I am not going to be drawn into the AI Doom conversation. @perrymetzger has been making many of the points I would; check out some of his threads if you want another perspective.
@id_aa_carmack
-
Academia as the Jedi Order: Academic Lineage and Mentorship
By
–
Thinking about academia as the Jedi Order: @woj_zaremba Padawan of @ylecun
, Padawan of @geoffreyhinton -
Quest System-Level Voice Filter Options and AI Models
By
–
Quest should probably have system level voice filter options. Frequency shifting and equalization would be straightforward, but it could eventually go to full AI models. Care would be needed to make sure it doesn’t introduce a frame of latency.
-
CUDA Graph Capture and PyTorch Integration Performance Benefits
By
–
CUDA graph capture is like old school OpenGL display lists, but clearly specified as just kernels operating on already allocated buffers. The PyTorch integration is very well done; if you can be 100% graphs, the python overhead basically vanishes.
-
Specialized AI applications beyond graphics processor implementations
By
–
I haven’t used any processor in memory implementations, but I can imagine more direct uses for specialized AI than graphics work.
-
Government oversight of large-scale AI training compared to rocket regulations
By
–
The comment “we think it’s important that major world governments have insight about training runs above a certain scale” makes me recall the 200,000 lb-sec FAA amateur rocketry limit.
-
Criticism of Greek letters in algorithm naming conventions
By
–
Being cute and putting Greek letters in your algorithm name is not appreciated. Looking at you, Maximal Update Parameteization / (µP)!
-
Non-linear image distortion for Gaussian pixel statistics in ML models
By
–
I wonder if there would be any benefit for ML models to non linearly distorting image pixel statistics into a true Gaussian distribution instead of just the traditional scale and bias.
-
DGX Cluster Commitment: Long-term Investment vs Hourly Purchase
By
–
That was my argument to buy the DGX in the first place, but a committed cluster is a 6+ month deal, not an hourly purchase.
-
Regret over NVLink choice versus conventional PCI A100 setup
By
–
In hindsight, I would have been much better off with a more conventional system with PCI A100 cards — I still haven’t done anything dramatic with the extra NVLink bandwidth, and it could have had twice the GPUs.