The first LLaMA is what really opened the floodgates on the hobbyist/"open source" side of LLM development. Facebook ostensibly required users be researchers who signed an agreement promising...
The first LLaMA is what really opened the floodgates on the hobbyist/"open source" side of LLM development. Facebook ostensibly required users be researchers who signed an agreement promising non-commercial usage, although in practice they let anyone with a .edu email in (even I got in), and the weights appeared in a torrent not long after they started releasing them.
Despite that, the fact that it was not allowed for commercial usage put a damper on the full extent of what people used it for.
Now, it seems Meta has gone full "who cares" and you can do whatever you want with the models.
They even provide a model fine-tuned on chat instructions!
The genie is well and truly out of the bottle now, for better or for worse. I imagine there's going to be a lot of LLMs popping up in any and every product and application you can think of.
As someone who was following that space before, I disagree with you a little bit there. Llama has been an amazing contribution, but the open source LLMs groups were very alive and having fun with...
The first LLaMA is what really opened the floodgates on the hobbyist/"open source" side of LLM development
As someone who was following that space before, I disagree with you a little bit there. Llama has been an amazing contribution, but the open source LLMs groups were very alive and having fun with models well before llama, to the point that they took a while to get support in after it was finally leaked.
Great contribution, but it was around and would be around still without. Open AI are unfortunately the ones who kicked that off, and who are now trying to kill it.
The competition. They try to argue only "experts" like them should have a right to work with AI. Their strategy ironically involves scaremongering about AI in general:...
Just saw this and came to tildes to see what people think. Having a large well trained model that's open to commercial use is really going to be a game changer. This is what that "we have no moat"...
Just saw this and came to tildes to see what people think. Having a large well trained model that's open to commercial use is really going to be a game changer. This is what that "we have no moat" memo was really worried about.
Looking forward to playing with it. I wonder how easy it will be to add additional training to fine tune it for custom applications and sectors? Llama had a flood of custom models built on top of it so I'm sure we'll see the same with this.
There have been a few products I've been working on that I thought "hmm an llm assistant so users can get more info based on their specific context before reaching out would be handy" but the current closed source offerings weren't really viable.
Offering it for free to any commercial applications with less than 700 million users is insanely generous. Meta aren't a cloud provider like the other big players, so I'm curious what their strategic thinking is? If Microsoft or Google open sourced and said "the best place to run your models is in our cloud" I would get it. I'm welcome for the open source nature of this model, but I don't understand it which makes me suspicious.
EEE is for when the other party is smaller, but between Microsoft (who de facto is aligned with OpenAI) and Google, Facebook is the smaller party by far in terms of resources. This is just good ol...
EEE is for when the other party is smaller, but between Microsoft (who de facto is aligned with OpenAI) and Google, Facebook is the smaller party by far in terms of resources.
I'm suspicious, EEE could still be the end-goal because good new like this is how Facebook sets itself to become the standard in offline AI. In the end it is about being competitive and winning...
I'm suspicious, EEE could still be the end-goal because good new like this is how Facebook sets itself to become the standard in offline AI. In the end it is about being competitive and winning over developers... but how they choose to change the ecosystem if they are top-dog is what matters in the long-run. In the meantime, everyone wins.
This is the less but still nefarious tactic. Give the product that most people can't be bothered to self manage away for free so you become the defacto player in the space Sell that product as a...
This is the less but still nefarious tactic.
Give the product that most people can't be bothered to self manage away for free so you become the defacto player in the space
Sell that product as a cloud service and reap the benefits when the enterprise players need a cloud offering.
That’s not really the case here. Facebook isn’t offering any managed ways to run LLama - the closest is that they partnered with Microsoft so it’s available on the azure model catalog, but if...
That’s not really the case here. Facebook isn’t offering any managed ways to run LLama - the closest is that they partnered with Microsoft so it’s available on the azure model catalog, but if anything that’s entrenching you into Microsoft’s business.
They’re just offering the model weights and architecture.
