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

OpenAI’s mission is to ensure Artificial General Intelligence (AGI) benefits all of humanity. The initial vision of AGI as a singular, oracular system has evolved. The new focus is on creating powerful AI tools to empower people to build the future. The long-term goal is a personal AGI that integrates with various services to assist in work and personal life, ultimately accelerating scientific discovery and societal progress.

This vision rests on three core pillars: Research, Product, and Infrastructure.

Research: The Path to Superintelligence and Automated Science

OpenAI’s core focus is on scaling deep learning with the belief that superintelligence—systems smarter than humans on many critical axes—could be less than a decade away. The research program is organized around the potential for these systems to accelerate scientific discovery, which is seen as the most significant long-term impact of AI.

Measuring and Projecting Progress

Progress is measured by the time horizon of tasks AI can perform. Current models can accomplish tasks that would take a top human expert about five hours (e.g., performance in informatics Olympiads). This horizon is expected to expand rapidly through algorithmic innovation and scaling, particularly along the axis of “in-context compute” or “test-time compute,” where there are orders of magnitude for growth.

Public Research Goals and Timelines

OpenAI shared its internal goals to provide transparency about its planning:

• AI Research Intern: By September of next year, the goal is to have a capable AI research intern that can meaningfully accelerate researchers by leveraging significant compute.
• Automated AI Researcher: By March 2028, the aim is to have a system capable of autonomously delivering on larger research projects—a fully automated AI researcher.

A Multi-Layered Approach to Safety and Alignment

As systems become more capable, safety becomes paramount. OpenAI structures its thinking around five layers:

1. Value Alignment: The most critical long-term question. This concerns the AI’s fundamental principles and what it “cares about,” especially when given unclear objectives or operating beyond human ability. The goal is a deep alignment with principles like “love for humanity.”
2. Goal Alignment: How the AI interacts with people and follows instructions.
3. Reliability: The AI’s ability to calibrate its predictions, express uncertainty, and handle unfamiliar environments.
4. Adversarial Robustness: The system’s resilience against targeted attacks from human or AI adversaries.
5. Systemic Safety: Guarantees about the overall system’s behavior that don’t rely on the AI’s intelligence, such as security controls and data access limitations.

A promising technical direction for studying value alignment is Chain-of-Thought Faithfulness. The idea is to keep parts of the model’s internal reasoning free from supervision during training, allowing it to remain a more representative “faithful” view of the model’s actual thought process. This requires preserving a “controlled privacy” for the model, a principle that influences both algorithm and product design at OpenAI.

Product: Evolving from Assistant to Platform

The product vision is to evolve from an AI super assistant within ChatGPT to an “AI Cloud”—a platform that others can build upon. The goal is to enable a world where more value is created by people building on the platform than by the platform creator itself.

Two foundational principles guide this platform strategy:

• User Freedom: Recognizing that users worldwide have different needs, the platform will offer significant control and customization. The aim is to treat adult users like adults, allowing for a broad range of human expression while maintaining broad safety bounds.
• Privacy: AI privacy is considered especially critical because users share intimate details with the technology. This necessitates strong technical protections and new policy concepts like “AI privilege” to safeguard user data.

Infrastructure: Building the Foundation for AGI

To power this vision, a massive infrastructure buildout is underway.

• Current Commitments: Total over 30 gigawatts of infrastructure, representing a $1.4 trillion financial obligation over many years. This requires extensive partnerships with companies like AMD, Google, Microsoft, Nvidia, Oracle, and many others across the supply chain.
• Future Aspiration: The long-term goal is to create an “infrastructure factory” capable of producing 1 gigawatt of compute per week, with a target cost of $20 billion per gigawatt over a 5-year lifecycle. This ambitious plan will require significant innovation in robotics and supply chain management.

A New Corporate Structure

To facilitate this scale and align with the mission, OpenAI has simplified its corporate structure:

• The OpenAI Foundation (Non-Profit): This entity is in control of the overall mission. It governs the Public Benefit Corporation (PBC) and will initially own about 26% of its equity, which can increase over time.
• OpenAI Group (Public Benefit Corporation): This entity will operate more like a traditional company to attract the necessary capital for the infrastructure buildout, while remaining bound to the non-profit’s mission.

The Foundation’s initial focus will be on two major initiatives:

1. Curing Disease: A $25 billion commitment to use AI for scientific discovery in medicine, from generating data to providing compute grants.
2. AI Resilience: A new field, broader than AI safety, focused on creating an ecosystem of organizations to mitigate the risks and disruptions of advanced AI (e.g., bio-risks, job displacement, mental health). This is analogous to the cybersecurity industry that grew to protect the internet.

• Tags:
• AI Safety
• Feature
• Superintelligence
• Scientific Discovery
• AGI
• Deep Learning
• Platform Strategy
• AI Alignment
• Compute Infrastructure