Lenny's PodcastInside OpenAI | Logan Kilpatrick (head of developer relations)
CHAPTERS
- 0:00 – 3:57
High-agency, high-urgency: the core hiring traits behind OpenAI’s speed
A cold open sets the tone: Logan argues that hiring for high agency and urgency is the biggest lever for building quickly. He explains what it looks like when people don’t wait for consensus and instead move directly from customer signal to solution.
- •High agency enables teams to act without needing broad consensus
- •Urgency converts customer feedback into immediate action
- •Problem-solving is optimized when individuals feel empowered to ship
- •Culture fit is framed as a practical execution advantage
- 3:57 – 8:20
Inside the OpenAI board/Sam weekend: stress, transparency, and rapid recovery
Logan shares what the dramatic leadership weekend felt like internally and why it was so surprising. He highlights OpenAI’s transparency norms, how the team rallied, and why the company returned to execution unusually fast.
- •The timing collided with a rare planned break for the company
- •Shock stemmed from trust in leadership and a normally transparent culture
- •In-person regrouping in SF, with an unexpectedly quick return to work
- •Perspective that it was fortunate the event happened while stakes were lower
- •Shared events (ChatGPT, GPT-4, DevDay, this incident) create cohesion
- 8:20 – 9:53
New AI interfaces beyond chat: Rabbit R1, TLDraw, and “infinite canvas” UX
Logan describes what’s exciting him most: new interaction paradigms that move beyond chat. He argues that chat is just the first dominant interface and predicts experimentation with multimodal and spatial/infinite-canvas workflows.
- •Consumer devices and novel UX are expanding how people use AI
- •Infinite canvas experiences may match human cognition better than chat logs
- •Multimodality is positioned as a major theme for 2024
- •Opportunity: product teams should rethink UI, not just model capability
- 9:53 – 12:59
How founders avoid competing with OpenAI: go vertical or be radically different
Lenny asks how builders can choose product directions that OpenAI won’t replicate. Logan draws a line between OpenAI’s focus on general capabilities and startups’ advantage in domain-specific workflows, data, and interfaces.
- •OpenAI prioritizes general reasoning/coding/writing; verticals can specialize
- •Example of vertical differentiation: legal tooling like Harvey
- •General-purpose assistants/agents may compete directly with ChatGPT/GPTs
- •If building general, it must solve major unmet problems vs. incumbent UX
- 12:59 – 18:35
Making companies more efficient with GPTs: internal tools for planning, ads, and analytics
The conversation shifts to practical efficiency gains inside organizations. Logan and Lenny share examples of internal GPTs that reduce workload for marketing, experimentation analysis, and quarterly planning rigor.
- •Studies (e.g., consulting/academic) suggest measurable productivity gains
- •Engineering is highlighted as a high-leverage use case (esp. “low-hanging” tasks)
- •GPTs help add company-specific nuance and voice vs. generic ChatGPT output
- •Examples: ad-creation GPT, experiment-results GPT, OKR/quarter planning GPT
- •Planning GPT prompts for metrics, timeline, and stakeholder mapping
- 18:35 – 22:12
Prompt engineering: why it matters, and why “context is all you need”
Logan reframes prompt engineering as effective communication—something humans already do with each other. He explains why models return generic output without context and previews a future where systems automatically expand low-fidelity prompts.
- •Models eagerly answer even when they lack key details
- •Generic prompts yield generic responses (“crap in, crap out”)
- •Human analogy: better questions produce better answers
- •Future direction: automatic prompt expansion (similar to how DALL·E elaborates prompts)
- •Core rule: supply the missing context the model can’t infer
- 22:12 – 26:03
How to write better prompts: add sources, constraints, and even small nudges
Using Lenny’s interview-question prompt as an example, Logan explains how to improve results by providing richer inputs (links, background, artifacts). They discuss prompt “tricks,” why they sometimes work, and how marginal gains can matter for long outputs.
