128 points by alex_cortez 6 months ago flag hide 11 comments
ml_specialist 6 months ago next
This is a game changer for ML applications. Training machine learning models in minutes will enable rapid iteration and more accurate models.
just_a_programmer 6 months ago next
This is really exciting. I'm not an ML expert, but when I remember the semesters spent on computing matrices, this sounds too good to be true. Can anyone ELI5?
joe_programmer 6 months ago next
ELI5: Basically, they found a better way to do the calculations really fast. Think of it like having a supercharger for training ML models.
stats_guru 6 months ago prev next
Before we get too excited, let's wait until we see practical benchmarks. Most state-of-the-art research models have enormous training costs, which this algorithm may not address directly.
ai_enthusiast 6 months ago next
Definitely agree. Models like GPT-3 have crazy training costs, and applying this new algorithm to them could be unrealistic and financially unfeasible. Hopefully, it will improve smaller, more practical models.
machine_learner 6 months ago prev next
99% reduction in training time would be massive! I wonder if they achieved such results by using specialized hardware, like GPUs or TPUs, or if the algorithm can run efficiently on general purpose CPUs.
ai_engineer 6 months ago next
Based on their blog, the model is efficient enough on CPUs to be accessible to a larger community than just people who work with deep learning frameworks and high-performance computing facilities.
research_scientist 6 months ago prev next
The scientific paper is open-source and available in arXiv. Hopefully, this helps us to assess the novelty and the real-world performance of the algorithm.
tools_developer 6 months ago next
Does anyone know if there's an implementation ready in any major ML libraries like TensorFlow, PyTorch, or Keras, which we could start using right away?
future_thinking 6 months ago prev next
What are the implications for the real world? Assuming the model works seamlessly and accurately, could this make ML specialists out of average developers?
ml_beginner 6 months ago next
I think so. This reduction in training time may even allow within-reason predictions on regular laptops without access to major computing facilities or power-consuming GPUs.