I think this is sort of the core of all of the arguments currently, at least the ones that are broadly reasoned out and don't come from a "FUCK AI!" or "LOVE AI!" standpoint. And even though the...
"I know that it might feel like refusing to vote for any AI-made game is a mistake. But it’s all mistakes. I just want us to make the right ones."
I think this is sort of the core of all of the arguments currently, at least the ones that are broadly reasoned out and don't come from a "FUCK AI!" or "LOVE AI!" standpoint. And even though the conceit of this post is that Walker does take a hardline no-AI stance in his voting behaviour, I think I can respect this line of reasoning much more than I'd initially expect to.
While I do share the concerns listed in the IGF statement, I am perhaps a little less fatalistic about them? No, I broadly do not think that AI usage is "causing cognitive decline" on the whole, in the same way that I don't think the concrete saw causes "lower average limb count per capita", even though it would be perfectly technically correct to make that assessment from a data driven standpoint, but I will agree that I don't think "AI is here to stay", even if (here comes my favourite phrase recently) mostly for practical reasons. I can't say that it is. I don't know what the future holds, and frankly, I wouldn't mind if it didn't.
And frankly, I say this as someone who does think AI has helped them personally, and would have little reservation about letting it do code (code, only code, nothing but the code) if I were to develop a video game. If AI disappeared tomorrow, I think I'd still have learned enough to reasonably start doing this myself, by hand, like in the olden days. It's certainly possible. I do not design systems smarter than me, which is just a good rule of thumb, no matter what tooling you're using.
What I will say is that this position is a rare one, mainly because it does require engaging with the topic, whether you want to or not. I can't blame you if you don't want to, it's a messy topic, and in the same way I don't want to discuss veganism it's perfectly okay to not want to discuss AI and sticking with a basic, maybe a little polemic, opinion of "FUCK ALL AI". Or "LOVE AI FOREVER". And even if I still hold a (marginally) different position to Walker, I'm really glad there's some nuanced thinking on the issue from a seasoned professional.
That’s fascinating. Will programmers somehow stop using it en masse? Why will they do that? Or what do you mean by AI no longer staying? To give you some idea of where I’m coming from, imagine...
I will agree that I don't think "AI is here to stay”
That’s fascinating. Will programmers somehow stop using it en masse? Why will they do that? Or what do you mean by AI no longer staying?
To give you some idea of where I’m coming from, imagine OpenAI and Anthropic go bankrupt tomorrow and their services shut entirely down. Deepseek weights are open source and virtually as capable, but I can run them on my desktop. There’s no way to put this genie back in the bottle, is there?
Mainly, I'm hedging my bets. Should have clarified, I don't think we can, with the sufficient conviction, assert that AI is here to stay. We simply don't know. I'll grant that it probably is, but...
Mainly, I'm hedging my bets.
Should have clarified, I don't think we can, with the sufficient conviction, assert that AI is here to stay. We simply don't know. I'll grant that it probably is, but the economy might change again. What if inference does, against all odds, become too expensive? What if the Pareto curve becomes a Pareto line and we just don't see improvements? What if this all just sorta... dies, like NFTs did?
Now again, I'm not saying that I believe that, I personally think that the chance of "AI staying" is probably higher than the inverse, but I also bet on no more on-the-ground military conflicts in Europe, and a few days after I made that statement to a couple of friends, Russia invaded Ukraine.
Forgive me if I've misunderstood, but it seems largely irrelevant if improvement stops. If AI capability stopped improving right here right now forever, we'd still be left with some incredibly...
What if the Pareto curve becomes a Pareto line and we just don't see improvements?
Forgive me if I've misunderstood, but it seems largely irrelevant if improvement stops. If AI capability stopped improving right here right now forever, we'd still be left with some incredibly useful tools that are adaptable to a huge range of situations. I respect the viewpoint of "we cannot know anything for certain", but I also struggle to find a theoretical reason that would cause mass desertion of AI like we saw with NFTs.
(I'm not unkz, just dropping in) If I may lovingly place a single word into unkz' mouth, I think that their point is that local inference -- or even just inference sat in someone else's data...
(I'm not unkz, just dropping in)
What if inference does, against all odds, become too expensive?
If I may lovingly place a single word into unkz' mouth, I think that their point is that local inference -- or even just inference sat in someone else's data centre, at the pre-hype rates -- is already feasible, so it'd take like a bacteria that eats silicon wafers to make it stop. Apple's apparently shipping an app to run local models, too, so I can't see how things won't only get more stuck in from here.
There's some wild stuff happening with inference engines now where people are reporting 80 tok/s with Deepseek V4 (Flash?) on a CPU. I honestly don't know if I believe that. But the Qwen model I...
There's some wild stuff happening with inference engines now where people are reporting 80 tok/s with Deepseek V4 (Flash?) on a CPU. I honestly don't know if I believe that. But the Qwen model I host gets a few % faster every week with llama.cpp updates. There's a MacOS inference engine that's highly optimized for Apple Silicon combined with Qwen models. I just ordered a new M5 Max Macbook and am looking forward to running models that previously would have been impossible to run on any laptop. Local inference is the future.
