Greg Brockman on OpenAI's Road to AGI
Greg Brockman on OpenAI's Road to AGI
Greg Brockman on OpenAI's Road to AGI
In this episode of the Latent Space podcast, Greg Brockman, co-founder and president of OpenAI, joins the conversation to explore the cutting-edge developments shaping the future of AI. From the evolution of reasoning in large language models to the practical applications of reinforcement learning, Brockman offers insights into how OpenAI is pushing the boundaries of what's possible in artificial intelligence.
The discussion covers the progression from GPT-4 to GPT-5, emphasizing improvements in reasoning, reliability, and real-world application. Brockman highlights the importance of reinforcement learning and online learning in achieving more sample-efficient training. He also touches on the concept of supercritical learning, where systems surpass critical thresholds and begin to learn autonomously. The conversation delves into model architecture innovations, including hybrid systems and compute-efficient routing. Practical advice for developers includes structuring prompts, leveraging agent-based tools, and understanding model limitations. Brockman also explores the future of AI engineering, the evolving role of developers, and the broader implications of achieving AGI, including economic shifts and the long-term value of compute power. Throughout, he emphasizes the importance of perseverance, mission-driven work, and strategic focus in AI development.
00:04
00:04
Greg Brockman proud of the team's achievements
01:04
01:04
After training GPT-4, the team realized it could do chat and began questioning why it wasn’t AGI.
05:10
05:10
Reinforcement learning enables models to learn from fewer examples
06:44
06:44
Compute remains the limiting factor in AI despite efficient algorithms
08:16
08:16
Compute is a fundamental fuel for intelligence, turning energy into reusable models
13:21
13:21
Algorithms can now handle complex environments like Dota.
16:34
16:34
A 40B neural net trained on DNA achieves early GPT-level performance
19:33
19:33
GPT-5 can perform great intellectual feats in hard domains like math and is more reliable than previous versions
22:51
22:51
GPT-5 excels at solving complex intellectual problems and uses interactive coding with user feedback for training.
29:43
29:43
OpenAI uses defense in depth with instruction hierarchy for agent robustness
31:48
31:48
Models are trained on human thought and refined through reinforcement learning.
38:17
38:17
Leveraging preference data improves model training
39:13
39:13
GPT-5's router selects between reasoning and non-reasoning models based on application needs
45:20
45:20
1,000x cost improvement for the same intelligence since GPT-4
48:37
48:37
Practical engineering decisions demonstrate the model's capabilities using cutting-edge techniques
49:11
49:11
GPT-5 can work with GPT-OSS and Codex infrastructure enabling seamless interplay and multiplayer
51:34
51:34
AI tools offer a productivity boost, enabling teams to accomplish more work with current structures
54:16
54:16
Building codebases around LLM strengths with modular design and quick unit tests
55:27
55:27
AI projects are compared to the New Deal and Apollo program in scale, signaling a major economic shift.
58:42
58:42
OpenAI focuses on long-term bets for major breakthroughs in AI research.
1:01:45
1:01:45
Teams shifted from robotics to digital tools for faster progress.
1:03:05
1:03:05
Connecting AI models to real-world applications is valuable, despite the feeling that all ideas are taken.
1:04:20
1:04:20
I haven't done angel investing for years as it's a distraction from OpenAI.
1:06:00
1:06:00
Compute will be the key resource shaping future societies
1:07:07
1:07:07
Greg wishes he had internalized earlier that amazing tools are now available to revolutionize fields.
