
Actually, I’ve been toying around with doing just that.
https://github.com/karpathy/nanochat
Given the surplus of power from solar I have, my set up (discussed elsewhere), I can run the training process end to end over about 50 days. It could be fun to make something truly bespoke from scratch.
Returning to my prior example - it was only overkill for this task (and only because the problem turned out to be relatively simple), but I wanted to answer against the claim that people were using LLMs and agents for just programming.
The example I cited IS silly, but it happened - IRL - 3 days ago. I can cite several much more “serious” use cases if you’d like.
On the broader topic of costs: none of us can really comprehend the embedded costs in the things we use everyday.
However unlike (say) the recurring embedded costs of buying bottled water, a cup of coffee or streaming a movie, the costs of my set up are paid once and pay off multiple times.
Ultimately, I get why there’s a FuckAI sentiment… but really we should expand that to “Fuck Corpo AI”. My feeling is that it’s not AI that people are against per se, it’s the corpos and what they’re doing with them. Too much of that gets flattened and tribal.
The “true cost” argument only holds if you assume the only valid infrastructure is someone else’s. I don’t make that assumption, and I’d say that neither do others in the enthusiast community.