SemiAnalysis has calculated how big that gap really is. After testing subscription tiers from both OpenAI and Anthropic – running long-horizon coding and agentic tasks until weekly...
If you’ve got a toy project that you want “AI” to give you a hand with, do it now.
Pretty soon all these companies are going to have to pay for all that investment in compute resources they’ve been busily soaking up over the last few years, and then they’re going to have to pay back their investors, and then they’re going to have to try and make a profit
This is the golden time for cheap commercial AI. Already the noose is starting to tighten, and it will never again be as cheap as it is now.
Yeah they’ve been pushing Claude code at work for us non coders jobs to come up with stuff that would help us. We’ve gotten a few surprisingly useful programs out of it, but our assumption is perfect them now before pricing goes through the roof. We are also only creating programs that do not require ongoing AI use. Just a bunch of relatively simple things that make our jobs easier.
I am still pushing my boss for some local hw as I think as a group we’ve spent a couple grand in the last month and that is the least of my reasons for wanting a local llm vs subscription.
I suppose, but small open weight models with more advanced coding frameworks optimized for them are catching up fast and you can do it privately at home on a mostly affordable consumer graphics card.
If you have solar it’s basically free, minus the graphics card CapEx you may want for gaming anyway, as well as some setup time and a bit of patience.
Yes, it’s trending in that direction, and I’ve been experimenting with pretty small models on my PC as I don’t really have the hardware to go large. If you’ve got the coding chops to set it up, it’s definitely something to keep an eye on.
There’s actually scope for someone to set up / sell local compute hardware+software packages, similar to all those coin miners. Give the end user a way to update models, or push models out to them or something, it seems it would be a good middle ground between manually typing code like a peasant and total corporate AI apocalypse.
There’s actually scope for someone to set up / sell local compute hardware+software packages, similar to all those coin miners.
I think that’ll be a viable target in the future, and have little doubt some are jumping on it already. However, I also think it’s too much of a moving target currently, a near optimal setup changes almost entirely month to month.
I find myself targeting last months setup, as then there’s enough literature out there to get it set up in a day or two and most of the kinks have been worked out. Otherwise, I lose too much coding time to debugging the bleeding edge.
IMO, at the moment, if you’re not capable of setting it up yourself you likely don’t have the experience to use it reasonably safely nor an adequate understanding of its limitations. You’ll find yourself using more time fixing the blunders than you gain, and / or the project will spiral out of control in maintainability, security, readability, and so forth. You could get away with small projects written as ‘write only’ code ala Perl though, keep the prompts and tests, when it needs to change rebuild with the newest hotness. Inefficient and unsatisfying though.
What’s your setup, if I may ask? I’m using llama.cpp router with vscode kilo.ai and qwen3.6-35B-MoE-MTP as a model mostly. It’s surprisingly good as a coding assistant, but I think you have to know what you are doing and know your stuff(aka be an experienced developer) to make it useful. just letting it vibe leads to crap code
Yup, vibe is occasionally useful for proof of concept stuff, but disastrous for maintainability, security, readability, or large codebases. Without experience it’s still a foot gun for anything even slightly serious.
Best approaches for a learner are to consider it autocomplete that needs research. Look up what it’s suggesting, see if it’s hallucinating, with luck it’ll point you in a useful direction where you can learn a good solution, as it has no idea what that is. Also makes a pretty good rubber duck for hashing out architectural decisions, finding alternative approaches etc, though you’ll have to point it at a web search for that. Spin up an e.g. vane instance for this, as small models don’t have enough world knowledge. Use it to write (or preferably copy from its system prompt examples) boilerplate and unit tests, perhaps descriptive comments (doublecheck).
One thing to do is put everything you learn about coding style into your system prompt as they’re dogshit at consistent style without significant beatings around the head. Finding your own comfortable, consistent style is super useful for future readability. The joke about when I wrote this only God and I understood it, now only God does, will come clear in a month or two. Learn to work around it. Simple beats fancy unless you truly need the speed.
While I do use agent iterative approaches, probably best to approach that organically as you grow, monsters lurk there. If you must, containerize / vm / isolate the hell out of something like opencode to muck around with.
FWIW I still write most of my code by hand, it’s simpler and more consistent, but I’m keeping an eye on the development of LLMs, and I will let it write scut code (that I edit later). Code and Mathematics are super structured languages, pretty much ideal for large language models, so I can see them maybe, eventually getting good. More general thought, not so much without significant architectural upgrades.
