Professor Jason Eshraghian - Neuromorphic LLMs
Event details
| Date | 10.03.2026 |
| Hour | 11:15 › 12:15 |
| Speaker | Professor Jason Eshraghian - UCSC Neuromorphic Computing Group |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
This talk explores the collision between neuromorphic computing and large language models. I will walk through our lab's journey building frontier reasoning LLMs, from software to neuromorphic hardware, covering the wins, the losses, and everything in between. The path runs through spiking LLMs (SpikeGPT), MatMul-free LLMs on Loihi 2, and culminates in our latent reasoning model, Ouro, trained end-to-end on a full production pipeline spanning data curation, pre-training, post-training, and alignment. At roughly 2 billion parameters, Ouro outperforms every leading model from Meta, Google, and the rest of the over-resourced crowd in the ~10B regime, punching about 5x above its weight. We've deployed several of these models on neuromorphic hardware at 2 watts, bringing state-of-the-art reasoning from the datacenter to the edge.
Practical information
- Informed public
- Free
Organizer
- Professor Wulfram Gerstner
Contact
- Elke Hewitt