Greg Brockman, co-founder and President of OpenAI, discusses his journey from building the first reverse Turing test game in 2008 to co-founding one of the world's leading AI companies. His early fascination with Computing Machinery and Intelligence by Alan Turing shaped his vision for building learning machines rather than programmed systems.

The conversation covers the pivotal 2012 ImageNet Classification with Deep Convolutional Neural Networks breakthrough that launched the deep learning revolution, Brockman's time at Stripe as employee #4, and the founding of OpenAI in his San Francisco living room in 2016. He explains the technical mechanics of how ChatGPT works, from pre-training on internet-scale data to post-training for specific behaviors.

Brockman details OpenAI's transition to an agent-first company by March 2025, the compute shortage limiting AI deployment, and recent breakthroughs where AI systems are discovering new knowledge in physics. The discussion explores the future of coding, the nature of AGI, and how AI will transform knowledge work across industries.

The Reverse Turing Test Game That Started Everything

In 2008, Brockman built a competitive game where two humans each talk to another human and an AI, trying to identify the other human first while acting bot-like themselves to avoid detection.

After two weeks of sitting alone in the game lobby, Stumble Upon sent 1,500 visitors in one day, creating constant gameplay with 3-4 games running simultaneously.

The bot used a database of previous conversations to match similar exchanges and reply with what humans had said before, working well for casual chat but failing on sophisticated topics.

Turing's 1950 Vision Predicted Modern AI Perfectly

Computing Machinery and Intelligence outlined building learning machines rather than programming explicit rules: 'You will never program it. Instead, you will need to build a machine that can learn its own answer.'

Turing described the exact process used today: build unsupervised models that observe the world, then use reinforcement learning with human rewards and punishments to achieve objectives.

The delay from 1950 to now came down to one factor: 'Very simple answer. It's compute. There just was not enough compute.'

AlexNet's 2012 Breakthrough Launched Deep Learning

The ImageNet Classification with Deep Convolutional Neural Networks paper by Hinton, Sutskever, and Krizhevsky didn't just win the ImageNet competition - it 'blew everything else out of the water.'

Alex Krizhevsky was writing fast GPU kernels while others dismissed it as 'just an engineering project,' but Ilya Sutskever recognized the breakthrough potential when combined with ImageNet.

Jeffrey Hinton used a management trick: 'Each week that you get a 1% improvement on the data set, I will push back the deadline on your review paper by one week' - done dozens of times.

Before 2012, neural nets were considered fraud science according to a 1995 paper on neural network booms and busts: 'These neural net people have no new ideas. They just want to build bigger computers.'

From Stripe Employee #4 to OpenAI Co-Founder

Brockman joined Stripe as the fourth employee after an instant connection with Patrick Collison over technical details like split keyboards and Dvorak layouts.

After 4.5 years at Stripe, Brockman left because 'if it really is so hard to build that group of people who can accomplish significant things, I got to get started now.'

OpenAI began in Brockman's San Francisco living room in 2016 with Sam Altman, Ilya Sutskever, Wojciech Zaremba, and others, unified by the goal of building beneficial AGI.

The founding team created a three-step plan on a flip chart: '1) solve RL (reinforcement learning), 2) solve UL (unsupervised learning), 3) gradually learning more complicated things' - which they've followed for a decade.

How ChatGPT Actually Works Under the Hood

Pre-training involves month-long runs costing millions, where models learn by predicting what comes next in sequences: 'If you can predict every word out of Einstein's mouth, you are at least as smart as Einstein.'

Post-training takes just days to teach behavior through reward models that judge outputs: 'Pre-trained models are less like a human and more like a humanity - everything's in there.'

The training process requires 2 AM wake-up calls when GPU runs crash: 'Every minute that it's down, you just look at the number of GPUs sitting idle and think about how many dollars are being wasted.'

Models now train on their own generated data through reinforcement learning: 'The machine goes out and tries to solve a task and you learn from everything it learns.'

OpenAI Goes Agent-First by March 2025

By March 31st, OpenAI aims for two goals: agents become the default tool over text editors, and all usage follows explicitly evaluated safety protocols.

The company mandates 'say no to slop' - human reviewers must hold higher quality bars for AI-generated code than human-written code.

Best engineers control interfaces and file structure by hand while outsourcing intricate implementation details to AI: 'The size of building blocks is going to increase over time.'

Future productivity will be measured by 'how much compute does an individual human marshal' as people become managers of agent fleets.

AI Discovers New Physics Knowledge

A physics professor who was an AI skeptic gave OpenAI's unreleased system a quantum physics hypothesis he planned to work on all year with collaborators.

The AI proved the opposite of the expected answer was true: 'His reaction was that this is the first time that it's felt like the system is thinking. There's new knowledge in there.'

The paper is being submitted for publication, representing 'a very significant moment and very representative of things to come' in AI-driven scientific discovery.

The Compute Shortage Limiting AI's Potential

Individual OpenAI employees now want 1,000 dedicated GPUs each: 'If you have 1,000 such individuals, you're at a million GPUs already. And there's not 10 million GPUs in existence.'

OpenAI cuts prices 100x year-over-year while demand explodes: 'Where we were for GPT-3 in 2020... you can get that level of intelligence... on your phone for free.'

'Compute will be a basic human right' because 'the more compute they have, the higher quality of life they can have' for economic productivity.

AI represents 'a manufacturing process from electricity to intelligence' where electricity serves as the fundamental input driving the entire system.