I'll begin with noting that glyph writes prolifically. This can be a challenge for some of us with ADHD š š£. Unfairly, those of us with ADHD can write walls of text and often struggle to read a few...
I'll begin with noting that glyph writes prolifically. This can be a challenge for some of us with ADHD š š£.
Unfairly, those of us with ADHD can write walls of text and often struggle to read a few paragraphs unless highly motivated.
Also, no mud to sling here: he's always struck me as a decent person on the internet. (Please note, by way of aside: I did not use an emdash earlier and I could have. Therefore, I am not an LLM masquerading as @elight. They abhor the use of the colon).
And, so, after a point, I skimmed.
Environmental impact
Humanity has this absurd optimism that the cost can be deferred because, surely, we'll be smarter in the future and the problem will magically be solved. We focus on the benefits of technology and all too rarely consider the costs and the perverse use cases. There are too many logical fallacies along this line of thinking to list. Besides that, I'd strain to find them all if I could remember more than their descriptions...
Disinformation
I'll be brief: we are just so fucked.
Long term value (at least for technology)
As new technologies are developed (if they're developed...?), LLMs will have far fewer samples to plag^H^H^H^Hbe trained from. As new technologies are introduced, won't LLMs become akin to ivory tower academics: all knowledge with no professional judgement?
(Almost) Conclusion
So, ok, LLMs will destabilize societies (maybe already have), wreck nature, and turn us into gibbering idiots but there's some...
Optimism?
Like most technologies, whatever their questionable qualities, they can almost always also be deployed for good.
For example, I'm developing an app (this is not an advertisement) to empower neurodivergent people with a compassionate tool to help them navigate their constraints with more ease. (Ok, so it sounds like a Silicon Valley TechCrunch Disrupt "... making the world a better place" pitch).
Point is: there is potential here to provide real help to real people.
Now if only we can survive long enough as a species to help them.
P.S. As you likely can infer, I'm always a riot at parties.
I read most of the post, but as with others skimmed a few paragraphs that felt like they were repeating. I don't know Glyph, but they seem like a decent person. They approach the topic by...
I read most of the post, but as with others skimmed a few paragraphs that felt like they were repeating.
I don't know Glyph, but they seem like a decent person. They approach the topic by expressing their feelings, but clarify they're not stating an absolute truth. They acknowledge they might be wrong, or they might be right, but after two years of trying to figure it out, they're left exhausted and are ready to move on. I think that's completely fair. I like generative AI, but even I feel some exhaustion at the constant march of news, progression, and discussion about it online.
I do think they have some misunderstandings about the technology, and I'll address those specifically, even if I don't disagree with their view as a whole (or right to hold it).
First, I think they undersell local models. Local models are actually very good, and suitable for many use cases. Various quants of Gemma and Llama work great for coding and conversation. While cloud models are constantly leapfrogging each other, the release of DeepSeek's R1 was fully open-weight, and considered state-of-the-art for its time. Yes it required substantial hardware to run, but it is local, and you can run quants or distills to still get most of it in a much smaller package.
Local models are also quite easy to use, at least for techies. It's not as simple as opening a website, but tools like LM Studio provide an easy frontend for downloading and interacting with models. More on the development side, ollama (as powered by llama.cpp) also provides an easy API layer for coding tools. These are pretty close to plug and play experiences.
Models are also being quantized specifically to run on smaller devices like phones. I actually think that's very cool, and empowering for those with poor connectivity. It reminds me of when I'd stored text dumps of Wikipedia on my first gen iPod Touch, so I could look up anything on the go. Having a natural language interface though is a much more accessible approach.
A little further down, I think Glyph is also a little confused about how the energy demands are distributed. Running inference locally doesn't make any real difference in terms of energy cost. In fact, cloud providers are probably more efficient due to better token caching, and using specialized hardware (as opposed to our gaming GPUs). However, inference is not the major energy consumer in any case. What's actually expensive is the training step. This requires a significant amount of compute, and every company is running their own version of it.
In some ways, this does make it easier to excuse as a customer. Individually, nobody is using that much electricity. Outside of those long "deep reasoning" queries, most prompts can be measured in fractions of a penny. However, I don't like this argument because it ignores that we're creating a market for LLM providers as well. In the same way that consumers blame "oil companies" for creating all the carbon pollution, while also driving gas guzzling vehicles. If we create a market, we are also responsible for the energy costs of training these models.
So I do recognize the problem. At the same time, I also recognize the tradeoff. LLMs provide significant utility, at least to me, and seemingly to the one billion ChatGPT users. I want these tools to continue to get smarter, more accurate, and faster. So I guess I am the car user in this case.
