It was a fun exercise recently to just open up a completely blank file and write an RL agent from scratch, without looking at any of my prior code. There is a point of scale where rewriting things from scratch is a bad idea, but it is a blessing when you can! By “from scratch” I
@id_aa_carmack
-
Running Convolutions in Frequency Space Like a Federation Science Officer
By
–
It is a trope that TV technobabble usually involves “reversing the frequencies” or “modulating the frequencies”, so I feel like a federation science officer when I declare “I am running the convolutions in frequency space!”
-
Amdahl’s Law: Limits of Parallel Computing and Sublinear Speedup
By
–
Amdahl’s law gets you there — you can linearly scale throughput on independent tasks, or scale up problems to utilize more parallel resources, but scaling a fixed task across parallel resources always has sublinear speed up. Also significantly more work.
-
GPU Count Power of Two for Task Distribution Optimization
By
–
It actually does matter here — number of GPUs in a system is almost always a power of two, and to maximize your utilization you will want that to evenly divide into the number of tasks.
-
Modern Recommendation Systems: Reinforcement Learning vs Single-Step Transactions
By
–
Are modern recommendation systems treated like a reinforcement learning problem, with a sum of discounted future rewards, or as strictly single step transactions? Many products do significant offline data analysis on actions taken to inform changes, but it seems under
-
Instagram’s AI 3D Photo Integration Beta Impresses with Seamless Stereoscopic Conversion
By
–
The beta 3D photo integration with Instagram is very well done! Every static photo becomes an AI generated stereoscopic 3D photo, and there is a “3D” button that lets you toggle the feature on and off for comparison. Every photo I looked at “just worked”, with no glaring
-
Old AI prediction from 2019 aging poorly after five years
By
–
This old post that I responded to popped back into my mentions today. Made in 2019, predicting 2029. More than halfway there, doesn’t fare well at all. Decadal AI doom predictions will fare similarly.
-
Challenges in Multi-Task Learning and Transfer Learning for Atari
By
–
I agree that chasing score on an Atari game has very limited value, but continuous online multi task learning with sparse rewards and transfer learning is definitely not solved for environments even as simple as Atari. Do you not think those are problems worth grappling with?
-
Upper Bound 2025 Talk Slides and Research Notes Available
By
–
The full video of my Upper Bound 2025 talk about our research directions should be available at some point, but here are my slides: https://
docs.google.com/presentation/d
/1GmGe9ref1nxEX_ekDuJXhildpWGhLEYBMeXCclVECek/edit
… And here are the notes I made while preparing, which are more extensive than what I had time to say: -
Latency Impact on Game AI Performance Algorithms
By
–
On emulated games. Adding real world latencies and uncertainties actually breaks many of the high performance algorithms that condition anything on the taken action having an immediate effect. The real research focus is around long term continuous multitask online learning, but