While this advice is true for all models, when it comes to agentic tasks (add this small feature/write this test harness/find bugs/suggest improvements), open source models are still way behind, vibe code or not.
Claude Fable or even Opus in an editor like Zed have a 1 million token context window and will “think” through the goals of the application, test their changes, work through debugging processes the way a programmer would, stop to ask for clarification, check diagnostic tools and linters, prompt to run test code, etc.
Llama, Gemma and Qwen etc. Do lack a lot of the world knowledge to get the goals of the application, but they also just don’t have the debugging skills, won’t test their code, don’t always tool call correctly, get confused as the context increases and nobody has enough vram to run on large context sizes locally.
They can do autocomplete on small functions but aren’t really there for more complex tasks.
On top of that, the biggest problem is that the best open source models are trained and released by the same giant tech conglomerates that have an interest in not competing with their own products. Qwen is Alibaba, Llama is Meta, gpt-oss is OpenAI. Even the more “independent” ones, kimi (Moonshot) and GLM (z.ai) are mostly funded by Alibaba and Tencent. They’re released for research and marketing purposes and to please their corporate backers with inflated stock. Almost nobody has the resources to train new models from scratch. People make lots of merges and fine tunes but AI is not democratised the way that traditional programming tools have been.
Maybe some day there will be enough cheap compute for open source communities to pool together resources to build competing models but they’re not really there yet :(
Context management is a huge part of making smaller models viable (and likely a big part of making frontier models better). Tricks like structured context libraries for thinking improve things a lot, I like approaches that output things like an Obsidian vault that let you dig in and correct bad assumptions easily, even if it’s a bit slower. It’s a useful deliverable that can (mostly) be reused with updated models.
Things like ‘the debugging skills, won’t test their code, don’t always tool call correctly’ are tangibly improving model to model, framework to framework, and are problems that will be solved in time, but yes they need handholding ATM.
Things like
test their changes, work through debugging processes the way a programmer would, stop to ask for clarification, check diagnostic tools and linters, prompt to run test code
are mostly down to framework, not model (except for failing to tool call, which is improving), and falling at a respectable rate.
That said, sure, frontier models get more in one go, personally I’m fine with only a 3-4x force multiplier instead of 10 to keep it local, but YMMV. For a business with resources for a bigger server it’ll be more like 8 times. Remember that some businesses handle sensitive data and can’t (or damn well shouldn’t) use frontier models, so the market is there.
Maybe some day there will be enough cheap compute for open source communities to pool together resources to build competing models but they’re not really there yet :(
Not wrong, decentralized inference is mostly solved (with latency penalties), but without decentralized training true democratization will remain out of reach. Hopefully a breakthrough will ensue, but until then we are dependent on the kindness of corporations (or them rugpulling competitors).
This could also be a part of the RAMpocalypse thing, ‘if there’s not a moat I’ll fucking dig one, damn everyone else’ (and damn SamA). I doubt that’s sustainable long term, but it might get them through to IPO, more’s the pity.
In five years once this RAM nonsense is over you’ll be able to run a comparatively high quality local LLM for very little money. I can’t see how these companies will ever make their money back.
I’m slightly optimistic that manufacturers will return to the retail market eventually. Every AI company is racing to hyperscale right now but there will be a point where the infrastructure is built and at that point the growth will slow down quite a bit. In that scenario there will be ongoing demand for components to be replaced as they become obsolete but I can’t imagine the demand will be the same level it is right now as everyone rushes to build.
That’s assuming this all works the way they want it to. If the economics aren’t viable and the bubble bursts…
Their Datacenter buildout doesn’t work they want to. Most projects are very much delayed, and those that even started getting built are over budget. OpenAI and Anthropic will collapse in the next years, and this is coming from someone who absolutely sees the good things about the technology itself.
There is no way, absolutely NO WAY to recuperate the amount of cash burnt on those two companies, and that is not even counting the amount of AI Startup whose cash is currently flowing towards to those two.
How about bailouts? GPT integration with Visa? IPOs sucked into funds? There’s a lot of money they can and will try to vacuum, don’t you worry about that. And banks will do their damnest best to help with all of that - just look at SpaceX.
If you’ve got a toy project that you want “AI” to give you a hand with, do it now.
