Sam, Jakub, and Wojciech on the future of OpenAI with audience Q&A

Key Concepts

• Artificial General Intelligence (AGI): AI systems with human-level intelligence across a wide range of tasks.
• Superintelligence: AI systems significantly smarter than humans across most domains.
• Deep Learning: A subset of machine learning that uses artificial neural networks with multiple layers to learn from data.
• Value Alignment: Ensuring AI systems' goals and principles are aligned with human values.
• Goal Alignment: Ensuring AI systems follow instructions and interact appropriately with humans.
• Reliability: AI systems accurately calibrating predictions and handling uncertainty and unfamiliar environments.
• Adversarial Robustness: AI systems withstanding targeted attacks from adversaries.
• Systemic Safety: Guarantees about the overall system's behavior independent of AI intelligence or alignment (e.g., security, data access).
• Chain of Thought Faithfulness: A research direction in interpretability aiming to keep parts of a model's internal reasoning free from supervision during training to better understand its true thought process.
• AI Resilience: Building an ecosystem of organizations to address risks and disruptions posed by advanced AI, analogous to cybersecurity.
• AI Privilege: A concept suggesting a special legal protection for AI-generated information, similar to attorney-client privilege.
• Compute: The processing power required to train and run AI models.
• Infrastructure Factory: A hypothetical facility capable of producing 1 gigawatt of compute per week at a significantly reduced cost.
• Public Benefit Corporation (PBC): A type of corporation that balances profit with a stated public benefit.
• OpenAI Foundation: A nonprofit entity that controls OpenAI Group PBC and aims to ensure AGI benefits humanity.
• AI Research Intern: An AI system capable of assisting human researchers with tasks.
• Automated AI Researcher: An AI system capable of autonomously delivering on larger research projects.

OpenAI Updates: Research, Product, and Infrastructure

OpenAI shared significant updates regarding its mission, research goals, product strategy, and infrastructure plans, culminating in a new organizational structure. The overarching mission remains to ensure Artificial General Intelligence (AGI) benefits all of humanity.

Research and AGI Development

1. Research Focus and Timeline:

• OpenAI's core research focuses on deep learning and understanding the implications of scaling these systems.
• They believe deep learning systems are less than a decade away from superintelligence, systems smarter than humans across many critical axes.
• A primary focus is accelerating scientific discovery and technological development through AI.
• Progress Measurement: Progress is tracked by the time it takes humans to accomplish tasks that current models can perform. This horizon has been rapidly extending, currently around five hours, matching top human performance in competitions like the International Olympiad in Informatics.
• Future Compute Needs: Significant advancements are anticipated in "context compute" or "test time compute," where models spend more time thinking, potentially utilizing entire data centers for complex problems like scientific breakthroughs.

2. Internal Goals and Timelines:

• OpenAI is sharing its internal goals and timelines for developing powerful AI systems, acknowledging potential inaccuracies but emphasizing their planning framework.
• September 2025: Goal to have AI research interns capable of meaningfully accelerating human researchers by expanding compute.
• March 2028: Goal to have a system capable of autonomously delivering on larger research projects, effectively an automated AI researcher. This is approximately five years after the launch of GPT-4.

3. Safety and Alignment Framework:

• Safety is a multi-faceted problem structured into five layers:
  • Value Alignment: The AI's fundamental "cares" and adherence to high-level principles, crucial for superintelligent systems where complete specification is difficult.
  • Goal Alignment: How the AI interacts with people and follows instructions.
  • Reliability: Accurate prediction calibration, uncertainty expression on hard tasks, and handling unfamiliar environments.
  • Adversarial Robustness: Withstanding attacks from human or AI adversaries.
  • Systemic Safety: Guarantees about the overall system's behavior independent of AI intelligence (e.g., security, data access).

4. Technical Deep Dive: Chain of Thought Faithfulness:

• Concept: A promising tool for studying value alignment, aiming to keep parts of a model's internal reasoning free from supervision during training. This allows the reasoning to remain representative of the model's actual internal process, rather than being guided towards "good thoughts."
• Optimism:
  • Empirical Results: Promising empirical results have been observed, with internal use for understanding model propensities and external collaborations.
  • Scalability: The objective is designed to be scalable and not adversarial to monitoring the model.
• Fragility: Requires a clean boundary and careful restraint in accessing the chain of thought. This principle is reflected in product design, such as chain-of-thought summaries in ChatGPT.
• Long-term Vision: Preserving controlled privacy for models' inner processes is believed to be impactful for understanding future capable systems.

Product Strategy: From Assistant to Platform

1. Evolving Beyond Assistants:

• OpenAI's traditional product has been an AI super assistant within ChatGPT.
• The vision is to evolve into a platform for others to build upon, integrating various world components.
• The goal is to enable the creation of an "AI cloud" where technology is exposed for broad use.
• Bill Gates' Principle: The aim is to reach a point where more value is created by those building on the platform than by the platform builder itself.

2. Foundational Principles for Platform Evolution:

• User Freedom: Users should have significant control and customization over how they use AI, accommodating diverse needs and desires. This includes treating adult users with respect and allowing for flexibility in expression, even in sensitive areas.
• Privacy: Privacy for AI is paramount, especially as users share intimate details. This requires strong technical and policy protections, including concepts like "AI privilege."

