NOTEBOOK
My writing
I write about language models, agent systems, and the tools I build to test them.
Featured article
Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5
I compiled a RISC-V CPU into Qwen 3.5 4B's weights, using its Gated DeltaNet layers as RAM. It runs C and MicroPython inside the forward passes, and Qwen uses it for exact arithmetic.
Read article ↗My writing
6 entriesTelling identical GPUs apart with one line of PyTorch
Physically identical GPUs leave a stable fingerprint in the order their atomic adds land. A custom probe identifies 19 identical A40s at 99.1%, and a plain torch.index_add_ call leaks it too.
Building oversight tools into a model's weights
A lie detector and log compiled into Qwen's weights that the model can't switch off, and an answer key planted inside a real Qwen 3.5 to test interpretability tools.
Hybrid models can hide reasoning in their RNN state
Under answer-only training, OLMo-Hybrid learns to track state silently in its RNN layers and generalise past its training length. Capping Gated DeltaNet's write strength at 1, as Qwen 3.5 does, stops it.
Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5
I compiled a RISC-V CPU into Qwen 3.5 4B's weights, using its Gated DeltaNet layers as RAM. It runs C and MicroPython inside the forward passes, and Qwen uses it for exact arithmetic.
Consensus Is Not Corroboration
Why language models need to distinguish a manufactured echo from genuinely independent confirmation.
Attention Lab: Eight Ways a Model Looks
Interactive field guides to the attention mechanisms shaping modern language models.