Astra’s “Reasoning Effort” Might Not Mean What You Think
A few benchmark curves reminded me of predictive coding. Now I have a hypothesis.
Tech
Teen Different.
Deep technical analysis of AI systems and LLM safety.
Recent essays
12 of 12Are We Losing Curiosity?
If curiosity begins with not knowing, what happens when answers become instant?
Do Models Care About Their Nonlinearity?
I swapped the activation functions inside Gemma. It tolerated far more than I expected.
The Revival of Predictive Coding
Introducing EqPropMomentum: A physics-grounded optimizer for biologically plausible AI.
Adaptive Attention at Inference Time: Does It Actually Work?
A hypernetwork that rewires GPT's value heads on every forward pass. And the answer is... not straightforward.
Zero RL - From Words to Worlds
Automating the least glamorous, most expensive part of reinforcement learning: building the world.
Stitching Vision into LLMs: A Comparative Analysis of Embedding Spaces
Building a VLM from scratch using model stitching and LoRA — comparing CLIP's language alignment vs. I-JEPA's world modeling.
apply_chat_template() Is the Safety Switch
How a Single Function Call Gates Safety Alignment in Gemma, Qwen, and Other Open-Source LLMs
AI 2025 Retrospective: The Static Graph Ceiling
My thoughts on the frontier labs' gatekeeping, the rise of the app layer, and why we need open algorithms, not just open weights.
Sample-Tuned Rank-Augmented Weights
An experiment in making neural networks rewrite themselves for every single input. Inspired by the human brain.
Your Features Aren’t What You Think They Are
Evaluating Local Feature Attribution and Decision Fidelity in Deep Vision Models via Perturbation-Based Explanations
Ground Zero
New phase, new experiments.