Fireside Chat, Interview, Conference Presentation
Opening Fireside: Ramin Hasani, Liquid AI & Ion Stoica, Arena | RAISE Summit 2026
- Liquid AI aims to maximize intelligence within the smallest possible compute units, enabling efficient foundation models to operate on CPUs, GPUs, and NPUs outside data centers with linear rather than exponential energy scaling.
- Enterprises are expected to shift from prototyping with powerful proprietary models to adopting smaller, open-source models that offer greater control, reliability, and lower energy and memory costs.
- The adoption of open-source technology is projected to explode as these models reach a capability threshold comparable to top proprietary models like GLM 5.2 and DeepSeq, which current open-weight versions are approaching.
- A future architecture will rely on orchestration of specialized models forming "armies" or "cities" to handle determinism, inputs, and outputs reliably, rather than isolated large models.
- Governments and regulators may soon extend restrictions on powerful proprietary models to open-weight and open-source variants, creating uncertainty for organizations and potentially targeting models a few months behind the latest proprietary releases.
- Sophisticated enterprises, particularly in Europe, are urged to build and own their intelligence capabilities immediately to avoid risks associated with future regulation and to ensure reliable access.
- Intelligence is predicted to emerge through social evolution and communication in agentic pipelines, with many specialized open-source models capable of handling languages, vision, and audio within the next year or two.
- Security concerns are anticipated to dissipate over the next two years, though reliability remains the primary challenge for AI to drive industry-wide change.
- Humans will remain ultimate arbiters for the foreseeable future, requiring specialized domain expertise to be liable for deploying AI capabilities as models become more powerful.
- Despite the predicted explosion of open-source adoption, the risk of regulations extending to already downloaded open-weight models presents a significant constraint on distribution.