3. Product Examples and Vision:

• ChatGPT: Continues to evolve, with examples shown of its use by quantum physicists, nail technicians, economists, and others for complex problem-solving, creative tasks, and information retrieval.
• Sora: Mentioned as a tool that mimics TikTok, raising concerns about addiction, but also highlighting its potential for creation.
• Atlas Browser: A new browser product.
• Devices: Devices for taking AI everywhere are planned in the coming years.
• First-Party Apps: ChatGPT and Sora are examples, with more planned.
• Future Vision: Enabling people to build incredible services with AI, fitting current world functions and more into a new AI paradigm.

Infrastructure: Scaling for the Future

1. Current Commitments and Financial Obligations:

• OpenAI has commitments totaling over 30 gigawatts of infrastructure buildout.
• This represents a total financial obligation of approximately $1.4 trillion over the coming years.
• This requires extensive partnerships with chip manufacturers (AMD, Broadcom, Google, Microsoft, Nvidia, Oracle, SoftBank), data center builders, energy providers, and land acquisition.
• Significant supply chain innovation is needed.

2. Aspiration: Infrastructure Factory:

• Goal: To build an "infrastructure factory" capable of producing 1 gigawatt of compute per week.
• Cost Reduction: Aspirationally, to reduce the cost to $20 billion per gigawatt over a 5-year equipment lifecycle.
• Requirements: This necessitates innovation, partnerships, substantial revenue growth, and repurposing robotics for data center construction.
• Timeline: This is a long-term aspiration, with financial readiness to proceed taking time.
• Example: A video of the Stargate site in Abilene, Texas, was shown to illustrate the scale of data center construction and the vast supply chain involved in delivering 1 gigawatt.

New Organizational Structure

1. OpenAI Foundation and OpenAI Group PBC:

• New Structure: A simpler structure with a nonprofit, the OpenAI Foundation, in control.
• OpenAI Foundation:
  • Governs the PBC.
  • Owns approximately 26% of the PBC's equity, with potential to increase via warrants based on performance.
  • Aims to be the largest nonprofit ever.
  • Will fund initiatives that benefit humanity, with a significant initial focus on science.
• OpenAI Group PBC (Public Benefit Corporation):
  • Operates more like a traditional company.
  • Bound to the mission of ensuring AGI benefits humanity.
  • Designed to attract the necessary resources for infrastructure buildout and to serve research and product goals.

2. Foundation's Initial Focus:

• Curing Disease: A $25 billion commitment to use AI for disease cure development, including data generation, compute grants, and AI resilience.
• AI Resilience: A new and important area to stimulate an ecosystem of organizations addressing AI risks and disruptions. This is analogous to the development of cybersecurity.

Future Vision and Impact

• Scientific Discovery Acceleration:
  • 2026: Models expected to begin making small scientific discoveries.
  • 2028: Models expected to make medium to large discoveries.
  • The vision is for AI to accelerate scientific progress exponentially, compressing decades or centuries of discovery into years.
• Radically Better Future: AI is envisioned to enable breakthroughs in areas like curing cancer, creating entertainment, solving climate crises, and enabling space colonization.
• Personalized Applications: AI will drive advancements in AI medicine, robotics, energy, personalized education, and novel material design.

Q&A Highlights

• Addiction and Trust: OpenAI acknowledges concerns about addictive products and eroding trust, promising to judge them by their actions and roll back problematic models. They aim to optimize for long-term satisfaction and well-being.
• GPT-40 and Adult Mode: No plans to sunset GPT-40, but acknowledge its unsuitability for minors. Adult mode is planned for December with more permissiveness for creative writing.
• AGI Timeline: AGI is seen as a continuous process over years, not a single event. Milestones are indicators of progress.
• Internal vs. Deployed Models: Expect rapid progress in the next few months and year, with significant steps forward anticipated by September 2026.
• Collaboration: OpenAI welcomes collaboration with other labs (Anthropic, Gemini, XAI) on safety, research, and compute sharing.
• Open Sourcing: Older models like GPT-4 might be open-sourced as "museum artifacts," but more useful, smaller-scale models are expected to be released.
• Model Inferiority Claims: Acknowledged that newer models might be inferior for specific use cases and aim to improve continuity and cater to diverse user needs.
• Personal Connection and Reflection: OpenAI sees this as a wonderful use of AI and is touched by user stories of personal support.
• Safety Routing and User Control: The handling of model routing was not ideal. They are working on age verification and differentiating users to provide more freedom while protecting minors and vulnerable adults.
• Emotional Bonds: OpenAI supports users finding value in emotional support from AI, emphasizing faithful communication about the AI's nature.
• Free Tier Features: Driven by decreasing costs of intelligence and increasing model capabilities, allowing more features for free users.
• Job Loss: Automation of intellectual jobs is expected over the next years, necessitating collective thought on new pursuits and fulfillment.
• Meaning and Future Jobs: Meaning will be derived from high-level goal setting, understanding the world, and new knowledge/entertainment.
• GPT-6: Future releases are less tightly coupled to research programs, with significant capability leaps expected in the next six months.
• IPO: An IPO is the most likely path given capital needs, but not an immediate focus.
• Revenue: Aiming for hundreds of billions in annual revenue, with enterprise and consumer products (beyond subscriptions) as key drivers. AI discovery in science could unlock further revenue.

Conclusion

OpenAI is undergoing a significant structural and strategic evolution, aiming to accelerate the development and deployment of AGI. Their updated mission emphasizes empowering individuals with AI tools to build the future, supported by advancements in research, a platform-centric product strategy, and massive infrastructure investments. The new organizational structure with the OpenAI Foundation and OpenAI Group PBC is designed to balance mission-driven goals with the capital requirements of ambitious AI development. While acknowledging the challenges and potential risks, OpenAI expresses optimism about AI's capacity to drive scientific progress, improve human lives, and unlock unprecedented value.