Zoinks, that's, uh. Yeah kinda wild. MoE models mitigate how awful local hardware is w.r.t. memory bandwidth, but that's probably like a 1.375 bpw model. Or else one of those crazy sixteen slot...
80 tok/s with Deepseek V4 (Flash?) on a CPU.
Zoinks, that's, uh. Yeah kinda wild. MoE models mitigate how awful local hardware is w.r.t. memory bandwidth, but that's probably like a 1.375 bpw model. Or else one of those crazy sixteen slot Xeons.
But yeah, things will only get faster. There are hard upper limits due to memory bandwidth, as noted, but better draft models, improved quants (e.g. trellis, autoround, etc.), and -- someday -- maybe a change in architecture (e.g. diffusion models) should push local hardware even further.
Local inference is the future.
Fingers crossed! That'll mean this is a bubble and I get to watch OpenAI/Anthropic implode in a fiery economic fireball (admittedly, this will singe everyone else too, but I really want to live to see them go down).
(edit) there's promising stuff coming for CAD modelling, and hopefully robotics will follow that. Fingers crossed, but that stuff is far more theoretical at this point than I'd be willing to bet on (although I'm hoping to get directly involved in it ...)
The Apple Silicon stuff seems to rely on token diffusion. There's all sorts of architectural changes like Qwen 3.8 Flash Next's embeddings table which can be offloaded to disk to save you 10-20...
The Apple Silicon stuff seems to rely on token diffusion. There's all sorts of architectural changes like Qwen 3.8 Flash Next's embeddings table which can be offloaded to disk to save you 10-20 GiB of VRAM.
But it's already profitable to sell people tokens. Hundreds of companies are doing that right now. Look at openrouter and see how many providers there are. Only a small handful have hundreds of...
What if inference does, against all odds, become too expensive?
But it's already profitable to sell people tokens. Hundreds of companies are doing that right now. Look at openrouter and see how many providers there are. Only a small handful have hundreds of billions in VC funding with which they can serve trillions of tokens at a loss.
What if the Pareto curve becomes a Pareto line and we just don't see improvements?
If the models stop improving right now then we're still left with AI models that make most traditional graphic design work not worth the cost. We're still left with LLMs that can write pretty decent code for a variety of use cases.
Additionally, I don't think it makes sense to vote against AI submissions for environmental reasons. We've never seen judges vote against contestants that flew in internationally to appear in a...
I broadly do not think that AI usage is "causing cognitive decline" on the whole, in the same way that I don't think the concrete saw causes "lower average limb count per capita", even though it would be perfectly technically correct to make that assessment from a data driven standpoint
Additionally, I don't think it makes sense to vote against AI submissions for environmental reasons. We've never seen judges vote against contestants that flew in internationally to appear in a competition. Or drove a truck 800 miles to get there.
AI is absolutely here to stay. People need to re-enter reality and realize people will almost always opt for the lowest effort option available. I'm not here to fight on behalf of AI companies (cue "leave the billion dollar corporation alone" meme). I don't need to do anything to end up being right.
I certainly hope so. I think we're starting to see cracks in it. Maybe this year, at the New Year's gathering, someone else will bring up AI, and I can see what other peoples' reactions are. 'Cuz...
People need to re-enter reality and realize people will almost always opt for the lowest effort option available.
I certainly hope so. I think we're starting to see cracks in it. Maybe this year, at the New Year's gathering, someone else will bring up AI, and I can see what other peoples' reactions are. 'Cuz last year it was rough. Hopefully people understand that more than just scammers and ex-tech bros can use these tools, though, and that genuinely good things can come from them as well.
As someone that has been mentally on-the-fence a lot about varying levels of AI usage, and trying to reckon with the different implications, I really enjoyed Austin's take here. As usual with him,...
As someone that has been mentally on-the-fence a lot about varying levels of AI usage, and trying to reckon with the different implications, I really enjoyed Austin's take here.
As usual with him, it's considered, well-stated, and thoughtful. I thought others on this site might enjoy reading it.
I think this is sort of the core of all of the arguments currently, at least the ones that are broadly reasoned out and don't come from a "FUCK AI!" or "LOVE AI!" standpoint. And even though the conceit of this post is that Walker does take a hardline no-AI stance in his voting behaviour, I think I can respect this line of reasoning much more than I'd initially expect to.
While I do share the concerns listed in the IGF statement, I am perhaps a little less fatalistic about them? No, I broadly do not think that AI usage is "causing cognitive decline" on the whole, in the same way that I don't think the concrete saw causes "lower average limb count per capita", even though it would be perfectly technically correct to make that assessment from a data driven standpoint, but I will agree that I don't think "AI is here to stay", even if (here comes my favourite phrase recently) mostly for practical reasons. I can't say that it is. I don't know what the future holds, and frankly, I wouldn't mind if it didn't.
And frankly, I say this as someone who does think AI has helped them personally, and would have little reservation about letting it do code (code, only code, nothing but the code) if I were to develop a video game. If AI disappeared tomorrow, I think I'd still have learned enough to reasonably start doing this myself, by hand, like in the olden days. It's certainly possible. I do not design systems smarter than me, which is just a good rule of thumb, no matter what tooling you're using.
