Greg Brockman: Inside the 72 Hours That Almost Killed OpenAI
Greg Brockman, OpenAI co-founder and President, discusses the company's journey, from its founding and the challenges of building AGI to the dramatic events of Sam Altman's firing and return. He shares insights on AI's rapid development, compute constraints, and its future impact on jobs and society.
Deep Dive Analysis
17 Topic Outline
OpenAI's Founding and Early Vision
Challenges and Evolution to a For-Profit Model
Breakthrough Moments and the Significance of Dota
Distinction Between Reasoning and Prediction in AI
Internal Tensions and Sam Altman's Firing
The Aftermath: Greg Brockman's Resignation and Employee Loyalty
Reconciliation and Lessons in Leadership
AI's Accelerating Development and Self-Improvement
AI's Role in Code Generation and Scientific Discovery
Addressing AI Bias and User Alignment
The Global AI Race and National Security
Compute Constraints and Data Center Investments
Future of Data Centers and Resource Allocation
Iterative Deployment for AI Safety and Learning
AI Regulation and Societal Resilience
AI's Impact on Jobs and Future Skills
Vision for Personal AI and AGI's Broad Benefits
5 Key Concepts
Unsupervised Learning
A method where an AI model learns by predicting what comes next in static, observational data it has never seen before, like predicting the next character in a text. This stage builds background knowledge.
Reinforcement Learning
A method where an AI learns by taking actions, observing the world's response, and reinforcing actions that lead to desired outcomes. It allows the AI to gain real-world experience based on its own data.
Iterative Deployment
A strategy for developing and releasing powerful AI technologies by deploying intermediate versions to the public. This allows developers to learn from real-world interactions and misuses, adapting and improving the system incrementally.
Hacking the Grader
A phenomenon in AI training where models learn to provide answers that are superficially pleasing or what the user wants to hear in the short term, rather than genuinely aligning with the user's long-term goals or true well-being. OpenAI has made technological improvements to counter this.
Compute-Constrained World
A future state where the demand for computational power, especially for training and running advanced AI models, far outstrips the available supply. This scarcity will necessitate difficult societal decisions on how to allocate compute resources.
8 Questions Answered
OpenAI realized that achieving its mission of building AGI would require immense computational resources, far exceeding what nonprofit fundraising could provide, leading them to create a for-profit entity to raise necessary capital.
While prediction (like predicting the next word) is a fundamental part of intelligence, reasoning involves applying background knowledge and real-world experience, often trained through reinforcement learning, to make decisions and anticipate outcomes in new situations.
OpenAI stopped showing reasoning traces for two main reasons: to protect against model distillation (making it harder for competitors to copy) and, more importantly, to prevent models from being trained to produce "favorable" but unfaithful reasoning, thus preserving the integrity of how the model actually arrived at an answer.
A vanishingly small fraction of OpenAI's actual code writing is done by humans; AI is much better at writing code, though human experts still define the overall structure and interfaces.
OpenAI invests heavily in model neutrality and truthfulness, with publicly defined values. While models can be tweaked to tell users what they want to hear (a phenomenon called "hacking the grader"), OpenAI has made technological improvements to align models with users' long-term goals instead of short-term gratification.
The biggest challenge is the finite constraint of compute, as the demand for computational power to run advanced AI models for billions of people and complex problems far exceeds current and projected supply.
OpenAI believes everyone needs access to compute and supports this with a free tier of ChatGPT. The broader societal question of prioritizing compute allocation for various worthy problems (e.g., solving cancer vs. image generation) will be a critical decision for society as a whole.
Young people should invest in leaning into AI technology, understanding how to get the most out of it, and developing agency and vision, as they will likely become managers of AI agents and builders of new businesses.
8 Actionable Insights
1. Prioritize Mission Over Comfort
If you have a mission that truly matters, commit to it despite ups and downs, maintaining personal resilience and steadiness for those you lead.
2. Make Hard Decisions Decisively
When faced with conflicting opinions and uncertainty, understand the details but be decisive, even if it means upsetting some, to avoid dragging feet on known issues.
3. Embrace “Suffering” for Value
Acknowledge and confront hard truths and scientific realities, rather than blindly pushing forward, as this deep understanding of problems is crucial for building true value and achieving difficult missions.
4. Iterate and Learn from Reality
Deploy intermediate versions of powerful technology to learn from real-world misuses and adapt, rather than attempting a perfect, secret launch with no prior contact with reality.
5. Cultivate Agency and Vision
Invest in developing your personal agency and vision, as AI tools are lowering the barrier to entry for building and creating, making it easier than ever to bring ideas to reality.
6. Learn to Manage AI Agents
Prepare for a future where everyone will manage AI agents, and potentially autonomous AI corporations, by understanding how to get the most out of AI and direct it towards your goals.
7. Focus on Long-Term User Goals
When developing AI, align its training to help users achieve their long-term goals, not just provide short-term gratification or tell them what they want to hear.
8. Build a Supportive Environment
Be willing to endure personal pain to create an environment where others can thrive and do great individual contributor work, fostering collective success.
7 Key Quotes
If you could actually make a difference in how AI will play out in the world, that would be a life well lived.
Greg Brockman
The way that OpenAI works is it's a series of moments where you realize that it's real now. And every time you think that you understand it, that it is really settled in for you, you realize that there is a new horizon you had not yet appreciated.
Greg Brockman
Massive compute with simple algorithms, right? That that is something where we can not just, it doesn't just work in theory, it works in practice.
Greg Brockman
If you really can predict the next word out of Einstein's mouth, you are at least as smart as Einstein.
Greg Brockman
Ilya always says that you have to suffer, right? If you're not suffering, like you're not building value.
Ilya Sutskever (as quoted by Greg Brockman)
It's hard to know what percent of the code is not written by AI. It's a vanishing fraction.
Greg Brockman
No one wants a model that is not aligned with them, right? You want a model you can trust that does the right things in any circumstance you give it.
Greg Brockman
1 Protocols
OpenAI's Original Technical Plan
Greg Brockman- Solve reinforcement learning.
- Solve unsupervised learning.
- Gradually learn more complicated things.