For better or for worse Meta has never really had serious B2B offerings.
FWIW, the MAU limit is measured at the point in time that Llama was released. Small thing, but it's clearly written to prevent their large competitors from using the model.
FWIW, the MAU limit is measured at the point in time that Llama was released. Small thing, but it's clearly written to prevent their large competitors from using the model.
The first LLaMA is what really opened the floodgates on the hobbyist/"open source" side of LLM development. Facebook ostensibly required users be researchers who signed an agreement promising non-commercial usage, although in practice they let anyone with a .edu email in (even I got in), and the weights appeared in a torrent not long after they started releasing them.
Despite that, the fact that it was not allowed for commercial usage put a damper on the full extent of what people used it for.
Now, it seems Meta has gone full "who cares" and you can do whatever you want with the models.
They even provide a model fine-tuned on chat instructions!
The genie is well and truly out of the bottle now, for better or for worse. I imagine there's going to be a lot of LLMs popping up in any and every product and application you can think of.
As someone who was following that space before, I disagree with you a little bit there. Llama has been an amazing contribution, but the open source LLMs groups were very alive and having fun with models well before llama, to the point that they took a while to get support in after it was finally leaked.
Great contribution, but it was around and would be around still without. Open AI are unfortunately the ones who kicked that off, and who are now trying to kill it.
I missed this, what is OpenAI trying to kill?
The competition. They try to argue only "experts" like them should have a right to work with AI. Their strategy ironically involves scaremongering about AI in general: https://arstechnica.com/information-technology/2023/05/openai-execs-warn-of-risk-of-extinction-from-artificial-intelligence-in-new-open-letter/
Oh my goodness. Thank you for the link!
Just saw this and came to tildes to see what people think. Having a large well trained model that's open to commercial use is really going to be a game changer. This is what that "we have no moat" memo was really worried about.
Looking forward to playing with it. I wonder how easy it will be to add additional training to fine tune it for custom applications and sectors? Llama had a flood of custom models built on top of it so I'm sure we'll see the same with this.
There have been a few products I've been working on that I thought "hmm an llm assistant so users can get more info based on their specific context before reaching out would be handy" but the current closed source offerings weren't really viable.
Offering it for free to any commercial applications with less than 700 million users is insanely generous. Meta aren't a cloud provider like the other big players, so I'm curious what their strategic thinking is? If Microsoft or Google open sourced and said "the best place to run your models is in our cloud" I would get it. I'm welcome for the open source nature of this model, but I don't understand it which makes me suspicious.
What do you folks think their thinking is?
Embrace, Extend, Extinguish?
Nothing, absolutely nothing, mega corps do for "free" is ever really free.
EEE is for when the other party is smaller, but between Microsoft (who de facto is aligned with OpenAI) and Google, Facebook is the smaller party by far in terms of resources.
This is just good ol fashioned competition.
I'm suspicious, EEE could still be the end-goal because good new like this is how Facebook sets itself to become the standard in offline AI. In the end it is about being competitive and winning over developers... but how they choose to change the ecosystem if they are top-dog is what matters in the long-run. In the meantime, everyone wins.
This is the less but still nefarious tactic.
Give the product that most people can't be bothered to self manage away for free so you become the defacto player in the space
Sell that product as a cloud service and reap the benefits when the enterprise players need a cloud offering.
That’s not really the case here. Facebook isn’t offering any managed ways to run LLama - the closest is that they partnered with Microsoft so it’s available on the azure model catalog, but if anything that’s entrenching you into Microsoft’s business.
They’re just offering the model weights and architecture.
For better or for worse Meta has never really had serious B2B offerings.
Nothing yet. I'm sure they plan to do something to make money here.
just on your last point- isn't their whole revenue model ads, I don't see how that's not b2b?
That makes sense given the history of this company.
FWIW, the MAU limit is measured at the point in time that Llama was released. Small thing, but it's clearly written to prevent their large competitors from using the model.