Goodfire maps the geometry inside neural networks: color wheels, time curves, and more
burny_tech · x · 2026-10-20
Goodfire (Atticus Geiger, Ekdeep Singh Lubana, Tom McGrath, et al.) published a long-form essay arguing that neural network internals mirror the structured geometry of the world. Key evidence: in LLMs, numbers, weekdays and months form circular loops; years and characters trace smooth curves. In image models, objects' spatial layout is recapitulated in activation space and colors lie on hue/saturation/lightness surfaces. In a genomic model, the tree of life sits on a manifold, and an unknown class of Alzheimer's biomarkers appeared as a clean curve. The authors argue neural geometry is the key to understanding and controlling AI.
More from Research
- Two New Robotics Benchmarks Released: MicroPlanBench and MicroAnnotationBench First Rankings — chris_j_paxton · 2026-10-20
- ICLR reviewer despairs: 5 assigned papers nearly all AI-written, calls peer review broken — peterbhase · 2026-10-20
- Meta: Co-Design for AMD GPUs Delivers Up to 1.5x Training Throughput vs Direct NVIDIA Ports — burny_tech · 2026-10-20
- Continuous Softmax Attention work shared, drawing attention — burny_tech · 2026-10-20
- Muon optimizer tweak with a polynomial grid saves ~0.5% of LLM training compute — burny_tech · 2026-10-20
- HeadEdit steers LLM behavior via frozen unembedding matrix, no fine-tuning needed — burny_tech · 2026-10-20