There was a time back in the years after dot com era where search servers were sold with the proprietary software loaded for enterprises to use. A google box running on your local network. Now there are advanced things like elastic and other search specific technologies that you can use.
I think LLMs might follow this market pattern where you can buy something to host yourself and then commoditization happens enough where open source solutions will also evolve to have good enough solutions.
An idea for a disrupting company would be to open source their LLM and offer support and feature development to enterprises as the paid offering, kinda like Red Hat or others doing that model. A key difference is running an LLM locally on decent sized compute is fine but it will be costly to scale on your own.
I feel like there's plenty of tech that has always been proprietary, and still is.
Yes, there's a lot of Open Source, and yes a lot of it is critical, but to see the current state as "open source " is not seeing the whole picture.
Google us not Open Source. Neither is Gmail, Google Maps etc. Then there's Windows, Office etc. What about AWS. Or iOS. Android is sorta open source, except for all the bits that actually make it useful.
And that's before we consider Facebook and friends. Or Netflix. Or Amazon shopping cart.
LLMs are following in the footsteps of pretty much every tech company, and every technology - it starts out life being proprietary. Maybe one day not just the software (llama) but the models will also be open source.
But no, it doesn't bother me at all. Most all Open Source is built on the ideas of proprietary software.
The state of LLMs right now seems analogous to the state of CAD software. There are free options available but they aren't as capable. The good part about this analogy is that the needs of the users aren't very dynamic -- if you are using a CAD software it is for design tasks. The interface may change, but the purpose doesn't. Because of this, the quality gap between free CAD packages and
LLMs are natural language models (what words are likely to come next given the context), not any sort of AGI. For that purpose, the gap between open and closed models is closing much faster in LLMs than CAD. I think LLMs will go the way of chess engines -- one of these models will become the Stockfish of LLMs and the proprietary models will end up being a waste of money and resources.
I mind, a lot. That is why I've built a cheap (in relative terms) rig that can run models up to approximately 600B parameters, although only extremely slowly once the model spills out of the GPUs. I would much rather be able to run open LLMs slowly than not at all.
> I don't like a future where I have to pay for every token just to write a program...
You don't really need to worry then. There is no future like that in the long run. No future at all.
If you think a little, you'll realize AI breaks a number of pillars holding society in a working state.
The thing forces the labor value of time to zero while simultaneously eliminating economic calculation in the factor markets, and eliminating capital formation.
There is a saying, businesses sometimes win so much that they lose. This is one of those times. All that lay ahead is a maelstrom of mathematical chaos, and the only ones that will survive are those not in it.
Windows has been the backbone of desktop computing for decades, and it certainly isn't free. People relying on open-source software seems to be a relatively new phenomenon, starting in the early-2000s.
For what it's worth, I absolutely share your concerns around the fact that LLM models are now proprietary. I want to see more open-source (or at least open-weight) models being built
> I know there are open source alternatives, but they’re not nearly as capable
They certainly are capable (DeepSeek being the obvious example), the problem is that they're still too expensive to run and there's no currently differentiator to compete with the big players who are likely selling inference at a loss.
You're in a bit of an echo chamber, I believe. The concern is real and valid. It just doesn't get spammed to death nearly as much as embracing narratives.
I don't think that's the solution. Proprietary LLMs will just keep growing, and it doesn't seem like the open source alternatives are gaining much traction. I guess it's because you need a lot of money to train high-quality LLMs (tons of energy, maybe?). Besides, as stated in the title of this post, us, as software engineers, the collective, don't seem to mind much the current state of things as far as I can tell.
Strictly speaking - we don't really know if this is true. There is no study proving AI gets smarter up to a certain point. It might keep scaling forever, or one day we might unknowingly reach the soft limit of LLM intelligence. I think assertions like the one you're making require specific evidence.
For comparison's sake, proprietary models like GPT-3 now pale in comparison to the results you get from a 7b Open Source LLM. The Open Source stuff really does move along, if not at the pace everyone would prefer.
LLMs are just a statistical analysis of a large body of written information. The proprietary part is mostly the training. It’ll end up being similar to OSM vs Google maps vs HERE. You can get close with crowd/open sourcing, but you can get significantly better at any given point in time by investing millions or billions. At some point the commodity version will be good enough and then no one will invest any more and we’ll stall out for a while until someone succeeds at the next moonshot.
There was a time back in the years after dot com era where search servers were sold with the proprietary software loaded for enterprises to use. A google box running on your local network. Now there are advanced things like elastic and other search specific technologies that you can use.
