Hi Amit, we used split the global AI compute demand into Training and and Inference;
Recap (old view):
Training = the training/creation of AI models.
Inference = running/applying created AI models.
It turns out that, as we reach the limits of available data, that there are additional venues for improving AI model performance. One such increasingly growing venue is post-training. It refers to training models on themselves. Eg. if you trained a model on playing a game, you can make it better by letting it play against itself. The post-training bucket also includes synthetic training data generation. This has been around for a while but is gaining importance. For these reasons, people (like Jensen) started to break training down into two buckets: pre-training and post-training.
So the new view is:
Training = the training/creation of AI models.
Post-training = model-on-model and synthetic data
Inference = running/applying created AI models.
I tried to keep it simple and not go into the technical minutiae.
Regarding AMD/ASICs I'll cover this in detail in my upcoming research note.
SUMMARY OF FED MEETING MINUTES (5/22/24):
1. Various offices mention raising rates if inflation warrants it
2. Fed officials note “disappointing” inflation data in Q1 2024
3. Officials discussed holding rates higher for longer
4. Some officials worry that financial conditions are not sufficiently restrictive
5. Some officials think long run rates need to be higher
6. Many officials commented on uncertainty about degree of policy restrictiveness
Higher for longer is here to stay.