OpenAI's Greg Brockman: Why Human Attention Is the New BottleneckOpenAI's
In a Nutshell
OpenAI, led by co-founder Greg Brockman, thrives by aggressively scaling compute amid shortages, leveraging scaling laws and innovations like Transformers to push toward AGI, with models now 80% there—excelling at software engineering, overnight optimizations, and agentic coding that shifts 80% of work from humans. Human attention emerges as the new bottleneck, demanding robust primitives for security, governance, and EQ in agentic workflows, while organizations flatten into solopreneurs and small teams commanding vast AI agents. OpenAI prioritizes enterprise sales, consumer productivity via proactive AGI assistants, and frontier science breakthroughs in physics and biology, envisioning a future where coding evolves into natural, vision-driven oversight rather than typing drudgery.
These notes were generated by AI and may contain inaccuracies.
Greg Brockman was employee number four and first CTO at Stripe, which processes 1.6% of global GDP. OpenAI has almost a billion or more weekly active users. He is co-founder, president, and chief builder at OpenAI.
OpenAI's business: buy, rent, build compute, and resell at a margin. Scale as long as margin is positive because demand for intelligence is unlimited. Current AIs handle any problem. They do not have enough compute and are constantly hunting for more. Matt Garman says GPU compute availability in 2026 rounds to zero. When launching ChatGPT, Brockman told team to buy all compute; demand has outpaced supply ever since.
Scaling laws feel like a fundamental scientific truth, empirical like Newton's laws. Neural networks designed in the 1940s, now scaled with massive compute, continuously improving capabilities with no wall. Constant innovations: micro tweaks like data formatting, larger shifts like LSTM to Transformer (not yet surpassed from 2018 paper). OpenAI leads in long-term research on architectures, algorithms, and paradigm shifts, with fruit on the horizon.
OpenAI has a formal definition of AGI, but intuitions vary. Models are 80% of the way there: smart, very capable. Smarter than Brockman at writing software with context. No one feels better at writing software than GPT-5.4, even writing kernels. Internal results show massive gains on low-level tasks with right setup.
Systems engineer prepared design document for complicated systems optimization, gave to model, went to sleep. Model implemented spec, added instrumentation, ran code, profiled slowdowns, iterated to optimized result overnight.
Lean into tools. Agentic coding went from 20% to 80% of code in December. Codex shifting from software engineers to anyone doing computer work. New features make it more powerful. Announced Chronicle tool for Codex: sees everything on computer, forms memories. Answers questions about recent activity without explanation. Shift to context: include AI in meetings, provide information. One-time investment now in giving AI enough context; models will improve. Constant cycle of improvement, talk to friends on usage.
Live in the future, co-design models and harnesses. Started with software engineering: human accountable for all merged code (structure, maintainability). Going vertical: finance, sales, IT with dedicated teams building domain skills, modifying UI. Externalize once refined. Work with AI-forward customers to define revolution.
Still figuring out. Prototypes cheap (e.g., dashboard in hours vs. week). Bottleneck shifts to sharing. Internal work on easy dashboard/widget/bot building and sharing. Needs governance: IT visibility, data provenance. Example: internal knowledge dumps to wikis; fix permissions on derived artifacts (invalidate if source inaccessible). Build architecture aware of AI usage; changes team relations and bottlenecks.
Decade is long; ceiling hard to internalize. Companies will change. Solopreneurs can build incredible businesses. Jobs easier, more fun, but competitive. Find niche, unique angle. Hierarchies may flatten to small teams doing incredible things. Individuals using GPT-5.4 Pro solve unsolved math problems alone (normally needs team). AlphaGo move 37 made game more interesting for humans; similar for other domains.
Invest in primitives: security, observability, governance. Anecdote: Asked Codex to install package, ran into error, pinged author on Slack, then escalated to manager after 2 minutes. Proactive but lacked EQ; needs checks. Building model EQ: flag high-risk actions, auto-approve low-risk. Human attention is scarcest resource; doing is easy, alignment with values is bottleneck. Build systems accounting for human factor.
Internet security ratcheting up (from '90s viruses/worms). Moving to secure regime with internet-wide effort. Use models to scan codebases, end-to-end red teaming. Leverage trusted access programs, defender community. Models powerful but not magic; part of resilience ecosystem. Incorporate for assurance on patches, quick updates.
New normal from tech trends: more people, lower barriers, easier successes. Play with technology yourself; AI intuitive, adapts to you. Get finger on pulse: changes, possibilities, model lags.
Core mission: empower people thoughtfully, maximize benefits, mitigate risks (esp. cybersecurity, biocurity). Built mitigations, trusted access programs. Models increasingly capable across dimensions. Expanded trusted access for cyber program last week; more should apply. Upcoming expansions. Tune mitigations for broad access with observability, positive deployment. Community effort.
OpenAI decides what to build in the application layer by applying a focused strategy. The field of AI is full of opportunities, but as a single company, they can only do so much despite building compute and hiring people. They prioritize the 80/20 parts of the space for most impact. Currently, they are in an agentic transition. They are serious about enterprise, selling to big companies and building sales motions. For consumer, the focus is on productivity, achieving goals, eliciting goals, and having AI proactively help. This ladders into building an AGI you can talk to with full context for personal and work life—trustworthy for advice on health, finances, career. They made painful decisions on what not to do, pursuing only what fits this singular vision.
In a few years, coding won't involve command lines and agents in the current unnatural state of sitting behind a box typing, causing carpal tunnel and hunched shoulders. Bodies were not designed for this. People want more time with loved ones, talking, coming up with brilliant visions, understanding themselves. Becoming CEO of an organization of 100,000 agents seems good, enabling much more productivity. Mechanics will feel as different as going from writing by hand with a quill to sending a text message with people working on your goals.
Frontier models will push science and physical AI. LLMs follow strong scaling laws for digital intelligence but less so for robotics, physical intelligence, biology, and science, where verification is harder or takes longer.
Science is a domain OpenAI is leaning into with line of sight to incredible progress and signs of life. Grounded in today: a physics result where AI derived a beautiful formula physicists thought impossible after working on it for some time—a significant step toward quantum gravity, bigger than months ago. A year from now could be much further.
Biology involves messy reality outside simulations, but lessons from software engineering apply: competition-solving isn't enough; need exposure to real-world messy codebases, human interruptions, adversarial conditions. Expect a renaissance in science, possibly big results this year, next year totally wild.
Greg and Anna enjoy watching movies, going on hikes—not as much time as hoped, but more post-AGI. Enjoy the ride.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
