AI will live on but maybe as open weight models running on your laptop. There is no reason to believe that the cloud model is the only possible way for this to work. And there’s only so long that any company can lose billions of dollars a year.
I run a 122B model locally. I love local models. If you think they’re anywhere near as capable as the data center models you’ve either not tried them or are lying to yourself
qwen flash is great. Privacy/sovereignty and protection from others stealing your data is very valuable and equivalence to frontier 3-4 months ago is very impressive. Your point is correct, and larger models have the capacity to solve problems with more general knowledge baked in, where smaller models simply can’t find critical information from search, or develop a strategy as well if they are clueless, and just repeatedly randomly guessing.
Totally agree on the privacy sovereignty front. I suspect in the long run personal tasks will be handled by local models for this reason. I guess I’m thinking more on the scale of institutional compute. Just like every large university has their own super compute cluster I expect them to want their own AI instance (I know companies like Amazon already have such a thing) and in those instances the gap in skill really matters and these institutions are willing to pay a fair amount for the best out there.
Also I can’t speak on equivalence with frontier models 3-4 months ago because I didn’t have access to paid models back then but if that’s true and not just bench maxxing then it means the moat is widening not shrinking. Which suggests these labs are in a good position to make themselves an asset the US cannot afford to lose despite their comically costly business model.
Yeah Qwen 3.5 122B-A10B. It’s the most capable model I’ve found that I can run on my machine. It’s useful for simple tasks but honestly it’s around the border if correcting/checking its works is comparable effort to doing it myself. So I don’t really use it in reality it’s more a novelty. I’ve also messed around with Kimi K3 since the benchmarks said it was awesome. It’s definitely a league above Qwen but must be benchmaxxed to hell cuz it’s not even close to the same league as opus (I can’t speak to OpenAI models as my team has a Claude account so those are the paid models I know)
AI will live on but maybe as open weight models running on your laptop. There is no reason to believe that the cloud model is the only possible way for this to work. And there’s only so long that any company can lose billions of dollars a year.
I run a 122B model locally. I love local models. If you think they’re anywhere near as capable as the data center models you’ve either not tried them or are lying to yourself
qwen flash is great. Privacy/sovereignty and protection from others stealing your data is very valuable and equivalence to frontier 3-4 months ago is very impressive. Your point is correct, and larger models have the capacity to solve problems with more general knowledge baked in, where smaller models simply can’t find critical information from search, or develop a strategy as well if they are clueless, and just repeatedly randomly guessing.
Totally agree on the privacy sovereignty front. I suspect in the long run personal tasks will be handled by local models for this reason. I guess I’m thinking more on the scale of institutional compute. Just like every large university has their own super compute cluster I expect them to want their own AI instance (I know companies like Amazon already have such a thing) and in those instances the gap in skill really matters and these institutions are willing to pay a fair amount for the best out there.
Also I can’t speak on equivalence with frontier models 3-4 months ago because I didn’t have access to paid models back then but if that’s true and not just bench maxxing then it means the moat is widening not shrinking. Which suggests these labs are in a good position to make themselves an asset the US cannot afford to lose despite their comically costly business model.
I know that is the case today. But hardware and software will improve and the task sets they will be capable of will grow.
Are you using Qwen? Where do you find it useful/not useful?
Yeah Qwen 3.5 122B-A10B. It’s the most capable model I’ve found that I can run on my machine. It’s useful for simple tasks but honestly it’s around the border if correcting/checking its works is comparable effort to doing it myself. So I don’t really use it in reality it’s more a novelty. I’ve also messed around with Kimi K3 since the benchmarks said it was awesome. It’s definitely a league above Qwen but must be benchmaxxed to hell cuz it’s not even close to the same league as opus (I can’t speak to OpenAI models as my team has a Claude account so those are the paid models I know)