My hope - and I recognize this might come off as an excuse - is that this helps further the push for alternative energy investment. Solar panels are getting extremely good, and extremely cheap, but are still not suitable for data center use. Nuclear is, and although we should have been building more nuclear for decades now, big tech companies are finally taking an interest. I hope this renewed interest helps motivate countries to do the same, and gets some economies of scale going.
Half-read ... they certainly do wander down a long list of tangents and caveats and such. And I gave up reading when I realized that, halfway thru, I still could not tell if they were equating...
Half-read ... they certainly do wander down a long list of tangents and caveats and such.
And I gave up reading when I realized that, halfway thru, I still could not tell if they were equating modern-era LLMs with genAI, or frustrated with other people who were doing it.
They might be very clever, but when I've read half your post and still do not know what your thesis is, there's a problem.
Yep, I figured that out after posting; I thought they were talking about AGI. Still not clear on their actual thesis, though, at least, not beyond "I don't approve of most genAI discussions"....
Yep, I figured that out after posting; I thought they were talking about AGI.
Still not clear on their actual thesis, though, at least, not beyond "I don't approve of most genAI discussions".
Like, why not? Halfway thru the article, and still no idea what their gripe is. I mean, I probably agree with them, but I got tired of trying to find out.
I deeply relate to the author's struggle to focus on the core of your thoughts on a broad topic (as opposed to endlessly going into secondary tangents). And I think that's actually part of the...
I deeply relate to the author's struggle to focus on the core of your thoughts on a broad topic (as opposed to endlessly going into secondary tangents). And I think that's actually part of the issue: "AI" is such a broad topic that any attempt at meaningful discourse on the subject is often derailed from the get-go because people aren't on the same page regarding what the subject even is. Are we talking about decision making algorithms in their entirety? Specifically machine learning? More specifically those relying on transformers, the architecture which was central to the currently ongoing "AI boom"? Are we even more specifically talking about the subset used for generating content, like here? Are we talking about Artificial General Intelligence? The layman often doesn't even know all of these are meaningfully different topics, and I don't think anything useful can be gained from discussing "AI" without understanding at least this.
I think this is also what's happening with the "anti-antis" Glyph mentions, which outlines a pattern similar to what I've done myself trying to talk about AI; The amount of discussion in the wake of the AI boom combined with, frankly, most people talking about it not properly understanding the topic as a whole means that quite often the issue isn't even that you disagree with someone's arguments but more that they're going in the wrong direction entirely. I personally hold a very dim view of how generative AI is currently being implemented but I cannot meaningfully express my concerns with it if the criticism starts and stops at "ChatGPT is evil incarnate, will take over the world and boil away the oceans". And within this context, I would find it pointless to talk about something like the use of generative AI as a mass-disinformation tool or any other gripe I do have when the part I actually need to clear up is "No, I don't want to start the Butlerian Jihad". Though the thought becomes more appealing by the day...
For example, I believe that the data gathering/training process for generative AI models is ethically in a dismal state and would be favorable to heavily regulating it (the uncontrolled data scrapers alone are doing measurable damage to critical aspects of the Internet's infrastructure, for one, which this very article rightfully calls out), but meaningful discussion about how we should go about it to actually improve the situation requires a general awareness of what the problem is and right now that's definitely not the case. This makes clearing up misconceptions about flawed criticism of genAI an important part of working toward a proper solution to the very real risks posed by its currently unfettered proliferation, and I don't think that should just be dismissed as "getting mad at people that do not exist" because I can guarantee you that they do exist. If we are to ever curb the abuse tech companies are perpetrating with the AI craze, public awareness of what isn't a problem to address is IMO critical to make sure any effort at proper legislation doesn't end up focusing on the wrong thing, and there is a lot of precedent regarding hamfisted digital laws around (be it simply because of cluelessness or active interference from bad actors) that backs up this specific concern.
While I've been lucky enough to mostly dodge the various ways companies keep trying to force feed genAI to me (If I ever find a valid use for feeding my data into a turbocharged autocomplete I'll do it on my own terms and on my own hardware, thank you very much), I do agree with the author that it's been getting annoying enough that I'd love nothing more than for all the genAI server farms to spontaneously melt away and for everyone to shut up about it forever. But given how relentless the push has been for adopting this technology everywhere regardless of whether it's appropriate and the damage this will keep causing, I don't think we can afford to just ignore it. The technology is already there and ultimately does have its uses, so hoping for it to fully vanish would be wishful thinking. Pushback does needs to happen (and the very vocal debate on the subject thankfully shows that it is happening), but it needs to happen in the right direction and given how many people and companies have a vested interest in preventing that, we're unfortunately not free from this mess any time soon.
