Open Frontier Research
The Crisis
Open frontier AI research in Western democracies faces an existential threat. The gap between resources available to academic researchers and closed corporate labs has grown from 2x (a decade ago) to 10-1000x today. Major AI labs that once published openly - Google, OpenAI, DeepMind - have progressively closed their doors.
Meanwhile, China has emerged as the leader in open AI contributions, with companies like DeepSeek and Alibaba publishing detailed technical papers and releasing competitive open-weight models.
Why Open Research Matters
Historical Precedent
As Jeff Dean notes, Google itself was built on open academic research:
- TCP/IP from academic work
- PageRank from Stanford Digital Library Project funding
- Deep learning from neural network research 30-40 years ago
Every major technology paradigm has been accelerated by open collaboration.
The Manhattan Project vs. Internet Model
Yan Stoica contrasts two historical models for important technology development:
- Manhattan Project: Closed, nationalized, controlled
- The Internet: Open, collaborative, shared artifacts
Both achieved their goals, but the internet model created more distributed innovation and benefited more people.
Democracy and Power
Yoshua Bengio frames open AI as essential to democracy:
"The power of intelligence is going to give power to whoever controls it, including potentially AIs that could be smarter than us, including dictators. This is a threat to our democracies because democracy is about sharing power."
Sample Efficiency and Representation
Yejin Choi emphasizes that AI should be:
- Of humans - Reflecting human values
- By humans - Created by diverse societies, not just a few companies
- For humans - Benefiting all, not just those in power
The Current State
Closed Labs
Major US labs have progressively restricted publication:
- OpenAI: Once published everything, now publishes selectively
- Google/DeepMind: Reduced open publication, competitive pressure
- Anthropic: Limited public research sharing
Open Alternatives
The open ecosystem continues through:
- Meta AI: Llama models, open research culture
- Chinese Labs: DeepSeek, Qwen, detailed publications
- Academic Collaborations: LAUD Institute, Berkeley AI Research
- Startups: Some (like Mistral) commit to openness
What's Needed
Resources
Academic researchers need:
- Compute access: Comparable to what closed labs have
- Competitive compensation: 2x difference, not 10-1000x
- Long-term funding: 3-5 year research horizons
Collaboration Infrastructure
- Shared artifacts: Code, models, datasets
- Open benchmarks: Transparent evaluation
- Publication culture: Rewarding openness
Policy Support
- Government funding: NSF, DARPA investment in open AI
- International cooperation: Not just US, global collaboration
- Regulatory frameworks: That don't disadvantage open research
Key Voices
Yan Stoica (Berkeley):
"In order for the best minds to collaborate, they need to be able to share the information and work on shared artifacts."
Yoshua Bengio (Law Zero):
"We need a global agreement on what to do and not do, where countries agree not to use AI to dominate others."
Yejin Choi (Stanford):
"We need to democratize generative AI. Different countries and different social sectors should be able to create that AI."
Related Reading
- Yoshua Bengio - Founder of Law Zero, advocate for AI safety and openness
- Yejin Choi - Stanford professor advocating for democratic AI
- Jeff Dean - Google's Chief Scientist, advocate for academic funding