Pretty soon all these companies are going to have to pay for all that investment in compute resources they’ve been busily soaking up over the last few years, and then they’re going to have to pay back their investors, and then they’re going to have to try and make a profit
This is the golden time for cheap commercial AI. Already the noose is starting to tighten, and it will never again be as cheap as it is now.
Another way to look at it would be “if you’ve got a toy project to practice coding without AI on, do it now” before that is the only option.
Yeah they’ve been pushing Claude code at work for us non coders jobs to come up with stuff that would help us. We’ve gotten a few surprisingly useful programs out of it, but our assumption is perfect them now before pricing goes through the roof. We are also only creating programs that do not require ongoing AI use. Just a bunch of relatively simple things that make our jobs easier.
I am still pushing my boss for some local hw as I think as a group we’ve spent a couple grand in the last month and that is the least of my reasons for wanting a local llm vs subscription.
I suppose, but small open weight models with more advanced coding frameworks optimized for them are catching up fast and you can do it privately at home on a mostly affordable consumer graphics card.
If you have solar it’s basically free, minus the graphics card CapEx you may want for gaming anyway, as well as some setup time and a bit of patience.
Yes, it’s trending in that direction, and I’ve been experimenting with pretty small models on my PC as I don’t really have the hardware to go large. If you’ve got the coding chops to set it up, it’s definitely something to keep an eye on.
There’s actually scope for someone to set up / sell local compute hardware+software packages, similar to all those coin miners. Give the end user a way to update models, or push models out to them or something, it seems it would be a good middle ground between manually typing code like a peasant and total corporate AI apocalypse.
I think that’ll be a viable target in the future, and have little doubt some are jumping on it already. However, I also think it’s too much of a moving target currently, a near optimal setup changes almost entirely month to month.
I find myself targeting last months setup, as then there’s enough literature out there to get it set up in a day or two and most of the kinks have been worked out. Otherwise, I lose too much coding time to debugging the bleeding edge.
IMO, at the moment, if you’re not capable of setting it up yourself you likely don’t have the experience to use it reasonably safely nor an adequate understanding of its limitations. You’ll find yourself using more time fixing the blunders than you gain, and / or the project will spiral out of control in maintainability, security, readability, and so forth. You could get away with small projects written as ‘write only’ code ala Perl though, keep the prompts and tests, when it needs to change rebuild with the newest hotness. Inefficient and unsatisfying though.
What’s your setup, if I may ask? I’m using llama.cpp router with vscode kilo.ai and qwen3.6-35B-MoE-MTP as a model mostly. It’s surprisingly good as a coding assistant, but I think you have to know what you are doing and know your stuff(aka be an experienced developer) to make it useful. just letting it vibe leads to crap code
Yup, vibe is occasionally useful for proof of concept stuff, but disastrous for maintainability, security, readability, or large codebases. Without experience it’s still a foot gun for anything even slightly serious.
Best approaches for a learner are to consider it autocomplete that needs research. Look up what it’s suggesting, see if it’s hallucinating, with luck it’ll point you in a useful direction where you can learn a good solution, as it has no idea what that is. Also makes a pretty good rubber duck for hashing out architectural decisions, finding alternative approaches etc, though you’ll have to point it at a web search for that. Spin up an e.g. vane instance for this, as small models don’t have enough world knowledge. Use it to write (or preferably copy from its system prompt examples) boilerplate and unit tests, perhaps descriptive comments (doublecheck).
One thing to do is put everything you learn about coding style into your system prompt as they’re dogshit at consistent style without significant beatings around the head. Finding your own comfortable, consistent style is super useful for future readability. The joke about when I wrote this only God and I understood it, now only God does, will come clear in a month or two. Learn to work around it. Simple beats fancy unless you truly need the speed.
While I do use agent iterative approaches, probably best to approach that organically as you grow, monsters lurk there. If you must, containerize / vm / isolate the hell out of something like opencode to muck around with.
FWIW I still write most of my code by hand, it’s simpler and more consistent, but I’m keeping an eye on the development of LLMs, and I will let it write scut code (that I edit later). Code and Mathematics are super structured languages, pretty much ideal for large language models, so I can see them maybe, eventually getting good. More general thought, not so much without significant architectural upgrades.
While this advice is true for all models, when it comes to agentic tasks (add this small feature/write this test harness/find bugs/suggest improvements), open source models are still way behind, vibe code or not.
