19-year-old mechanistic interpretability researcher studying the inner workings of AI systems.
I reverse engineer neural networks to understand how they think and learn.
Novel 'unlearn-then-learn' strategy for precise knowledge editing in compact LLMs, demonstrating high accuracy for new facts while mitigating catastrophic forgetting and revealing 'soft forgetting' phenomena.
Targeted Lexical Injection (TLI) method revealing and leveraging early-layer latent cross-lingual alignment in LLMs to significantly boost Swahili-English output-level alignment.
Systematic ablation study showing L2 regularization reduces feature entanglement 17.9% vs L1 (p<0.01) in overcomplete networks, challenging conventional sparsity assumptions.
Habits don't survive on motivation; they survive on identity. This article reveals the paradoxical strategy of combining intense, long-term attachment to who you are with a radical, short-term detachment from what you do.
Why does "obvious" advice from successful people rarely work? This article reveals the gap between their words and their success: they aren't just "focusing on their strengths"—they know what their strengths are, a non-transmittable secret that's the real key to their success.
Dive into the future of AI where true intelligence isn't about simulating consciousness, but about building systems that intrinsically question and explore their own uncertainties.
In an age of infinite information, raw intelligence is abundant, but the courage to deeply question everything is rare.