What makes an AI model good in 2026 is no longer its architecture. It's the part the lab won't publish. Everyone converged on the same linear hybrid, so the edge moved to data, RL, and tool use. My new piece: byin-cwi.github.io/MatrixWeb… #LLM

I updated my personal website: byin-cwi.github.io/MatrixWeb… I’ll use it to share research notes and personal reflections on learning algorithms, brain-inspired AI, LLMs, agents, scaling laws, memory, and neural network dynamics. First posts are already online. More to come.

We never really knew how to train nonlinear RNNs well… BPTT struggled with vanishing grads (no long-range memory) and sequential rollout (hard to parallelizable). What if instead an oracle told us the optimal memory state m_t at each step? Then the RNN could do one-step supervised learning on (m_t, x_{t 1}) → m_{t 1} labels. We call this Supervised Memory Training (SMT): a replacement for BPTT that trains RNNs without unrolling them. SMT is time-parallelizable and solves vanishing gradients. Website: akarshkumar.com/smt/ arXiv: arxiv.org/abs/2606.06479

How do beta and gamma rhythms impact synaptic integration in pyramidal neurons? Our latest paper that examines this question studied this using a biophysically detailed pyramidal neuron model. Here’s what we found. A 🧵below. elifesciences.org/articles/9… 1/10

Our NeuroAI study made it onto the cover of Nature Machine Intelligence (@NatMachIntell) ❤️. In it, we demonstrate that a developmentally-inspired visual diet can drastically improve the robustness of ANN-based vision systems. open access, open code, open weights, open science

Be the scientist, not just the researcher. AI has already replaced much of research labor. BUT it hasn’t replaced taste, judgment, courage, or the ability to ask the question that changes everything.

The "model weights are unreadable" excuse just died in one paper. Neural networks have billions of parameters. Nobody really knows what each one does. A new paper introduces adVersarial Parameter Decomposition. The method splits a model's weights into small, single-purpose subcomponents. Each piece handles one specific job. Things like emoticon prediction or gender identification. Only a tiny fraction fires on any given input. How they pulled this off: 1. Break weights into simple parts 2. Ablate components adversarially during training 3. Keep only what preserves behavior The breakthrough is attention. The technique decomposes attention computations even when spread across multiple heads. That problem went unsolved for three years. It scales beyond toy networks to real four-layer language models. The authors now consider it a serious competitor to sparse autoencoders. You can trace information flow through attribution graphs. You can hand-edit specific behaviors and predict the outcome. VPD makes weights legible.

The entire RAG stack may be missing one layer. Vector RAG retrieves similar chunks. PageIndex @PageIndexAI retrieves document tree nodes. PageRankRAG retrieves connected evidence paths. I built a graph propagation layer on top of PageIndex-style document trees.

Human brain & AI Workshop at @IJCAIconf in August 2026, Bremen, Germany hbai2026.github.io Organized by@chen_guozhang and colleagues. Join to hear from @eckstein_maria @AToliasLab @NeuroNaud @Pieters_Tweet and others! Submit your papers by May 15 openreview.net/group?id=ijca…

🚀 OmniMouse is Accepted at ICLR 2026: Poster Presentation in Rio on Friday, April 24, 3:15PM, P3-#1808 📝 Blog: enigma-brain.github.io/omnim… 💻 Code: github.com/enigma-brain/omni… 🤗 Model: hf.co/the-enigma-project/omn… 🤗 Data: hf.co/datasets/the-enigma-pr… 📄 Paper: arxiv.org/abs/2604.18827 1/

The geometry of neural dynamics along the cortical attractor landscape directly reflects changes in attention, as large-scale brain activity shifts across its hills and valleys depending on the state. nature.com/articles/s41467-0…

境由心现，识依缘起。

📣 Submit your work to HBAI 2026! The 5th International Workshop on Human Brain and Artificial Intelligence, co-located with IJCAI 2026, invites papers on the frontier of AI and brain research. 🗓 Deadline: May 15 🔗 hbai2026.github.io #NeuroAI #BrainInspiredAI #IJCAI2026

Preprint out arguing that we should build the techology to translate (compile) molecularly annotated connectomes into dynamics. I think this is incredibly important. arxiv.org/abs/2603.25713

The weather's looking good for tomorrow's Artemis II launch, and our teams are getting the rocket ready for liftoff! Read the latest updates on our mission around the Moon: go.nasa.gov/4tiFY4P

Coo000l project, it deserves 10k stars.