📢 📢 Are 1.58-bit LLMs going to save the planet and destroy NVIDIA’s business?
Microsoft Research published a research paper:
👩🎓Introducing BitNet b1.58: 1.58-bit LLMs, in which every single parameter is ternary {-1, 0, 1}, that rival full-precision Transformer LLMs in performance while significantly boosting efficiency—paving the way for a new era of 1-bit LLMs.
👉 So why is this so important?
Well, instead of using a 16- or 32-bit model that needs huge GPU power and electricity, this 1.58-bit model performance is (nearly) equal to these power-hungry models but without the need for that GPU power.
🧙♂️ It sounds like magic and too good to be true, but if an established organization such as Microsoft Research is publishing it, we need to take this seriously.
This 1.58-bit LLM would enable AI to run on smartphones and other battery-charged Edge devices. Why? Because it would need a lot less computational power and wouldn’t drain your battery as much.
🤔 But wait a second.
🤖 Are these models already as good as some current LLMs? With AI advancements going at a rapid pace, would that mean that 1.58-bit models will largely replace the current 16 and 32-bit models?
👻 If that is the case, our current rising star, NVIDIA, has a problem. Its rise in company value is based on a growing need for GPU power to support the ever-growing AI models. NVIDIA will have to come up with new hardware supporting the 1.58-model and this could mean opportunities for competitors.
💚 With 1.58-bit models, there is no need for GPU power. And that would also be good news for the environment as much less electricity is needed.
You might wonder why it is called a 1.58-bit LLM? It comes from the mathematical equation:
👩🎓log(3)/log(2) = 1.5849625
A 1.58-bit LLM can have three values: -1, 0 , and 1. And that limited number of values makes it so energy efficient and no need for GPUs.
It is still early days for us to understand what impact this new technology will have on the further development of AI, but one thing is for sure: we are just getting started.