Nvidia's moment; Assyrian dog pack; hype & reality in contemporary GPT-LLM-ML; very briefly noted; & Golden Gate Lab Rescue, equilibrium-restoring economic forces AWOL, a low-pressure economy is a...
NVidia GPUs are an interesting market. The large scale ML use of high end GPUs can only have a limited market as few players can afford the scale needed for data and training. High end GPUs will not be suited to any consumer market for decades at the earliest. GPUs to accelerate graphics and smaller scale analytics, that is a more feasible growth market - declining prices driving volume. I would buy a better GPU for computational work, although the games market must be the main driver.
Is there a market for a high end GPU in every humanoid robot in the future?
Yes: this is the "training vs. inference" part. Nvidia's bull consumer case is that every smartphone acquires a local graphics box within bluetooth range (and over the net if you are out of the house) inside the privacy bubble of the user. Otherwise, it seems to rely heavily on a large demand for training models...
I am a bit old school. I like Pete Warden's approach of reducing the trained model to work on an smart phone by reducing the trained model size to be less granular, e.g. using fewer bits per weight, etc. This pushes the ML models out to the edge, using minimal CPUs and memory. Training needs to be done on beefier hardware, even using the cloud, but once done, the model can be ported to the smartphone and run on basic hardware. Models can also be run on a minimal Raspberry Pi and even Arduino hardware, potentially making AI much more ubiquitous.
I could see a good market for trained models that can be downloaded and use the smartphone camera and other potential sensors to run inference on data input. These might well appear as apps in the store to buy, or apps that work with specialist hardware to select a model and run it, perhaps even train the model with data. What I don't see is the user trying to train LLMs or other big data ML models. The training must be on limited data for the user, or using pre-trained models for applications like ChatGPT. Of course, running rules-based ML would be almost trivial by comparison if the target hardware would allow different languages to be run on it - very difficult with Apple's walled garden.
Your insight re 2016 is most perceptive and appreciated and dare I say invaluable for anyone writing the history of that awful outcome.
NVidia GPUs are an interesting market. The large scale ML use of high end GPUs can only have a limited market as few players can afford the scale needed for data and training. High end GPUs will not be suited to any consumer market for decades at the earliest. GPUs to accelerate graphics and smaller scale analytics, that is a more feasible growth market - declining prices driving volume. I would buy a better GPU for computational work, although the games market must be the main driver.
Is there a market for a high end GPU in every humanoid robot in the future?
Yes: this is the "training vs. inference" part. Nvidia's bull consumer case is that every smartphone acquires a local graphics box within bluetooth range (and over the net if you are out of the house) inside the privacy bubble of the user. Otherwise, it seems to rely heavily on a large demand for training models...
I am a bit old school. I like Pete Warden's approach of reducing the trained model to work on an smart phone by reducing the trained model size to be less granular, e.g. using fewer bits per weight, etc. This pushes the ML models out to the edge, using minimal CPUs and memory. Training needs to be done on beefier hardware, even using the cloud, but once done, the model can be ported to the smartphone and run on basic hardware. Models can also be run on a minimal Raspberry Pi and even Arduino hardware, potentially making AI much more ubiquitous.
I could see a good market for trained models that can be downloaded and use the smartphone camera and other potential sensors to run inference on data input. These might well appear as apps in the store to buy, or apps that work with specialist hardware to select a model and run it, perhaps even train the model with data. What I don't see is the user trying to train LLMs or other big data ML models. The training must be on limited data for the user, or using pre-trained models for applications like ChatGPT. Of course, running rules-based ML would be almost trivial by comparison if the target hardware would allow different languages to be run on it - very difficult with Apple's walled garden.
Does anybody know anything about Apple's currently-in-beta compressed transformer autocorrect, and how it runs on iPhones?
"Cry “Havoc!”:"
I'm disappointed. No "Dances With Wolves."?