Blog

Notes from my research on model internals, efficient GenAI, and hardware-level failure modes. I write when a result or engineering lesson deserves more room than a paper abstract allows.

Latest

What Bit-Flip Attacks Reveal About Multimodal LLMs

Five to seven targeted bit flips were enough to collapse several foundation models. I unpack how FlipLLM finds those bits, what the multimodal result means, and why targeted protection is the more useful outcome.

Read the article