Inside OpenAI Enterprise: Forward Deployed Engineering, GPT-5, and More | BG2 Guest Interview

Episode Summary

Executive Summary: In this podcast, OpenAI's head of engineering and product discuss the company's enterprise platform, highlighting successful deployments with T-Mobile, Amgen, and Los Alamos National Labs. They explore the challenges of AI agent autonomy compared to self-driving cars, the trade-offs in building GPT-5, and the potential of reinforcement fine-tuning. The conversation also covers long/short investment ideas, the future of software engineering, and personal AGI moments.

Main Topics: OpenAI's Enterprise Platform (Priority: 5/5): Discussion of OpenAI's B2B efforts, including the API, enterprise products, and government deployments. Emphasis on distributing AGI benefits through enterprise customers. AI Autonomy: Physical vs. Digital (Priority: 4/5): Comparison of self-driving cars (physical autonomy) and AI agents (digital autonomy). Analysis of why physical autonomy is ahead despite higher safety bars. GPT-5 Development and Trade-offs (Priority: 5/5): Insights into building GPT-5, including trade-offs between reasoning tokens and latency, instruction following, and customer feedback loops. Multimodal Models and Real-Time API (Priority: 4/5): Progress on voice, image, and video models. Challenges in making voice models as intelligent as text models and the benefits of the real-time API. Model Customization: Reinforcement Fine-Tuning (Priority: 4/5): Introduction of reinforcement fine-tuning (RFT) for creating best-in-class models for specific use cases, with examples from financial services and tax. Long/Short Investment Ideas (Priority: 3/5): Personal long and short picks: long on esports and healthcare, short on AI tooling and memorization-based education. Future of Software Engineering (Priority: 3/5): Prediction that more people will code, but not necessarily as professional software engineers. Emphasis on AI-native skills and critical thinking.

Key Arguments: Enterprise deployments succeed with top-down buy-in, a tiger team of technical and subject matter experts, and a focus on evals first. Physical autonomy (self-driving cars) is ahead of digital autonomy (AI agents) due to longer development timelines and existing scaffolding like roads and traffic laws. GPT-5's key improvements are in instruction following, reduced hallucinations, and reasoning, but trade-offs exist between thinking time and latency. Reinforcement fine-tuning (RFT) is more powerful than supervised fine-tuning for creating best-in-class models for specific tasks. Healthcare is the industry most likely to benefit from AI due to large amounts of structured/unstructured data and admin-heavy processes. The number of software engineers may not increase, but the amount of software engineering will grow as more people use AI tools to code.

Data Points: ChatGPT launch date: 2022 - ChatGPT was released in 2022, marking the beginning of widespread AI agent use. Waymo rides: Tens of millions - Waymo has completed tens of millions of autonomous rides. Tesla FSD rides: 3.5 billion - Tesla's Full Self-Driving has accumulated 3.5 billion miles. AI deployment failure rate: 95% - MIT report states 95% of AI deployments fail. GPT-5 thinking time: Up to 10 minutes - GPT-5 Pro can think for up to 10 minutes for complex problems. OpenAI intern productivity: Incredible - OpenAI's first intern class were extremely productive due to AI-native skills.

Pivotal Quotes: "Physical autonomy is ahead of digital autonomy in 2025." — Sherwin Wu: Comparing self-driving cars to AI agents in enterprise. "I think AI agents are like really in day one here. Like ChatGPT only came out in 2022. And the slope, I think, is incredibly steep." — Sherwin Wu: Discussing the early stage of AI agents relative to self-driving cars. "We literally had to bring the weights of the model physically into their supercomputer." — Sherwin Wu: Describing the on-prem deployment of O3 at Los Alamos National Labs.

Implications: Enterprise AI adoption requires deep integration and scaffolding. Healthcare and life sciences are poised for transformation. AI agents will rapidly improve, potentially surpassing physical autonomy. The future of work will involve more people coding, but not necessarily as professional engineers.