What I will say is that this position is a rare one, mainly because it does require engaging with the topic, whether you want to or not. I can't blame you if you don't want to, it's a messy topic, and in the same way I don't want to discuss veganism it's perfectly okay to not want to discuss AI and sticking with a basic, maybe a little polemic, opinion of "FUCK ALL AI". Or "LOVE AI FOREVER". And even if I still hold a (marginally) different position to Walker, I'm really glad there's some nuanced thinking on the issue from a seasoned professional.
That’s fascinating. Will programmers somehow stop using it en masse? Why will they do that? Or what do you mean by AI no longer staying?
To give you some idea of where I’m coming from, imagine OpenAI and Anthropic go bankrupt tomorrow and their services shut entirely down. Deepseek weights are open source and virtually as capable, but I can run them on my desktop. There’s no way to put this genie back in the bottle, is there?
Mainly, I'm hedging my bets.
Should have clarified, I don't think we can, with the sufficient conviction, assert that AI is here to stay. We simply don't know. I'll grant that it probably is, but the economy might change again. What if inference does, against all odds, become too expensive? What if the Pareto curve becomes a Pareto line and we just don't see improvements? What if this all just sorta... dies, like NFTs did?
Now again, I'm not saying that I believe that, I personally think that the chance of "AI staying" is probably higher than the inverse, but I also bet on no more on-the-ground military conflicts in Europe, and a few days after I made that statement to a couple of friends, Russia invaded Ukraine.
Forgive me if I've misunderstood, but it seems largely irrelevant if improvement stops. If AI capability stopped improving right here right now forever, we'd still be left with some incredibly useful tools that are adaptable to a huge range of situations. I respect the viewpoint of "we cannot know anything for certain", but I also struggle to find a theoretical reason that would cause mass desertion of AI like we saw with NFTs.
(I'm not unkz, just dropping in)
If I may lovingly place a single word into unkz' mouth, I think that their point is that local inference -- or even just inference sat in someone else's data centre, at the pre-hype rates -- is already feasible, so it'd take like a bacteria that eats silicon wafers to make it stop. Apple's apparently shipping an app to run local models, too, so I can't see how things won't only get more stuck in from here.
There's some wild stuff happening with inference engines now where people are reporting 80 tok/s with Deepseek V4 (Flash?) on a CPU. I honestly don't know if I believe that. But the Qwen model I host gets a few % faster every week with llama.cpp updates. There's a MacOS inference engine that's highly optimized for Apple Silicon combined with Qwen models. I just ordered a new M5 Max Macbook and am looking forward to running models that previously would have been impossible to run on any laptop. Local inference is the future.
Zoinks, that's, uh. Yeah kinda wild. MoE models mitigate how awful local hardware is w.r.t. memory bandwidth, but that's probably like a 1.375 bpw model. Or else one of those crazy sixteen slot Xeons.
But yeah, things will only get faster. There are hard upper limits due to memory bandwidth, as noted, but better draft models, improved quants (e.g. trellis, autoround, etc.), and -- someday -- maybe a change in architecture (e.g. diffusion models) should push local hardware even further.
Fingers crossed! That'll mean this is a bubble and I get to watch OpenAI/Anthropic implode in a fiery economic fireball (admittedly, this will singe everyone else too, but I really want to live to see them go down).
(edit) there's promising stuff coming for CAD modelling, and hopefully robotics will follow that. Fingers crossed, but that stuff is far more theoretical at this point than I'd be willing to bet on (although I'm hoping to get directly involved in it ...)
The Apple Silicon stuff seems to rely on token diffusion. There's all sorts of architectural changes like Qwen 3.8 Flash Next's embeddings table which can be offloaded to disk to save you 10-20 GiB of VRAM.
But it's already profitable to sell people tokens. Hundreds of companies are doing that right now. Look at openrouter and see how many providers there are. Only a small handful have hundreds of billions in VC funding with which they can serve trillions of tokens at a loss.
If the models stop improving right now then we're still left with AI models that make most traditional graphic design work not worth the cost. We're still left with LLMs that can write pretty decent code for a variety of use cases.
Additionally, I don't think it makes sense to vote against AI submissions for environmental reasons. We've never seen judges vote against contestants that flew in internationally to appear in a competition. Or drove a truck 800 miles to get there.
AI is absolutely here to stay. People need to re-enter reality and realize people will almost always opt for the lowest effort option available. I'm not here to fight on behalf of AI companies (cue "leave the billion dollar corporation alone" meme). I don't need to do anything to end up being right.
I certainly hope so. I think we're starting to see cracks in it. Maybe this year, at the New Year's gathering, someone else will bring up AI, and I can see what other peoples' reactions are. 'Cuz last year it was rough. Hopefully people understand that more than just scammers and ex-tech bros can use these tools, though, and that genuinely good things can come from them as well.
As someone that has been mentally on-the-fence a lot about varying levels of AI usage, and trying to reckon with the different implications, I really enjoyed Austin's take here.
As usual with him, it's considered, well-stated, and thoughtful. I thought others on this site might enjoy reading it.