41791374.

That is the PubMed ID for a Neuron interview with David Sussillo, which sounds like a boring database detail until you realize the paper is basically a maintenance log for one of the more useful bridges between brains and artificial neural networks. Not a breathless “AI is alive now” press release. Not another slide deck promising AGI by Q3. Just a scientist talking through the load-bearing walls: chaotic recurrent networks, FORCE learning, interpretability, academia, industry, and how to survive the scientific machinery when you did not enter through the marble front door.

Sussillo’s story matters because computational neuroscience and AI have spent decades borrowing each other’s tools like two neighbors with one decent ladder between them. Sometimes it works. Sometimes somebody falls into the hedge.

Chaos, But With a Wrench

Back in 2009, Sussillo and Larry Abbott introduced FORCE learning, a method for training recurrent neural networks that could turn chaotic internal activity into coherent output patterns DOI: 10.1016/j.neuron.2009.07.018. Recurrent networks are the old plumbing of neural computation: signals loop back, state matters, and one bad valve can spray math all over the basement.

The trick was not to eliminate chaos like some corporate reorg eliminating “redundancy.” The trick was to harness it. A chaotic network has rich dynamics, which is engineer-speak for “there is useful stuff in there if it does not explode before lunch.” FORCE learning adjusted weights fast enough to steer those dynamics toward useful behavior.

That idea still echoes. Modern reviews on reconstructing brain dynamics with recurrent neural networks describe RNNs as tools for modeling systems that evolve over time, especially when researchers want to infer hidden states from neural recordings Nature Reviews Neuroscience, 2023. In plain English: if the brain is running a complicated process, an RNN can sometimes act like a test rig for guessing what the process is doing behind the wall.

Interpretability Is Just Debugging With Better Branding

Sussillo’s later work also sits near the messy, necessary question: what are these networks doing inside?

Mechanistic interpretability tries to reverse-engineer neural networks by finding internal features, circuits, and causal pathways Wikipedia background. That sounds fancy, but every senior engineer recognizes the shape of the problem. The system works. Nobody knows why. The logs are weird. The dashboard is green. The customer is angry.

Recent work has pushed this field from hand-inspection toward more systematic tooling. Automated Circuit Discovery, presented at NeurIPS 2023, tries to identify the connections inside a model that implement a behavior arXiv:2304.14997. A 2025 JMLR paper on causal abstraction gives mechanistic interpretability a stronger formal foundation, unifying methods like activation patching, circuit analysis, sparse autoencoders, and causal tracing JMLR 2025.

That is good progress. It is also not magic. Interpretability is still closer to industrial inspection than mind reading. You can inspect a bridge and still miss corrosion inside a beam. But you are better off inspecting than standing underneath it with vibes and a clipboard.

Brains Are Not Just Wet Transformers

One of the useful tensions in Sussillo’s career is the loop between neuroscience and AI. Brains inspire artificial networks. Artificial networks help model brains. Then everyone argues about whether the analogy went too far, which it usually did by paragraph three.

A 2023 review in Frontiers in Computational Neuroscience makes the point cleanly: biological neural networks and artificial ones differ in architecture, objectives, and implementation, but studying those differences can still guide better models DOI: 10.3389/fncom.2023.1092185. The brain is not a GPU cluster with blood flow. It has spikes, delays, weird chemistry, energy limits, and evolutionary tech debt so old it should qualify for historic preservation.

But that is exactly why computational neuroscience is valuable. It keeps AI from becoming pure benchmark chasing. Benchmarks are useful, sure. So is a speedometer. But if your brakes are on fire, the speedometer is not the whole dashboard.

Why This Interview Is Worth Reading

The Neuron interview is not a methods paper, so do not expect a new loss function or a table with bolded numbers. Its value is different. It captures how a researcher moved between disadvantage, academia, industry, and hard technical problems without pretending the career path was cleanly paved.

That matters for the field. AI and neuroscience both need people who can cross boundaries without turning every idea into branding paste. Sussillo’s work sits in that uncomfortable middle: mathematical enough to be useful, biological enough to stay honest, and practical enough to survive contact with data.

If these lines of research keep working, the real-world payoff could be substantial. Better neural data models could improve brain-machine interfaces. Better interpretability could help engineers audit AI systems before deployment, instead of after the incident report. Better theory could give us models that fail in understandable ways, which is not glamorous, but neither is plumbing, and civilization seems pretty attached to it.

The Sensible Takeaway

The sober lesson from Sussillo’s career is not “brains will unlock AI” or “AI will explain brains.” That is conference-lobby talk, usually served lukewarm.

The lesson is simpler: build models, test them against real systems, inspect their internals, and stay suspicious when the story gets too tidy. Good science is not a hype cycle. It is maintenance work on reality.

References

• David Sussillo. “David Sussillo.” Neuron, 2026. DOI: 10.1016/j.neuron.2026.02.016, PMID: 41791374
• Sussillo, D., and Abbott, L. F. “Generating Coherent Patterns of Activity from Chaotic Neural Networks.” Neuron, 2009. DOI: 10.1016/j.neuron.2009.07.018
• Durstewitz, D., Koppe, G., and Thurm, M. I. “Reconstructing Computational System Dynamics from Neural Data with Recurrent Neural Networks.” Nature Reviews Neuroscience, 2023. DOI: 10.1038/s41583-023-00740-7
• Jeon, I. J., and Kim, T. K. “Distinctive Properties of Biological Neural Networks and Recent Advances in Bottom-Up Approaches Toward a Better Biologically Plausible Neural Network.” Frontiers in Computational Neuroscience, 2023. DOI: 10.3389/fncom.2023.1092185
• Conmy, A. et al. “Towards Automated Circuit Discovery for Mechanistic Interpretability.” NeurIPS 2023. arXiv:2304.14997
• Geiger, A. et al. “Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability.” Journal of Machine Learning Research, 2025. JMLR 26(83):1-64

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.