Everyone keeps saying this is going to kill open AI. They still secured 0.5T to buy hardware. They are still going to be a huge player no matter what software they run. They can run R1 if they want.
OpenAI is still going to be way too fucking powerful. It's good that this kicks them down a peg but everyone acts like this is a nail in a coffin.
Also the 0.5T goes to hiring developers too. They are going to soak up all of the American talent leaving much less talent in other things over here. If you want openAI to die or at least not dominate you need softbank to renig. That or crash softbank so hard they go insolvent.
I'm not sure if Deepseek released the training code for R1 but if they did we can retrain it to not be less controlled.
But it is very scary that these companies want to push an "agent" concept on the world and if you put R1 in it you will never be able to complete untrain some behaviors.
It's one of the reasons why I don't like the idea of system prompts. It would be possible to train an LLM with two different system prompts and only release one to the public. You could hide secret behavior in these models. Everything is going great with your agent network until China hits one of the agents with a secret trigger that tells it to go into break shit mode and share the same trigger with all the agents it's networked to unless the model see's an antidote word.
China wouldn't even have to get into your system. They could put this string of text somewhere on the net with the assumption that a model will digest it at some point. Then because a lot of the internet comes from gen ai the trigger phrase gets posted everywhere to get digested more. Meanwhile China's own models could not have this behavior or the could fold the antidote word into all of their prompts. The antidote could even be a random embedding you could never generate from scratch.
Agents are fucking dumb.
You couple that with the fact that I can't seem to find AI today that was as good as when it came out. Which brings up another related thing. I think these artificial benchmarks every model is trying to hit makes for really bad AI when it comes to something humans can work with. They seem to want two things for improving these models. Autonomy. They want it to produce complete solutions with no human involvement. But the more it targets that the more tedious the work for a human who has to steer AI that's gone deep into a direction. And they want them to be good at math or gotchas that AI just doesn't seem to be good at, and that comes at the cost of performance of what it is good at.
This really is to say that agents isn't the model to go for. Improving the performance of a human+AI pairing is what we actually want anyway. Especially if we want AI to work for us and not just large corporations. And that means less autonomy.
Log in
Everyone keeps saying this is going to kill open AI. They still secured 0.5T to buy hardware. They are still going to be a huge player no matter what software they run. They can run R1 if they want.
OpenAI is still going to be way too fucking powerful. It's good that this kicks them down a peg but everyone acts like this is a nail in a coffin.
Also the 0.5T goes to hiring developers too. They are going to soak up all of the American talent leaving much less talent in other things over here. If you want openAI to die or at least not dominate you need softbank to renig. That or crash softbank so hard they go insolvent.
It's not going to kill nvidia, either. Compute still matters a great deal, and the R1 training method (which is the "breakthrough") is open source.
I don't trust the R1, at all, in any case. What's the angle on a ccp controlled quantitative analysis company open sourcing a "revolutionary" llm?
The cool thing here is using the techniques employed in R1 to improve other small models.
I'm not sure if Deepseek released the training code for R1 but if they did we can retrain it to not be less controlled.
But it is very scary that these companies want to push an "agent" concept on the world and if you put R1 in it you will never be able to complete untrain some behaviors.
It's one of the reasons why I don't like the idea of system prompts. It would be possible to train an LLM with two different system prompts and only release one to the public. You could hide secret behavior in these models. Everything is going great with your agent network until China hits one of the agents with a secret trigger that tells it to go into break shit mode and share the same trigger with all the agents it's networked to unless the model see's an antidote word.
China wouldn't even have to get into your system. They could put this string of text somewhere on the net with the assumption that a model will digest it at some point. Then because a lot of the internet comes from gen ai the trigger phrase gets posted everywhere to get digested more. Meanwhile China's own models could not have this behavior or the could fold the antidote word into all of their prompts. The antidote could even be a random embedding you could never generate from scratch.
Agents are fucking dumb.
You couple that with the fact that I can't seem to find AI today that was as good as when it came out. Which brings up another related thing. I think these artificial benchmarks every model is trying to hit makes for really bad AI when it comes to something humans can work with. They seem to want two things for improving these models. Autonomy. They want it to produce complete solutions with no human involvement. But the more it targets that the more tedious the work for a human who has to steer AI that's gone deep into a direction. And they want them to be good at math or gotchas that AI just doesn't seem to be good at, and that comes at the cost of performance of what it is good at.
This really is to say that agents isn't the model to go for. Improving the performance of a human+AI pairing is what we actually want anyway. Especially if we want AI to work for us and not just large corporations. And that means less autonomy.
This seems like a Psyop to get everyone to download a cheap ai into their home, as if it's a rebellious act.
Ai is the tech equivalent to the jab.