# Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

Page: https://stenobird.com/podcast/the-cognitive-revolution/intelligence-on-the-edge-liquid-ai-s-ramin-hasani-on-the-search-for-device-native-foundation-models
Text version: https://stenobird.com/podcast/the-cognitive-revolution/intelligence-on-the-edge-liquid-ai-s-ramin-hasani-on-the-search-for-device-native-foundation-models.md
Podcast: ["The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis](https://stenobird.com/podcast/the-cognitive-revolution)
Published: 2026-07-04T06:14:22+00:00
Episode link: https://www.cognitiverevolution.ai/intelligence-on-the-edge-liquid-ai-s-ramin-hasani-on-the-search-for-device-native-foundation-models/
Audio file: https://pdst.fm/e/mgln.ai/e/1113/pscrb.fm/rss/p/traffic.megaphone.fm/RINTP7587791262.mp3
Processing state: processed
JSON: https://stenobird.com/v1/public/podcasts/the-cognitive-revolution/episodes/intelligence-on-the-edge-liquid-ai-s-ramin-hasani-on-the-search-for-device-native-foundation-models
Duration seconds: 6478

## Resource

Liquid AI CEO Ramin Hasani argues that the future of AI lies in efficient, hardware-aware architectures rather than just massive scale. The discussion explores how liquid neural networks can bring high-performance intelligence to edge devices like phones, cars, and wearables.

## Highlights
- Main idea: Scaling parameters alone is not the only path to intelligence; efficient, biologically-inspired architectures can outperform massive models on edge hardware
- Practical takeaway: Using automated foundation model design (AFMD) allows for optimizing neural networks specifically for target hardware like NPUs and mobile chips
- Technical insight: Input-dependent gating mechanisms, such as those in Liquid S4, allow models to adapt transformations based on real-time data inputs
- Failure mode: Relying on proxy metrics during model design can lead to architectures that perform poorly on actual downstream tasks and target hardware
- Market opportunity: The massive demand for local, private, and low-latency inference on consumer electronics presents a larger market than centralized data centers

## Topics

Liquid Neural Networks, Edge AI, Foundation Models, Machine Learning Architecture, Hardware-Aware Design, Neural Dynamics, On-device Intelligence, SSMs

## Chapters
- 1:00 — Liquid AI's Empirical Approach: How Liquid AI applies a neutral, empirical approach to designing networks for constrained environments like Shopify and Mercedes-Benz.
- 9:00 — Biological Origins of Liquid Networks: Exploring the first-principles approach of modeling how neurons exchange information, inspired by the brain of worms.
- 17:00 — Modeling Membrane Potential: The mathematical foundations of modeling ion propagation and neural dynamics in continuous-time systems.
- 25:00 — Neurons vs. Parameters: A deep dive into why counting neurons may be a more meaningful metric for efficiency than simply counting parameters.
- 34:00 — The Limits of Unstructured Scaling: Discussing the challenges of scaling unstructured architectures and the search for more efficient regimes.
- 42:00 — Automated Model Design (AFMD): How Liquid AI uses meta-learning to automate the search for optimal architectures for specific hardware.
- 50:00 — Optimizing for NPUs and Hardware: The importance of designing simplified, structured networks that align with the capabilities of modern AI chips.

## Actions

- request_transcript: `POST https://stenobird.com/v1/public/podcasts/the-cognitive-revolution/episodes/intelligence-on-the-edge-liquid-ai-s-ramin-hasani-on-the-search-for-device-native-foundation-models/transcription-requests` — Idempotently request low-priority transcript generation for this episode.
- read_markdown: `GET https://stenobird.com/podcast/the-cognitive-revolution/intelligence-on-the-edge-liquid-ai-s-ramin-hasani-on-the-search-for-device-native-foundation-models.md` — Read the agent-friendly Markdown representation of this episode resource.

A page view does not enqueue transcription. Agents should invoke `request_transcript` explicitly when they need this episode processed.

## Transcript

Full transcripts are not published on public pages unless there is a clear rights basis.