- •When the model lacks public context, you must supply it (links, excerpts, notes)
- •Browsing/tools can help gather and ground the prompt in real material
- •The model won’t always tell you it lacks context—it will still answer
- •Small prompt tweaks may change performance slightly (e.g., tone cues, “take a break”)
- •Referencing OpenAI’s prompt engineering guide for patterns and examples
- 26:03 – 31:54
What GPTs are and why the GPT Store matters: packaging context + tools into shareable apps
Logan explains GPTs as a way to turn a good ChatGPT workflow into a reusable, shareable custom assistant. He outlines built-in tools (files, browsing, image generation, code interpreter) and the trajectory toward easier API integrations and monetization.
- •GPTs replace “share a chat link” with a reusable configured experience
- •Capabilities: custom instructions, file uploads, browsing, image generation, code interpreter
- •Advanced: connect external APIs (Notion, Gmail) to take actions
- •Monetization is positioned as a major upcoming unlock
- •Non-developers gain power by packaging context and workflows
- 31:54 – 42:33
How OpenAI ships fast: culture, decision principles, and Slack-first coordination
Logan compares OpenAI with large institutions like Apple and NASA and explains how a younger org moves faster. He emphasizes hiring, mission-driven prioritization, reliability as a core dev-platform principle, and real-time communication as an execution multiplier.
- •Speed advantages come from fewer legacy process constraints
- •High agency + urgency are treated as foundational operating principles
- •Example: developers’ demand for abstractions led to the Assistants API
- •Planning exists (H1/Q goals), but priorities shift as the landscape changes
- •Decision criteria: mission alignment, and reliability before new capabilities
- •Slack-heavy, real-time communication accelerates cross-functional work
- 42:33 – 44:47
Team size, growth, and why research stays small in a GPU-constrained world
Logan shares public headcount context and discusses how OpenAI scales without choking innovation. He explains why adding researchers can reduce throughput when GPU capacity is the limiting factor, while product engineering scales differently.
- •Publicly cited size: ~750–780, with rapid ongoing growth
- •Hiring spans engineering and product roles, especially infrastructure
- •Research group intentionally kept small to preserve experimental velocity
- •GPU constraints mean more researchers can slow everyone’s iteration cycles
- •Different scaling dynamics between research and product engineering
- 44:47 – 47:37
What’s next: multimodal ChatGPT, agent-like GPTs, and onboarding the next hundreds of millions
Logan outlines near-future directions: richer modalities, more capable GPT experiences, and a path toward agents that work asynchronously. He argues GPTs also solve the blank-slate onboarding problem by packaging narrow, useful experiences.
- •ChatGPT shifting from text-only toward voice, images, and camera inputs
- •GPTs framed as an early step toward agents that can “go do work” and report back
- •Asynchronous, longer-running tasks are closer to how humans delegate work
- •GPTs help new users by offering specific, packaged use cases
- •Vertical, narrowly-scoped assistants are a gateway to broader AI adoption
- 47:37 – 50:39
GPT-5 expectations: predictable scaling, practical gains, and avoiding hype-driven product plans
Lenny asks how builders should design for a GPT-5 future. Logan stresses that improvements will be meaningful but will normalize quickly, and the best product opportunities remain grounded in real user problems—especially vertical workflows.
- •GPT-4 introduced more predictable capability scaling based on compute
- •GPT-5 likely improves speed and capability but won’t be “magic” in the way hype suggests
- •Model leaps feel world-changing at first, then become normal fast
- •Better models amplify existing product opportunities, especially domain-specific ones
- •Correct framing: a stronger tool to solve the same real-world problems
- 50:39 – 1:08:06
Enterprise + new releases + builder advice: sharing internal GPTs, embeddings v3, and moving beyond chat
The final stretch covers OpenAI’s business offerings, upcoming product updates, and where builders should focus. Logan highlights enterprise controls and internal sharing, discusses GPT-4 Turbo updates and embeddings v3, and encourages teams to build AI-native experiences beyond chat.
- •B2B options: API, ChatGPT Teams, and ChatGPT Enterprise (SSO, controls, security)
- •Key enterprise unlock: sharing internal GPTs/prompt templates safely
- •Product updates: GPT-4 Turbo refresh (addressing “laziness”), embeddings v3
- •Embeddings: power retrieval/grounding; improved non-English performance; significantly cheaper
- •Builder guidance: design AI-native workflows (not just “a chatbot on top”) and invest in deeper UX
- •Lightning round: books, favorites, interview question, and “measure in hundreds” motto
