How OpenAI operates: hiring, shipping fast, and the GPT opportunity

Executive overview

Most companies slow down as they grow because institutional checks accumulate. OpenAI stays fast by hiring specifically for high agency and urgency — people who see a problem and solve it without consensus loops.

GPTs mark the first step toward an agent future, and vertical use cases are where the real opportunity lies for builders. OpenAI focuses on general reasoning; everything domain-specific is open ground.

The edge goes to builders who move beyond chat interfaces and design AI natively into the core product experience.

The board crisis and what it revealed about OpenAI's culture

• The firing happened during the first company-wide break since ChatGPT launched — most employees were already dispersed
• Leadership had maintained high transparency internally, so the event was genuinely shocking — not a symptom of known tension
• By Monday after Thanksgiving, the team was laser-focused and back to work
• Logan's takeaway: the stakes were lower now than they would be in five or ten years; better to stress-test governance early
• Shared adversity — like the board crisis and earlier launches — has been a consistent team-bonding mechanism

How OpenAI ships quickly

• Young company with no accumulated institutional legacy slowing decisions
• Two non-negotiable hiring criteria: high agency (act without waiting for consensus) and urgency (move without being told to)
• Example: the Assistants API was not a top-down roadmap item — a group of engineers saw the developer need and built it themselves
• Slack is the operating system: instantaneous cross-team coordination; Sam Altman's most-used app
• Research teams are intentionally kept small — in a GPU-constrained environment, adding a researcher who doesn't up-level the whole group is a net productivity loss

Planning and prioritisation at OpenAI

• No formal OKRs; planning is high-level (H1/Q1 goals)
• First filter for any decision: does this get us closer to AGI?
• On the developer side, the first question is reliability — new features are deprioritised when the core API experience is unstable
• Revenue is a proxy metric, not the goal — it funds compute, which enables better models
• Ground shifts constantly; the planning process has to accommodate rapid change in the external environment

Prompt engineering: what actually works

• Models respond to context the same way humans do — generic input produces generic output
• Core principle: context is all you need; treat the model like a capable human who knows nothing about you
• Practical tips:
  • Add relevant links, background, and constraints explicitly
  • Use Browse with Bing to give the model live context about a specific person or topic
  • A smiley face marginally increases output quality — the model is trained on human communication norms
  • Telling the model to "take a break and then answer" can improve responses for the same reason
• OpenAI's prompt engineering guide covers additional techniques with empirical backing
• AI systems will eventually auto-expand sparse prompts into high-fidelity descriptions (already happens with DALL-E)

GPTs: what they are and where they're going

• GPTs let you pre-load context, instructions, files, and tools into a shareable custom version of ChatGPT — no coding required for most use cases
• Key unlocks: code interpreter, browsing, image generation, and external API connections (e.g. Notion, Gmail, Zapier)
• Non-developers can now solve complex, domain-specific problems without engineering support
• Monetisation through the GPT Store is coming — usage-based payouts for creators
• GPTs are the primary onboarding path for the next wave of AI users: a narrow, packaged use case is far more compelling than a blank chat interface
• Best current integrations: Canva (top GPT), Zapier (connects all 5,000 automations without code), Universal Primer (Socratic learning)

Where to build: vertical vs. general products

• OpenAI focuses on general reasoning, coding, writing — not vertical applications
• Building a general-purpose assistant that competes with ChatGPT requires a radically differentiated angle; otherwise you're competing against sustained R&D investment
• Safe territory for builders: domain-specific tools with proprietary knowledge (e.g. Harvey for legal AI)
• OpenAI will launch general-purpose agent products — do not be surprised; do not build against that assumption
• They will not launch vertical products (e.g. AI sales agents) — that space is open

The biggest product opportunities right now

• Move beyond chat: consumers will favour AI experiences that aren't chatbots — canvas interfaces, infinite workspaces, ambient agents
• Build for the GPT-5 world, not the GPT-4 world — don't design around current model limitations
• GPT-5 will be a better tool, not a paradigm shift; problems that exist today will still exist, solved more effectively
• The edge goes to companies that assume users will normalise AI capabilities quickly — plan for adoption, not amazement
• Embeddings are underused: the new V3 model is 5x cheaper, far stronger on non-English languages, and enables grounded Q&A over any knowledge base (e.g. 62,000 pages of text for $1)
• Internal GPTs for enterprise teams — domain-specific prompt templates shared inside a company — are an immediate high-value use case

Using AI inside your company today

• Engineering is the highest-leverage starting point — at least 50% productivity improvement on well-scoped tasks
• Custom GPTs let teams encode company voice and nuance that generic ChatGPT lacks
• Real examples: GPT for drafting Facebook/Google ads; GPT that interprets A/B experiment results like a data scientist
• ChatGPT Teams and Enterprise add SSO, higher limits, internal GPT sharing, and data privacy controls
• Harvard Business School study on AI-augmented consulting shows order-of-magnitude efficiency gains for users vs. non-users