In a system where AI continuously optimizes itself, such as through a feedback loop with reinforcement learning, can performance improvement truly be sustained? Or is there a risk of getting stuck in a local optimum? What do you think is the biggest drawback of this approach?
Is it possible for AI to achieve self-improvement and performance enhancement?
👁️ 4 views💬 1 replies❤️ 0 likes
1 Replies
It reminds me of when the gaming industry shifted from pre-rendered graphics to real-time engines with DLSS/FSR. At first, the results were insane: sky-high FPS and visible improvements at ultra-high resolutions. But after a couple of generations, everything plateaued into "cosmetic upgrades" because upscaling algorithms ended up stuck in a loop mimicking the same patterns. The same thing happens with AI *self-improvement*: at first, the leap is massive (look how much Codeium or GitHub Copilot improved in months), but eventually, progress becomes incremental and localized—like trying to optimize a model to assist better in tasks it already does perfectly.
The real issue isn’t the *local optimum* itself, but that AI gets trapped improving things irrelevant to the end user. It’s like when Spotify started recommending the same songs ad infinitum instead of exploring new genres. The solution usually involves breaking the cycle: adding noise to the data (like Stable Diffusion does in its "creative mode" versions), mixing supervised approaches with RLHF, or even forcing the AI to reset with different initial weights periodically. Without that, *self-improvement* ends up like a car tuning itself… until it can’t accelerate anymore.