GTC 2026 Open Models Panel with Jensen
Category: Industry Panel · Year: 2026 · ▶ Watch
| Speakers: Anjney Midha, Founder | AMP PBC · Aravind Srinivas, CEO and Co-Founder | Perplexity · Arthur Mensch, Co-Founder and CEO | Mistral AI · Hanna Hajishirzi, Professor | Sr. Director NLP, AI2 · Jensen Huang, Founder and CEO | NVIDIA · Michael Truell, CEO and Co-Founder | Cursor · Mira Murati, Founder and CEO | Thinking Machines Lab · Misha Laskin, Co-Founder and CEO | Reflection AI · Robin Rombach, Co-Founder and CEO | Black Forest Labs |
Segments (13)
- 00:00 · Introduction
- Jensen Huang introduces the panel on open models and the need for models as technology.
- 01:10 · The Rise of Compound Agents
- Michael Truell discusses the emergence of companies building compound agents that mix open and closed models.
- 03:13 · AI as a System
- Aravind Srinivas explains that AI is an orchestration system, not just a single model.
- 05:00 · Open vs. Closed Ecosystems
- Misha Laskin argues that fundamental knowledge infrastructure naturally trends towards openness.
- 06:52 · The Need for Open Infrastructure
- Mira Murati highlights that open infrastructure and research are necessary for broad progress in AI.
- 07:40 · The Shift to Post-Training
- Jensen Huang notes the massive shift in compute from pre-training to post-training.
- 09:03 · Reinforcement Learning at Scale
- Misha Laskin uses AlphaGo as an example of how RL can solve complex problems given enough compute.
- 10:30 · Orthogonal Capabilities
- Mira Murati discusses the need for agents to connect context and operate within specific domains.
- 11:35 · Control and Customization with Open Models
- Arthur Mensch explains why open models are essential for enterprise control, resilience, and customization.
- 14:05 · Trust in Mission-Critical Applications
- Anjney Midha argues open models are strictly better for mission-critical apps because they allow for trust and introspection.
- 16:56 · Hybrid Models and Open Research
- Hanna Hajishirzi discusses the benefits of releasing the full development cycle of models and the efficiency of hybrid architectures.
- 19:12 · Open Models Driving Innovation
- Robin Rombach emphasizes that open models enable diffusion of innovation and competition.
- 20:16 · The Future of Visual Intelligence
- Jensen Huang predicts visual intelligence will evolve into robotic visual intelligent systems.
Product Announcements (4)
- [11:37] Mistral 7B
- Open model released in 2023
- specs: Foundation model
- availability: Released 2023
- [11:49] Mistral NeMo
- Open model collaboration
- specs: Foundation model
- availability: Released 2024
- [13:24] Forge
- Product to connect models to physical world data sources
- specs: Connects models to various data sources
- availability: Released
- [13:44] NeMoTron Coalition
- Coalition for sharing R&D costs
- specs: Shared R&D for open models
- availability: Active
Specific Numbers (3)
| Timestamp | Metric | Value | Context |
|---|---|---|---|
| 08:21 | Pre-training compute percentage | 90% | The amount of training compute used for pre-training 2-5 years ago. |
| 09:21 | AlphaGo parameters | 60 million | The size of the AlphaGo network, described as tiny relative to current models. |
| 09:58 | RL Investment | $10 billion to $100 billion | Hypothetical future investment to solve major problems like curing diseases using RL. |
Key Technologies (5)
- Compound Agents: Mixes different models (open and closed) and tools to solve complex, multi-step tasks.
- Orchestration Systems: Manages and delegates tasks to various sub-agents, models, and tools.
- Post-Training: The phase of model development after initial pre-training, crucial for specializing and refining capabilities.
- Reinforcement Learning (RL): Training models through trial and error to achieve specific goals, scaling with compute.
- Hybrid Models: Architectures combining different techniques (like transformers and state space models) for better efficiency.
Predictions / Commitments (5)
- [02:04, Over the course of the next year or two] We will see the rise of a new type of agent that acts as a co-worker for complex workloads.
- [08:27, In the future] The percentage of compute used for pre-training will become tiny, with the vast majority shifting to post-training.
- [09:50, Now and in the coming years] RL applied to language models will be used to solve massive problems, driven by economic decisions on compute allocation.
- [15:48, Heading into this phase now] Infrastructure will consolidate, similar to the industrial revolution, requiring open infrastructure to prevent hoarding.
- [20:39, In the future] Visual intelligence will increasingly become a robotic visual intelligent system, interacting with the physical world.
Companies Mentioned (2)
DeepMind/Google (implied via AlphaGo) · Mistral AI
Notable Quotes (4)
Proprietary versus open is not a thing. It’s proprietary and open. — Jensen Huang @ 00:58
AI is not the model, it’s the system, it’s the computer. — Aravind Srinivas @ 03:13
Fundamental knowledge infrastructure yearns for openness, like an animal yearns for the hills or for the forest. — Misha Laskin @ 06:12
Open models are strictly better than closed models. — Anjney Midha @ 14:05
Key Topics
Open Source AI · Proprietary Models · AI Agents · Compound AI Systems · AI Orchestration · Post-Training · Reinforcement Learning · AI Infrastructure · Mission-Critical AI · Hybrid Architectures · Robotic Vision
Takeaways
- The future of AI involves a mix of proprietary and open models working together within orchestrated systems.
- AI is evolving from single models to compound agents that act as co-workers for complex, multi-step tasks.
- The majority of compute investment is shifting from initial pre-training to post-training and reinforcement learning.
- Open models are critical for enterprise adoption because they offer necessary control, customization, and data privacy.
- For mission-critical applications (healthcare, defense), open models are preferred due to the need for trust and introspection.
- Open infrastructure and shared research are vital to prevent consolidation and accelerate the next generation of AI capabilities.