I'll begin with noting that glyph writes prolifically. This can be a challenge for some of us with ADHD š š£.
Unfairly, those of us with ADHD can write walls of text and often struggle to read a few paragraphs unless highly motivated.
Also, no mud to sling here: he's always struck me as a decent person on the internet. (Please note, by way of aside: I did not use an emdash earlier and I could have. Therefore, I am not an LLM masquerading as @elight. They abhor the use of the colon).
And, so, after a point, I skimmed.
Environmental impact
Humanity has this absurd optimism that the cost can be deferred because, surely, we'll be smarter in the future and the problem will magically be solved. We focus on the benefits of technology and all too rarely consider the costs and the perverse use cases. There are too many logical fallacies along this line of thinking to list. Besides that, I'd strain to find them all if I could remember more than their descriptions...
Disinformation
I'll be brief: we are just so fucked.
Long term value (at least for technology)
As new technologies are developed (if they're developed...?), LLMs will have far fewer samples to plag^H^H^H^Hbe trained from. As new technologies are introduced, won't LLMs become akin to ivory tower academics: all knowledge with no professional judgement?
(Almost) Conclusion
So, ok, LLMs will destabilize societies (maybe already have), wreck nature, and turn us into gibbering idiots but there's some...
Optimism?
Like most technologies, whatever their questionable qualities, they can almost always also be deployed for good.
For example, I'm developing an app (this is not an advertisement) to empower neurodivergent people with a compassionate tool to help them navigate their constraints with more ease. (Ok, so it sounds like a Silicon Valley TechCrunch Disrupt "... making the world a better place" pitch).
Point is: there is potential here to provide real help to real people.
Now if only we can survive long enough as a species to help them.
P.S. As you likely can infer, I'm always a riot at parties.
"I would have written a shorter letter, but I did not have the time.", as Blaise Pascal so beautifully put it.
Sure. Except I also cannot read these failed shorter letters. š
I read most of the post, but as with others skimmed a few paragraphs that felt like they were repeating.
I don't know Glyph, but they seem like a decent person. They approach the topic by expressing their feelings, but clarify they're not stating an absolute truth. They acknowledge they might be wrong, or they might be right, but after two years of trying to figure it out, they're left exhausted and are ready to move on. I think that's completely fair. I like generative AI, but even I feel some exhaustion at the constant march of news, progression, and discussion about it online.
I do think they have some misunderstandings about the technology, and I'll address those specifically, even if I don't disagree with their view as a whole (or right to hold it).
First, I think they undersell local models. Local models are actually very good, and suitable for many use cases. Various quants of Gemma and Llama work great for coding and conversation. While cloud models are constantly leapfrogging each other, the release of DeepSeek's R1 was fully open-weight, and considered state-of-the-art for its time. Yes it required substantial hardware to run, but it is local, and you can run quants or distills to still get most of it in a much smaller package.
Local models are also quite easy to use, at least for techies. It's not as simple as opening a website, but tools like LM Studio provide an easy frontend for downloading and interacting with models. More on the development side, ollama (as powered by llama.cpp) also provides an easy API layer for coding tools. These are pretty close to plug and play experiences.
Models are also being quantized specifically to run on smaller devices like phones. I actually think that's very cool, and empowering for those with poor connectivity. It reminds me of when I'd stored text dumps of Wikipedia on my first gen iPod Touch, so I could look up anything on the go. Having a natural language interface though is a much more accessible approach.
A little further down, I think Glyph is also a little confused about how the energy demands are distributed. Running inference locally doesn't make any real difference in terms of energy cost. In fact, cloud providers are probably more efficient due to better token caching, and using specialized hardware (as opposed to our gaming GPUs). However, inference is not the major energy consumer in any case. What's actually expensive is the training step. This requires a significant amount of compute, and every company is running their own version of it.
In some ways, this does make it easier to excuse as a customer. Individually, nobody is using that much electricity. Outside of those long "deep reasoning" queries, most prompts can be measured in fractions of a penny. However, I don't like this argument because it ignores that we're creating a market for LLM providers as well. In the same way that consumers blame "oil companies" for creating all the carbon pollution, while also driving gas guzzling vehicles. If we create a market, we are also responsible for the energy costs of training these models.
