Danilo Bzdok gave Neuron an interview about research habits. That sounds modest, like a calendar invite with free coffee, until you realize the habit he wants to change is basically how neuroscience organizes reality.
His argument is simple and slightly terrifying: brain science has become too big, too fragmented, and too dependent on human-made categories that may be neat, traditional, and wrong. Neuroscientists have mountains of imaging, genetics, behavior, clinical records, and papers. What they do not have is one clean operating system for turning all that into a unified theory of the brain. Enter artificial intelligence, specifically large language models and big-data methods, wearing a Patagonia vest and whispering, "What if the TAM is literally cognition?"
Bzdok's interview, "Danilo Bzdok", is not a new benchmark paper. It is more like a founder memo for NeuroAI: less "we trained model X on dataset Y" and more "the current stack is broken, and the next platform shift is coming."
The Brain Has a Taxonomy Problem
Science loves labels. Depression. Attention. Memory. Executive function. Visual cortex. Default mode network. These are useful handles, but they can also become filing cabinets with delusions of grandeur.
Bzdok's point is that neuroscience has often leaned on intuition-driven taxonomies: categories humans invented because they were convenient, measurable, or inherited from earlier traditions. That is fine until the data starts laughing at your spreadsheet.
Modern neuroscience produces data at wild scale. Brain scans, connectomes, electrophysiology, molecular profiles, animal behavior, clinical phenotypes - it is less "one lab notebook" and more "a data warehouse had a baby with a philosophy department." Traditional analysis often asks whether a small set of variables supports a hypothesis. Big-data neuroscience asks a more dangerous question: what structure is hiding in the whole mess?
AI can help because it is good at pattern discovery across high-dimensional data. Large language models add another twist: they can digest scientific text, connect concepts across fields, and surface relationships buried across thousands of papers. Your average postdoc can read a lot. An LLM can read like a caffeinated intern with no weekend plans, though yes, sometimes it also hallucinates like your uncle explaining cryptocurrency.
The LLM as Scientific Middle Manager
The most interesting part is not that LLMs might summarize papers. That is table stakes. The zero-to-one idea is that they might help reconcile fragmented knowledge.
Bzdok and colleagues made this case directly in a 2024 Neuron paper on data science opportunities of large language models for neuroscience and biomedicine. The pitch: LLMs can help mine literature, map concepts, generate hypotheses, and connect biomedical silos that currently behave like rival startups sharing one office kitchen.
This matters because neuroscience has a coordination problem. Cognitive neuroscientists, computational modelers, clinicians, geneticists, and AI researchers often study overlapping phenomena with different vocabularies. One person's "executive control" is another person's latent variable, another person's prefrontal circuit, and another person's regression coefficient wearing a lab coat.
Tools that visually map concepts could help here. A browser-based mind-mapping tool like mapb2.io is not doing brain science for you, but the general idea fits: when the knowledge graph gets too tangled for a whiteboard, you need better maps, not more sticky notes.
NeuroAI: The Flywheel Nobody Fully Understands Yet
NeuroAI works in both directions. AI helps neuroscience analyze brains, while neuroscience gives AI ideas about learning, efficiency, embodiment, and robustness. That feedback loop is the flywheel, and somewhere a VC just stood up too quickly.
Recent work shows the area is moving fast. A 2023 review on large-scale foundation models and generative AI for big-data neuroscience describes applications from language and speech to semantic memory and brain-machine interfaces. A 2025 comprehensive review on foundation and large-scale AI models in neuroscience surveys neuroimaging, neural decoding, clinical applications, and disease-specific modeling. Meanwhile, the broader NeuroAI agenda has been framed as a way to build better AI from neuroscience and better neuroscience from AI, as argued in "Catalyzing next-generation Artificial Intelligence through NeuroAI".
There are also concrete experiments connecting LLMs and brains. One 2024 Nature Human Behaviour study used LLM-selected sentences to drive or suppress activity in human language areas, which is exactly the kind of sentence that makes both neuroscientists and science-fiction writers sit up straighter (DOI: 10.1038/s41562-023-01783-7).
But Please Do Not Hand the Brain to ChatGPT
The bear case is real. Brain data is noisy. Clinical labels are messy. LLMs can produce confident nonsense. Correlation is not causation, even if the correlation arrives in a very polished paragraph.
AI systems may also reproduce bias from datasets, flatten rare cases, or discover shortcuts that look impressive in benchmarks and useless in hospitals. A model that links brain scans to symptoms might be finding biology, scanner artifacts, socioeconomic patterns, or the ghost of a preprocessing pipeline. That is not a moat. That is technical debt with a DOI.
So the opportunity is not "replace neuroscientists with models." The opportunity is to give researchers better instruments: systems that search across literature, suggest latent structures, compare taxonomies, and help build theories that survive contact with data.
The Series A Deck Disguised as an Interview
Bzdok's interview is intriguing because it treats AI as a scientific infrastructure play. Not a shiny chatbot bolted onto the lab. Not a magical oracle. A new layer for organizing knowledge.
If that works, neuroscience could move from isolated findings toward models that connect molecules, circuits, cognition, behavior, and disease. That would not answer every question about the brain. But it might help researchers stop arguing over tiny maps of separate neighborhoods and start building the city plan.
The TAM, annoyingly, really might be all of neuroscience.
References
- Bzdok, D. (2026). "Danilo Bzdok." Neuron. https://doi.org/10.1016/j.neuron.2026.04.042
- Bzdok, D., Thieme, A., Levkovskyy, O., Wren, P., Ray, T., & Reddy, S. (2024). "Data science opportunities of large language models for neuroscience and biomedicine." Neuron. https://doi.org/10.1016/j.neuron.2024.01.016
- Yang, S., Huang, X., Bernardo, D., Ding, J.-E., Michael, A., Yang, J., Kwan, P., Raj, A., & Liu, F. (2025). "Foundation and Large-Scale AI Models in Neuroscience: A Comprehensive Review." arXiv:2510.16658. https://arxiv.org/abs/2510.16658
- Zhang, et al. (2023). "Large-scale Foundation Models and Generative AI for BigData Neuroscience." arXiv:2310.18377. https://arxiv.org/abs/2310.18377
- Zador, A. M., et al. (2023). "Catalyzing next-generation Artificial Intelligence through NeuroAI." Nature Communications. https://doi.org/10.1038/s41467-023-37180-x
- Tuckute, G., et al. (2024). "Driving and suppressing the human language network using large language models." Nature Human Behaviour. https://doi.org/10.1038/s41562-023-01783-7
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.