If LLMs can’t be made more power efficient, then the technology isn’t ready yet for general use. There may have to be fundamental engineering breakthroughs before what these systems do can be more than a very expensive and inefficient worldwide proof of concept.
Of course, the military and the surveillance industry will still want to keep it, even if no one else can afford it, and we’ll be the ones paying for it.
They CAN be made more efficient. Hell, AMD just bought a company that translated LLMs to direct silicon… which should enable scalability if it wasn’t for one of the key players owning the tech thus limiting its use.
It’s also estimated that some 70% of Claude tokens gets spent on recursive self-correction so the output is higher quality. Instead of bettering the model, Anthropic just keeps throwing money at it to make the problem go away.
Meanwhile Chinese models - as small as 36-40-80b params - are outdoing these fancy frontier models if the right harness is used.
The truly silly (and simultaneously neat and fascinating, and to a large extent, very logical and obvious) part is that US attempts to cripple Chinese ML/LLM development by putting most advanced GPU tech behind ITAR has more or less made the Chinese ML industry just write straight up better and more optimized code for the domain.
US and western companies are currently relying on throwing bleeding edge hardware at the problem (read: very obviously brute-forcing the problem), whereas Chinese companies and engineers, having to rely on comparatively less advanced (though still highly capable) hardware has meant they’re staying competitive by optimizing the software layer more effectively.
This has very interesting and likely negative long-term implications for western engineering practices in a lot of ways. And it’s also very much not a new problem - simply throwing compute at things is a mindset that became highly prevalent in the west around when The Cloud became a thing. Previously, you couldn’t just spawn another mainframe in your company’s datacenter, but when it’s as simple as aws ec2 start-instances ..., the barrier to entry for getting more compute is so low that most people were just going “hey, it’s faster”. I feel like the engineering expertise around embedded/baremetal optimization has just been largely evaporating over the last couple decades; in the vast majority of professional software domains, it’s simply not a valued skill anymore (or, more correctly, at the moment). It’s sad for me in several ways - not least of which because I did my degree in EECS, and hardware/software interplay and optimization is an area of engineering that I have always found very interesting.
I’ve spent the last month or so using nothing but a local 27B Qwen model as my coding assistant and it’s working fine. The big American AI companies could all collapse in ruins tomorrow and it’s not going to slow things down all that much, anyone who wants to use AI will be able to continue using it. Might even be better in the long run if that happens - the people who are scared of ASI will get their much-demanded temporary slowdown at the frontier while everyone else catches up with open-weight models that everyone can use without central control.
I run a 5 bit quantized Qwen 3.8 27b at home on a gaming PC with an RX 7900 XTX and RAG memory using Gemma MoE (i forget what one)running on an old 1080 ti on my Valheim server.
It takes 3-15 minutes to run depending on how many tool calls the model makes and that’s mostly memory or web scrapes since that runs on the slower hardware. I mostly just use it as a discord bot to look stuff up for me and to fetch the news.
1080 Ti? But that is 11 GB of VRAM? I know because that is my GPU and the reason I have not tried it. They all call for 12 GB VRAM. I have 64 GB RAM, so 50 GB for the model would be fine.
On the 1080ti I run gemma4:e4b to to do RAG and vision jobs. Not sure the quants and I just run it in Ollama that auto adjust the model because I’m lazy. Not 100% sure but I think mixture of experts load the parts that are needed so there easier on memory. Offloading to CPU memory or swap is a big performance hit.
I tried to run a 300b model on the main PC once, but it took half an hour to get a few words out of it. You can go past the VRAM on the card but sacrifice speed.
The kind of specs that would enable local LLM usage were available on laptops since 2020. The floor for acceptable performance for LLMs stopped literally just before 24+gb of VRAM would have become pretty common in consumer applications. The reason this technology isn’t more ubiquitously available is because ‘they’ thought it would be more profitable to hoover up literally every byte of RAM and GPU in the world rather than let the technology continue to naturally evolve to a place where any consumer could have run useful models on their home computers. In other words, your anger is misplaced.
Wanted to refactor a too large and hectic hacked together state machine into separate functions and reduce redundancy and increase code readability
Every model I’ve tried failed.
And that should be an easy task an actual AI should be able to handle.
Logic is done, just translate to a different structure.
That should be exactly what they are able to do at least.
None worked, because the project is just to big and interlinked.
I’ve got 24GB of VRAM, which helps a lot. I’m using the Q4_K_M version of this model, with the KV cache quantized to Q8, and that’s able to get me to 128000 tokens of context. I’m using the Qwen Code harness in VS Code.
If 128000 tokens still isn’t enough, break the task down into smaller chunks and help the harness out with MCP tools - I installed Codebase Memory to help it navigate large codebases more efficiently, for example. Ask the coding agent what’s causing problems and it may be able to help identify solutions too, I’ve had it troubleshoot itself several times now and it’s been quite good at inspecting its own settings and finding where the problems are with them. I even installed Firecrawl for it so it can do its own web searches.
There may have to be fundamental engineering breakthroughs before what these systems do can be more than a very expensive and inefficient worldwide proof of concept.
Controversial thought after reading this sentence: It’s cryptocoins all over again.
That’d be fine. Cryptocurrency as a whole is still going strong (total market cap is around 3 trillion dollars) and most of the major ones have switched over to energy-free proof-of-stake validation.
