thanks man its a good thread. i feel like people dont explain PPO very well in these threads, thats the main part that always feels handwavy. i’d also like an idea of order of magnitude of finetune data needed for instructgpt but havent seen good numbers
RESEARCH
-
Simulating Trump vs ordinary people: AI capability differences
By
–
Like he says ppl like Trump would be easy to simulate while I say normal people are easy to simulate
-
The Evolution of AI: From Artificial to Apparent Intelligence
By
–
The evolution of AI:
2000s: Artificial intelligence
2010s: Augmented intelligence
2020s: Apparent intelligence -
UW Leading the Way in Artificial Woke Intelligence
By
–
Proud that UW is taking the lead on AWI (Artificial Woke Intelligence).
-
94.5% Accuracy Fashion MNIST Achievement in Five Epochs
By
–
94.5% on Fashion MNIST in 5 epochs is pretty amazing!
-
Bypassing AI Detection: Natural Text and Watermarking Constraints
By
–
I'm not sure about that. E.g. tricking the ZeroGPT algo means actually writing more natural text (so it'll be better, not worse), and tricking the watermarking means not being limited by the constraints of the watermarking algo.
-
Accessible Explanation of Diffusion Models Through Physical Ink Process
By
–
I love the accessible explanation of diffusion models through the physical process of ink diffusing through water and following that trajectory and reversing it.
-
Cognition in LLMs: Associative Memory vs. Functional Awareness
By
–
That’s a good point. What humans see as cognition in LLMs is often just good associative memory, or (generously) intuition. A model that can echo “I am aware” is not necessarily, but in larger LLMs more of the functional implications of awareness begin to apply.