Greg Brockman: AI Is About to Go Parabolic! Here's What's Next
Greg Brockman, co-founder of OpenAI, discusses the founding of OpenAI in 2015, its evolution from nonprofit to for-profit, the dramatic firing and reinstatement of Sam Altman, and his vision for AI's transformative impact on society. He covers technical milestones, compute strategy, safety philosophy, and the future of personal AGI for all 8 billion people on Earth.
Summary
Greg Brockman recounts leaving Stripe to pursue AI, feeling that Stripe's mission wasn't personally his own. After a conversation with Sam Altman in 2015, the two decided to co-found OpenAI despite skepticism about competing with well-resourced incumbents like Google DeepMind. The founding team was assembled through a visionary offsite in Napa, where the core technical plan—solving reinforcement learning, unsupervised learning, and scaling complexity—was conceived. Notably, early candidates like Dario Amodei and Chris Olah ultimately went to Google Brain instead.
Brockman describes a series of technical milestones that reinforced belief in the mission: the DOTA reinforcement learning result, the unsupervised sentiment neuron paper in 2017 which showed semantics emerging from language modeling, and GPT-4, which blurred the lines of what AGI criteria even meant. He explains the transition from nonprofit to for-profit as a necessity driven by compute economics—realizing that achieving AGI would require massive capital that nonprofit fundraising could not support.
On the dramatic firing of Sam Altman by the OpenAI board in November 2023, Brockman describes being removed from the board simultaneously with no explanation given. He immediately decided to quit, inspiring a wave of employee loyalty—no one accepted competing offers during the chaotic weekend, a petition crashed Google Docs, and Ilya Sutskever's public reversal signaled the path to restoration. Brockman reflects on this as one of the most emotionally intense periods of his career, noting that Ilya's eventual departure from OpenAI was the only moment he considered quitting entirely.
Brockman articulates OpenAI's iterative deployment philosophy—deploying progressively more powerful systems rather than building in secret—as a core safety and learning mechanism. He notes that GPT-3's top misuse turned out to be medical spam, something no one had anticipated. He discusses the decision to hide chain-of-thought reasoning to preserve both competitive advantage and interpretability integrity, explaining that training models to make reasoning look good destroys its faithfulness as a window into model cognition.
Looking forward, Brockman envisions a compute-powered economy where everyone has a personal AGI, AI autonomously conducts scientific research targeting specific problems like cancer, and entrepreneurship becomes universally accessible. He acknowledges job displacement risks but argues the gains—empowerment, access to world-class medical advice, and the ability for anyone to build software—will outweigh the losses. He emphasizes that ensuring broad access to compute is as critical as the technology itself, and that OpenAI's mission remains ensuring AGI benefits all of humanity.
Key Insights
- Brockman argues that OpenAI's core technical plan—solve reinforcement learning, solve unsupervised learning, then scale complexity—was essentially conceived in a single day at a Napa offsite in 2015, before anyone had formally joined the company, and has guided the lab's direction for the subsequent decade.
- Brockman claims that truly predicting the next word out of Einstein's mouth would require being at least as smart as Einstein, arguing that prediction and intelligence are deeply connected—and that this same mechanism underlies both the unsupervised pretraining and reinforcement learning stages of modern AI training.
- Brockman reveals that OpenAI deliberately stopped showing chain-of-thought reasoning to users not primarily for competitive reasons, but because training models to make their reasoning look presentable destroys the faithfulness of that reasoning as an interpretability signal—meaning the visible thoughts would no longer reflect how the model actually arrived at its answer.
- Brockman states that the number one misuse of GPT-3 turned out to be medical spam advertising drugs to people—something the team never anticipated despite spending significant time thinking about grand misuse scenarios like misinformation—and uses this as the core justification for iterative deployment over secretive development.
- Brockman claims that the actual writing of code is now essentially done entirely by AI, with humans remaining superior only at higher-level structural decisions like module layout and interface definitions, and that this shift—already visible in tools like Codex—will extend to every field of computer-based knowledge work.
Topics
Full transcript available for MurmurCast members
Sign Up to Access