If LLMs can’t be made more power efficient, then the technology isn’t ready yet for general use. There may have to be fundamental engineering breakthroughs before what these systems do can be more than a very expensive and inefficient worldwide proof of concept.
Of course, the military and the surveillance industry will still want to keep it, even if no one else can afford it, and we’ll be the ones paying for it.
They CAN be made more efficient. Hell, AMD just bought a company that translated LLMs to direct silicon… which should enable scalability if it wasn’t for one of the key players owning the tech thus limiting its use.
It’s also estimated that some 70% of Claude tokens gets spent on recursive self-correction so the output is higher quality. Instead of bettering the model, Anthropic just keeps throwing money at it to make the problem go away.
Meanwhile Chinese models - as small as 36-40-80b params - are outdoing these fancy frontier models if the right harness is used.
The truly silly (and simultaneously neat and fascinating, and to a large extent, very logical and obvious) part is that US attempts to cripple Chinese ML/LLM development by putting most advanced GPU tech behind ITAR has more or less made the Chinese ML industry just write straight up better and more optimized code for the domain.
US and western companies are currently relying on throwing bleeding edge hardware at the problem (read: very obviously brute-forcing the problem), whereas Chinese companies and engineers, having to rely on comparatively less advanced (though still highly capable) hardware has meant they’re staying competitive by optimizing the software layer more effectively.
This has very interesting and likely negative long-term implications for western engineering practices in a lot of ways. And it’s also very much not a new problem - simply throwing compute at things is a mindset that became highly prevalent in the west around when The Cloud became a thing. Previously, you couldn’t just spawn another mainframe in your company’s datacenter, but when it’s as simple as
aws ec2 start-instances ..., the barrier to entry for getting more compute is so low that most people were just going “hey, it’s faster”. I feel like the engineering expertise around embedded/baremetal optimization has just been largely evaporating over the last couple decades; in the vast majority of professional software domains, it’s simply not a valued skill anymore (or, more correctly, at the moment). It’s sad for me in several ways - not least of which because I did my degree in EECS, and hardware/software interplay and optimization is an area of engineering that I have always found very interesting.I’ve spent the last month or so using nothing but a local 27B Qwen model as my coding assistant and it’s working fine. The big American AI companies could all collapse in ruins tomorrow and it’s not going to slow things down all that much, anyone who wants to use AI will be able to continue using it. Might even be better in the long run if that happens - the people who are scared of ASI will get their much-demanded temporary slowdown at the frontier while everyone else catches up with open-weight models that everyone can use without central control.
Pity about the economy in the meantime, I guess.
What kind of hardware do you need to run a model like that? And what do you use :-) ?
I run a 5 bit quantized Qwen 3.8 27b at home on a gaming PC with an RX 7900 XTX and RAG memory using Gemma MoE (i forget what one)running on an old 1080 ti on my Valheim server.
It takes 3-15 minutes to run depending on how many tool calls the model makes and that’s mostly memory or web scrapes since that runs on the slower hardware. I mostly just use it as a discord bot to look stuff up for me and to fetch the news.
1080 Ti? But that is 11 GB of VRAM? I know because that is my GPU and the reason I have not tried it. They all call for 12 GB VRAM. I have 64 GB RAM, so 50 GB for the model would be fine.
On the 1080ti I run gemma4:e4b to to do RAG and vision jobs. Not sure the quants and I just run it in Ollama that auto adjust the model because I’m lazy. Not 100% sure but I think mixture of experts load the parts that are needed so there easier on memory. Offloading to CPU memory or swap is a big performance hit.
I tried to run a 300b model on the main PC once, but it took half an hour to get a few words out of it. You can go past the VRAM on the card but sacrifice speed.
Anyone who’s privileged enough to have the hardware.
The kind of specs that would enable local LLM usage were available on laptops since 2020. The floor for acceptable performance for LLMs stopped literally just before 24+gb of VRAM would have become pretty common in consumer applications. The reason this technology isn’t more ubiquitously available is because ‘they’ thought it would be more profitable to hoover up literally every byte of RAM and GPU in the world rather than let the technology continue to naturally evolve to a place where any consumer could have run useful models on their home computers. In other words, your anger is misplaced.
I’m always running into context limit problems
Wanted to refactor a too large and hectic hacked together state machine into separate functions and reduce redundancy and increase code readability
Every model I’ve tried failed.
And that should be an easy task an actual AI should be able to handle. Logic is done, just translate to a different structure.
That should be exactly what they are able to do at least.
None worked, because the project is just to big and interlinked.
I’ve got 24GB of VRAM, which helps a lot. I’m using the Q4_K_M version of this model, with the KV cache quantized to Q8, and that’s able to get me to 128000 tokens of context. I’m using the Qwen Code harness in VS Code.
If 128000 tokens still isn’t enough, break the task down into smaller chunks and help the harness out with MCP tools - I installed Codebase Memory to help it navigate large codebases more efficiently, for example. Ask the coding agent what’s causing problems and it may be able to help identify solutions too, I’ve had it troubleshoot itself several times now and it’s been quite good at inspecting its own settings and finding where the problems are with them. I even installed Firecrawl for it so it can do its own web searches.
Controversial thought after reading this sentence: It’s cryptocoins all over again.
That’d be fine. Cryptocurrency as a whole is still going strong (total market cap is around 3 trillion dollars) and most of the major ones have switched over to energy-free proof-of-stake validation.