Unsupervised Learning• Dec 10, 2025• 57:37Interview

OpenAI's Head of Product: How the Best Teams Build, Ship and Scale AI Products

From Unsupervised Learning

Olivier Goodman•Head of Product for Enterprise, OpenAI

Executive Summary

OpenAI has released GPT-5.1 and a specialized Codex model, focusing on improving speed to address user feedback on GPT-5's latency for basic queries.
AI models are seeing significant adoption in scientific research, dramatically accelerating tasks like literature aggregation and reproducing complex mathematical proofs, cutting weeks of work down to minutes.
The cost of high-capability AI is rapidly decreasing, with OpenAI seeing a 1-2 order of magnitude reduction in GPT-4 level query costs over 2-3 years, consistently finding that price cuts drive more than offsetting volume increases.
The focus for successful AI implementation is shifting from the model alone to the surrounding 'harness'—the data pipelines, tooling, and infrastructure—which is becoming equally critical for enterprise success.

8 quotes

Concerns Raised

Building reliable, fully autonomous agents remains a very hard problem, likely a decade away from automating entire jobs.
A major blocker for enterprise AI is that most critical operational knowledge is undocumented and exists only in employees' heads.
Current AI voice models have not yet achieved true naturalness, struggling with aspects like interruptions and conversational cadence.

Opportunities Identified

Massive productivity gains in software engineering through specialized coding models like Codex.
Accelerating scientific discovery, particularly in complex fields like physics and life sciences/drug design.
Untapped demand for AI applications that will be unlocked as inference costs continue to fall dramatically.
Providing not just model APIs but also the associated 'harness' and infrastructure as a service to enterprises.

Key Themes

Model Evolution and Specialization

OpenAI's latest release, GPT-5.1, was specifically designed to be faster than its predecessor for common queries without sacrificing intelligence. This release also included GPT-5.1 Codex, a specialized model for coding that has driven a 70% increase in pull requests among OpenAI's own engineers, signaling a trend towards more domain-specific, optimized models.

This highlights the industry's move beyond raw capability to address practical usability issues like speed and to create specialized tools that offer dramatic productivity gains in high-value verticals like software development.

AI as a Scientific Accelerator

Large language models are being increasingly adopted by the scientific community to accelerate research. Use cases range from aggregating literature to testing hypotheses, with one notable example showing GPT-5 Pro reproducing the complex math from a new physics paper in 30 minutes, a task that took the human author weeks.

This demonstrates AI's potential to fundamentally change the pace of scientific discovery, acting as a powerful tool for researchers and potentially leading to breakthroughs in fields like physics and drug design.

The Economics of AI Inference

The cost of using powerful LLMs has plummeted by one to two orders of magnitude in recent years. OpenAI has found that every price cut to its API has been met with a surge in usage that more than compensates for the lower price, indicating massive, untapped demand that is currently gated by cost.

This deflationary trend is a key enabler for widespread AI adoption. As costs continue to fall, previously uneconomical applications (like real-time website personalization for every user) will become viable, unlocking new markets and business models.

The Importance of the AI 'Harness'

The speaker emphasizes a shift in perspective from 'the model is everything' to recognizing the critical importance of the 'harness' built around it. This includes the data pipelines, tooling, and operational infrastructure, which are essential for getting high-quality results, especially in enterprise settings where critical knowledge is often undocumented.

This signals a maturation of the AI development landscape. For businesses, it means that competitive advantage will come not just from accessing the best model, but from building superior systems to feed, prompt, and manage the AI's interaction with proprietary data and workflows.

Enterprise AI Adoption and Frontiers

The speaker predicts 2025 will be 'the year of coding in the enterprise,' with AI tools seeing meaningful adoption. While full job automation via agents is still difficult, strong use cases are emerging in customer support and sales. Future progress in enterprise AI is as dependent on standardizing data infrastructure as it is on model improvements.

This provides a roadmap for near-term enterprise AI strategy, focusing on high-ROI areas like developer productivity and customer experience, while highlighting the foundational data work required to unlock more advanced automation in the future.

Get started free

Topics

GPT-5.1 GPT-5.1 Codex OpenAI Large Language Models AI Coding Assistants Software Engineering Productivity Scientific Research AI in Science Drug Discovery Life Sciences AI Economics API Pricing Cost Reduction Enterprise AI AI Adoption Model Harness AI Agents

Processed Apr 3, 2026 yt-dlp + mlx-whisper + Gemini