Claude Fable or even Opus in an editor like Zed have a 1 million token context window and will “think” through the goals of the application, test their changes, work through debugging processes the way a programmer would, stop to ask for clarification, check diagnostic tools and linters, prompt to run test code, etc.
Llama, Gemma and Qwen etc. Do lack a lot of the world knowledge to get the goals of the application, but they also just don’t have the debugging skills, won’t test their code, don’t always tool call correctly, get confused as the context increases and nobody has enough vram to run on large context sizes locally.
They can do autocomplete on small functions but aren’t really there for more complex tasks.
On top of that, the biggest problem is that the best open source models are trained and released by the same giant tech conglomerates that have an interest in not competing with their own products. Qwen is Alibaba, Llama is Meta, gpt-oss is OpenAI. Even the more “independent” ones, kimi (Moonshot) and GLM (z.ai) are mostly funded by Alibaba and Tencent. They’re released for research and marketing purposes and to please their corporate backers with inflated stock. Almost nobody has the resources to train new models from scratch. People make lots of merges and fine tunes but AI is not democratised the way that traditional programming tools have been.
Maybe some day there will be enough cheap compute for open source communities to pool together resources to build competing models but they’re not really there yet :(
Context management is a huge part of making smaller models viable (and likely a big part of making frontier models better). Tricks like structured context libraries for thinking improve things a lot, I like approaches that output things like an Obsidian vault that let you dig in and correct bad assumptions easily, even if it’s a bit slower. It’s a useful deliverable that can (mostly) be reused with updated models.
Things like ‘the debugging skills, won’t test their code, don’t always tool call correctly’ are tangibly improving model to model, framework to framework, and are problems that will be solved in time, but yes they need handholding ATM.
Things like
are mostly down to framework, not model (except for failing to tool call, which is improving), and falling at a respectable rate.
That said, sure, frontier models get more in one go, personally I’m fine with only a 3-4x force multiplier instead of 10 to keep it local, but YMMV. For a business with resources for a bigger server it’ll be more like 8 times. Remember that some businesses handle sensitive data and can’t (or damn well shouldn’t) use frontier models, so the market is there.
Not wrong, decentralized inference is mostly solved (with latency penalties), but without decentralized training true democratization will remain out of reach. Hopefully a breakthrough will ensue, but until then we are dependent on the kindness of corporations (or them rugpulling competitors).
This could also be a part of the RAMpocalypse thing, ‘if there’s not a moat I’ll fucking dig one, damn everyone else’ (and damn SamA). I doubt that’s sustainable long term, but it might get them through to IPO, more’s the pity.
Sounds like it’ll never be worth it.
In five years once this RAM nonsense is over you’ll be able to run a comparatively high quality local LLM for very little money. I can’t see how these companies will ever make their money back.
If manufacturers are willing to sell components to us in five years that is.
Of course if the colllapse happens before then the story might be different…
I’m slightly optimistic that manufacturers will return to the retail market eventually. Every AI company is racing to hyperscale right now but there will be a point where the infrastructure is built and at that point the growth will slow down quite a bit. In that scenario there will be ongoing demand for components to be replaced as they become obsolete but I can’t imagine the demand will be the same level it is right now as everyone rushes to build.
That’s assuming this all works the way they want it to. If the economics aren’t viable and the bubble bursts…
“Hyperscale” is utterly meaningless MBA jargon at this point. Equivalent of verbal slop from industry shills and CNBC/Bloomberg sell side simps.
Sorry if that’s true. I understood the word to mean aggressive growth at any cost to try and shut out competition before they can get established.
Their Datacenter buildout doesn’t work they want to. Most projects are very much delayed, and those that even started getting built are over budget. OpenAI and Anthropic will collapse in the next years, and this is coming from someone who absolutely sees the good things about the technology itself.
Stop, I can only handle so much good news!
There is no way, absolutely NO WAY to recuperate the amount of cash burnt on those two companies, and that is not even counting the amount of AI Startup whose cash is currently flowing towards to those two.
How about bailouts? GPT integration with Visa? IPOs sucked into funds? There’s a lot of money they can and will try to vacuum, don’t you worry about that. And banks will do their damnest best to help with all of that - just look at SpaceX.
🤞
Sounds like price hikes to communicate costs are coming and resources are going to be redistributed to productive uses.
deleted by creator