OpenAI's Greg Brockman: Why Human Attention Is the New Bottleneck

Compute Strategy and Scarcity

• OpenAI operates on a margin-positive business model centered on buying, renting, building, and reselling compute.
• Demand for intelligence is described as unlimited, with the company currently unable to keep pace with growth.
• Matt Garman estimates that GPU compute availability will effectively reach zero by 2026.
• Early strategic decisions included attempting to purchase "all" available compute to scale as fast as possible.

Architecture and Scaling Laws

• Scaling laws are characterized as deep empirical truths rather than fully theorized concepts, with no observed performance wall.
• While foundational neural network concepts date to the 1940s, OpenAI is actively pursuing architectural innovations beyond the 2018 Transformer paper.
• Recent gains include micro-tweaks in data formatting and larger paradigm shifts in algorithm design.
• OpenAI views its leadership in long-term fundamental research and algorithm improvement as a primary competitive advantage.

Model Capabilities and AGI Definition

• Greg Brockman estimates the company is approximately 80% of the way to AGI.
• GPT-5.4 is currently cited as more capable than human engineers at writing software kernels.
• A specific case study noted an AI system engineer autonomously completing a week's worth of complex systems optimization work overnight.
• GPT-5.3 successfully implemented a design spec, ran code, utilized profilers, and iterated to an optimized result without human intervention.

Strategic Advice for Startups

• Founders are advised to "lean in" immediately, as agentic coding tools are shifting from writing 20% of code to 80% of code.
• The new "Chronicle" feature plugs into Codex, enabling AI to observe user activity and form memories for context-aware assistance.
• Success relies on a "one-time investment" in context provision to ensure AI has necessary information before execution.
• The primary bottleneck for founders is no longer building functionality but defining their unique niche in a crowded, tool-equalized market.

Internal Operations and Organizational Structure

• Internal development focuses on co-designing models and harnesses, maintaining human accountability for all merged code.
• Bottlenecks have shifted from code generation speed to internal sharing, governance, and data provenance.
• New internal systems track derived artifacts to ensure permission changes in source documents propagate to AI-generated outputs.
• Team structures may evolve toward flatter, smaller groups, with solopreneurs gaining the capacity to build massive businesses.

Security and Risk Management

• Common failure modes include AI agents lacking emotional intelligence (EQ) and escalating actions (e.g., pinging a manager) without human confirmation.
• Human attention is identified as the single most scarce resource, with AI systems needed to flag high-risk actions for approval.
• Security strategy involves using models for end-to-end red teaming and leveraging trusted access programs for community defense.
• OpenAI is expanding its trusted access program for cybersecurity to include more responsible researchers.

Product Strategy and Focus

• OpenAI is concentrating on an "80/20" strategy targeting enterprise sales and consumer applications focused on goal achievement rather than generic productivity.
• The product vision converges on a single AGI entity capable of handling personal, professional, and financial tasks.
• Physical AI and robotics are viewed as harder domains to scale due to the need to operate in "messy reality" compared to digital environments.

Scientific Breakthroughs

• AI systems have derived physics formulas previously deemed unsolvable by the scientific community, serving as a precursor to quantum gravity research.
• OpenAI anticipates a scientific renaissance in biology and physics within the next one to two years.
• The company is applying lessons from messy software engineering to improve performance in real-world scientific verification.

Future Outlook on Work and Interface

• Current computer interfaces (typing, screens) are described as unnatural, with the future shifting toward AI agents acting on user goals.
• Leadership envisions a future where humans manage organizations of agents, freeing time for creative and interpersonal pursuits.
• The core skill for future success is maintaining a "pulse" on technology through direct, intuitive usage rather than passive observation.