So I do recognize the problem. At the same time, I also recognize the tradeoff. LLMs provide significant utility, at least to me, and seemingly to the one billion ChatGPT users. I want these tools to continue to get smarter, more accurate, and faster. So I guess I am the car user in this case.
My hope - and I recognize this might come off as an excuse - is that this helps further the push for alternative energy investment. Solar panels are getting extremely good, and extremely cheap, but are still not suitable for data center use. Nuclear is, and although we should have been building more nuclear for decades now, big tech companies are finally taking an interest. I hope this renewed interest helps motivate countries to do the same, and gets some economies of scale going.
Half-read ... they certainly do wander down a long list of tangents and caveats and such.
And I gave up reading when I realized that, halfway thru, I still could not tell if they were equating modern-era LLMs with genAI, or frustrated with other people who were doing it.
They might be very clever, but when I've read half your post and still do not know what your thesis is, there's a problem.
Pretty sure he means "generative AI" to include the whole class of contemporary AI.
Yep, I figured that out after posting; I thought they were talking about AGI.
Still not clear on their actual thesis, though, at least, not beyond "I don't approve of most genAI discussions".
Like, why not? Halfway thru the article, and still no idea what their gripe is. I mean, I probably agree with them, but I got tired of trying to find out.
I deeply relate to the author's struggle to focus on the core of your thoughts on a broad topic (as opposed to endlessly going into secondary tangents). And I think that's actually part of the issue: "AI" is such a broad topic that any attempt at meaningful discourse on the subject is often derailed from the get-go because people aren't on the same page regarding what the subject even is. Are we talking about decision making algorithms in their entirety? Specifically machine learning? More specifically those relying on transformers, the architecture which was central to the currently ongoing "AI boom"? Are we even more specifically talking about the subset used for generating content, like here? Are we talking about Artificial General Intelligence? The layman often doesn't even know all of these are meaningfully different topics, and I don't think anything useful can be gained from discussing "AI" without understanding at least this.
I think this is also what's happening with the "anti-antis" Glyph mentions, which outlines a pattern similar to what I've done myself trying to talk about AI; The amount of discussion in the wake of the AI boom combined with, frankly, most people talking about it not properly understanding the topic as a whole means that quite often the issue isn't even that you disagree with someone's arguments but more that they're going in the wrong direction entirely. I personally hold a very dim view of how generative AI is currently being implemented but I cannot meaningfully express my concerns with it if the criticism starts and stops at "ChatGPT is evil incarnate, will take over the world and boil away the oceans". And within this context, I would find it pointless to talk about something like the use of generative AI as a mass-disinformation tool or any other gripe I do have when the part I actually need to clear up is "No, I don't want to start the Butlerian Jihad".
Though the thought becomes more appealing by the day...For example, I believe that the data gathering/training process for generative AI models is ethically in a dismal state and would be favorable to heavily regulating it (the uncontrolled data scrapers alone are doing measurable damage to critical aspects of the Internet's infrastructure, for one, which this very article rightfully calls out), but meaningful discussion about how we should go about it to actually improve the situation requires a general awareness of what the problem is and right now that's definitely not the case. This makes clearing up misconceptions about flawed criticism of genAI an important part of working toward a proper solution to the very real risks posed by its currently unfettered proliferation, and I don't think that should just be dismissed as "getting mad at people that do not exist" because I can guarantee you that they do exist. If we are to ever curb the abuse tech companies are perpetrating with the AI craze, public awareness of what isn't a problem to address is IMO critical to make sure any effort at proper legislation doesn't end up focusing on the wrong thing, and there is a lot of precedent regarding hamfisted digital laws around (be it simply because of cluelessness or active interference from bad actors) that backs up this specific concern.
While I've been lucky enough to mostly dodge the various ways companies keep trying to force feed genAI to me (If I ever find a valid use for feeding my data into a turbocharged autocomplete I'll do it on my own terms and on my own hardware, thank you very much), I do agree with the author that it's been getting annoying enough that I'd love nothing more than for all the genAI server farms to spontaneously melt away and for everyone to shut up about it forever. But given how relentless the push has been for adopting this technology everywhere regardless of whether it's appropriate and the damage this will keep causing, I don't think we can afford to just ignore it. The technology is already there and ultimately does have its uses, so hoping for it to fully vanish would be wishful thinking. Pushback does needs to happen (and the very vocal debate on the subject thankfully shows that it is happening), but it needs to happen in the right direction and given how many people and companies have a vested interest in preventing that, we're unfortunately not free from this mess any time soon.