
#150 From a Worm's Brain to Mercedes: Building AI That Runs Without the Cloud // Mathias Lechner, Liquid AI
Shownotes
Can a model small enough to run on a Raspberry Pi be good enough for real products? Mathias Lechner thinks so. As co-founder and CTO of Liquid AI, he builds foundation models that run on phones, laptops and cars without the cloud, and his team starts with the hardware before anything else.
- From C. elegans to liquid neural networks
- No long-term technical moat, and why silicon partnerships matter
- Speed as the North Star for on-device models
- The memory triangle: bandwidth, size and compute
- AI in cars and the challenge of 10-year hardware
- Training data, evals and research culture at Liquid AI
00:00:51 Meet Mathias: from a worm's nervous system to Liquid AI
00:01:41 Why computers felt like magic, and how he found C. elegans
00:04:39 A 19-neuron self-driving demo while the industry scaled up
00:07:34 Robustness vs knowledge: what a car really needs to know
00:08:38 Is vision enough for autonomous driving?
00:09:28 From the MIT test track to Liquid AI's LFM2 models
00:09:57 Is there any worm left in the model?
00:12:10 No long-term technical moat: why silicon partnerships matter
00:15:01 What chipmakers get in return, and access to pre-release hardware
00:16:14 Inference speed as the North Star
00:17:12 Cloud vs local: where workloads will run
00:19:02 Memory bandwidth, memory size and FLOPS: the on-device triangle
00:22:35 How many tokens per second a phone assistant really needs
00:23:34 Is the local bet really about tools?
00:24:59 What runs on device and what runs in the cloud by 2028
00:28:11 Unmetered intelligence: token maxing with a local agent
00:28:29 His setup: Claude Code, Codex and OpenCode in YOLO mode
00:31:15 Why Mercedes uses Liquid AI for its in-car assistant
00:33:06 Why not a Chinese open model? Speed is non-negotiable
00:34:45 Why cars only get Raspberry Pi-level compute
00:35:41 Cars live 10 years, models six months: continual updates
00:38:00 Testing Liquid AI's open-weight models: where to start
00:40:00 Inside the training cluster: thousands of GPUs, academic workflow
00:42:47 Data: the unglamorous work behind model quality
00:45:06 How to judge data: would a user actually ask this?
00:47:12 From use cases to evals and benchmarks
00:50:04 Learning resources he recommends
00:51:35 His recent aha moment: tiered memory systems
00:53:39 Tobi's surprise: a message to his 2017 self
00:55:21 Hardware awareness from day one
- Which AI workloads are moving on-device and which will stay in the cloud
- Why memory bandwidth and memory size, not just FLOPS, decide what can run locally
- How hardware partnerships can matter more than model architecture as a moat
- What it takes to ship AI into cars with long hardware lifecycles
- How to judge training data by asking whether real users would need it
- Where to start if you want to test small open-weight models yourself
"For me at least, there's no long-term technical moat anywhere... It's just innovation. It's a race." – Mathias Lechner [00:12:10]
"The way we think about the moat is actually about silicon partnerships." – Mathias Lechner [00:14:14]
"For us, speed is non-negotiable. We start actually with the speed." – Mathias Lechner [00:34:13]
"If you have a good local agent, you can token max because it's basically free." – Mathias Lechner [00:28:11]
Liquid AI: https://www.liquid.ai
Mathias on Liquid AI: https://www.liquid.ai/team/mathias-lechner
Liquid AI models on Hugging Face: https://huggingface.co/LiquidAI
Join alphalist: https://alphalist.com
alphalist is the biggest community of CTOs in Germany. If you're an experienced CTO, join 600+ peers sharing insights in our Slack space and at our events. Apply at alphalist.com.
Alle Folgen von alphalist.CTO Podcast – For CTOs and Technical Leaders