I think LLMs might follow this market pattern where you can buy something to host yourself and then commoditization happens enough where open source solutions will also evolve to have good enough solutions.
An idea for a disrupting company would be to open source their LLM and offer support and feature development to enterprises as the paid offering, kinda like Red Hat or others doing that model. A key difference is running an LLM locally on decent sized compute is fine but it will be costly to scale on your own.
I feel like there's plenty of tech that has always been proprietary, and still is.
Yes, there's a lot of Open Source, and yes a lot of it is critical, but to see the current state as "open source " is not seeing the whole picture.
Google us not Open Source. Neither is Gmail, Google Maps etc. Then there's Windows, Office etc. What about AWS. Or iOS. Android is sorta open source, except for all the bits that actually make it useful.
And that's before we consider Facebook and friends. Or Netflix. Or Amazon shopping cart.
LLMs are following in the footsteps of pretty much every tech company, and every technology - it starts out life being proprietary. Maybe one day not just the software (llama) but the models will also be open source.
But no, it doesn't bother me at all. Most all Open Source is built on the ideas of proprietary software.
The state of LLMs right now seems analogous to the state of CAD software. There are free options available but they aren't as capable. The good part about this analogy is that the needs of the users aren't very dynamic -- if you are using a CAD software it is for design tasks. The interface may change, but the purpose doesn't. Because of this, the quality gap between free CAD packages and
LLMs are natural language models (what words are likely to come next given the context), not any sort of AGI. For that purpose, the gap between open and closed models is closing much faster in LLMs than CAD. I think LLMs will go the way of chess engines -- one of these models will become the Stockfish of LLMs and the proprietary models will end up being a waste of money and resources.
I couldn’t care less as long as one of them gives me a lot of value.
I mind, a lot. That is why I've built a cheap (in relative terms) rig that can run models up to approximately 600B parameters, although only extremely slowly once the model spills out of the GPUs. I would much rather be able to run open LLMs slowly than not at all.
What hardware? I have been considering doing the same.
> I don't like a future where I have to pay for every token just to write a program...
You don't really need to worry then. There is no future like that in the long run. No future at all.
If you think a little, you'll realize AI breaks a number of pillars holding society in a working state.
The thing forces the labor value of time to zero while simultaneously eliminating economic calculation in the factor markets, and eliminating capital formation.
There is a saying, businesses sometimes win so much that they lose. This is one of those times. All that lay ahead is a maelstrom of mathematical chaos, and the only ones that will survive are those not in it.
You don't have to use LLMs, of course, and when you want help coding, LLMs are vastly cheaper than human programmers.
I see LLMs as a faster stackoverflow rather than a cheaper human
Windows has been the backbone of desktop computing for decades, and it certainly isn't free. People relying on open-source software seems to be a relatively new phenomenon, starting in the early-2000s.
For what it's worth, I absolutely share your concerns around the fact that LLM models are now proprietary. I want to see more open-source (or at least open-weight) models being built
> I know there are open source alternatives, but they’re not nearly as capable
They certainly are capable (DeepSeek being the obvious example), the problem is that they're still too expensive to run and there's no currently differentiator to compete with the big players who are likely selling inference at a loss.
You're in a bit of an echo chamber, I believe. The concern is real and valid. It just doesn't get spammed to death nearly as much as embracing narratives.
You can host Llama if you are concerned about this. If you have something so valuable you will not give it for free to everyone.
I don't think that's the solution. Proprietary LLMs will just keep growing, and it doesn't seem like the open source alternatives are gaining much traction. I guess it's because you need a lot of money to train high-quality LLMs (tons of energy, maybe?). Besides, as stated in the title of this post, us, as software engineers, the collective, don't seem to mind much the current state of things as far as I can tell.
> Proprietary LLMs will just keep growing
Strictly speaking - we don't really know if this is true. There is no study proving AI gets smarter up to a certain point. It might keep scaling forever, or one day we might unknowingly reach the soft limit of LLM intelligence. I think assertions like the one you're making require specific evidence.
For comparison's sake, proprietary models like GPT-3 now pale in comparison to the results you get from a 7b Open Source LLM. The Open Source stuff really does move along, if not at the pace everyone would prefer.
Llama is also proprietary and not open source.
The "open source means open training pipeline" issue is not relevant to the OP's question.
LLMs are just a statistical analysis of a large body of written information. The proprietary part is mostly the training. It’ll end up being similar to OSM vs Google maps vs HERE. You can get close with crowd/open sourcing, but you can get significantly better at any given point in time by investing millions or billions. At some point the commodity version will be good enough and then no one will invest any more and we’ll stall out for a while until someone succeeds at the next moonshot.