rajap
Stability AI and EA partner to reimagine game development
RL algorithms are less bitter-lesson-pilled than 2015-era deep learning
The real issue isn't reward shaping or curriculum learning - everyone complains about those. The deeper problem is that we're hardcoding the credit assignment timescale into our algorithms.Discount factors (γ), n-step returns, GAE λ parameters - these are human priors about temporal abstraction baked directly into the learning signal. PPO's GAE(λ) literally tells the algorithm "here's how far...
Disconnect between AI capability and public perception thereof
I am increasingly concerned about the disconnect between the capabilities of AI and the public perception thereof. While GPT-5 has closed that gap somewhat by providing free users with much better reasoning models at least a fraction of the time, I’m still hearing the same stories over and over again from across the societal spectrum: AI is overhyped. Agents are not a real thing. Nobody is...
Re:MIT Study Finds AI Use Reprograms the Brain, Leading to Cognitive D...
on Wednesday September 03, 2025 @02:06PM (#45118770)
with proper testing you can make suer that given A the returned value is B