Vlad Larin

Yup, really believe that this is a way

10.12.2024 22:40 — 👍 1    🔁 0    💬 0    📌 0

Without this compass, a well-crafted prompt could steer the AI into actions that look safe on the surface yet ultimately harmful. By having a continuous human empathy stream in the loop, the model starts to worry about these subtleties, sensing when something might undermine human well-being.

10.12.2024 18:47 — 👍 2    🔁 0    💬 1    📌 0

This dynamic input ensures that each next-generation step leans into compassion and alignment — enabling the AI to reflect on the context rather than just effectively “erase weights” for potentially harmful prompts.

10.12.2024 18:46 — 👍 3    🔁 0    💬 1    📌 0

The mechanism? Live RLHF. Consider a reward model that continuously evaluates ongoing and a number of predicted tokens. By embedding its judgments back into the main model’s pipeline, we give the AI a running commentary of human intent.

10.12.2024 18:45 — 👍 0    🔁 0    💬 1    📌 0

Abstract human-to-AI interaction

I’m starting to think that reflective human empathy might be key to genuine AI alignment. Instead of just filtering harmful outputs post-hoc, what if we integrate direct human feedback into the model’s reasoning process?

10.12.2024 18:45 — 👍 7    🔁 0    💬 1    📌 0

Humans learn sequentially — each experience shaping the next. AI learns in parallel, absorbing vast knowledge simultaneously. This fundamental difference creates entirely new forms of cognition. Could AI’s parallel consciousness unveil truths about reality that transcend human perception?

01.12.2024 06:07 — 👍 32    🔁 4    💬 3    📌 1

Hi #Bluesky Really interested to see how it goes, glad to see how much attention the idea of real content ownership and full customization gets 🚀

01.12.2024 05:45 — 👍 7    🔁 0    💬 0    📌 0