Goodfire's Matthew Kowal to talk interpretability across the model design pipeline at Cohere Labs
Cohere_Labs · x · 2026-10-22
Cohere Labs' AI Safety and Alignment group is hosting Matthew Kowal, MTS at Goodfire, for an October 23rd talk titled "Interpretability Across the Model Design Pipeline."
Key points from the abstract:
- Interpretability framed through model affordances: the actions a system makes possible, giving researchers and engineers new ways to intervene throughout the design pipeline.
- Recent methods for building better "knobs" to understand and control model behavior by decomposing activations into multidimensional manifolds rather than straight lines.
- Ongoing work using these interpretable knobs during training itself, leveraging a model's own representations to shape learning while keeping the process interpretable and monitorable.
More from Research
- UW Baker Lab and NVIDIA launch RFD4-Proteina, a 13x faster protein design model — AllThingsApx · 2026-10-22
- NVIDIA opens RFD4-Proteina BioNeMo model for 13x faster protein design — AllThingsApx · 2026-10-22
- Pinal on Nature: Designing Functional Proteins Directly from Natural Language — bravo_abad · 2026-10-22
- 136-Country, 118M-Article Dataset Audits State Media Share in LLM Training Corpora — deliprao · 2026-10-22
- PyTorchCon talk: QR, eigendecomposition and Cholesky kernels decide if better optimizers actually train faster — PyTorch · 2026-10-22
- Multi-agent LLM robot search: 6 agents cut search time 2x, theory bound within 3-7x of observed — peter_richtarik · 2026-10-22