Inside OpenAI | Logan Kilpatrick (head of developer relations)
Logan Kilpatrick, Head of Developer Relations at OpenAI, discusses how OpenAI operates, their rapid shipping, and key hiring attributes. He shares insights on prompt engineering, building vertical AI products, and leveraging AI for internal efficiency, along with future opportunities.
Deep Dive Analysis
16 Topic Outline
Inside OpenAI During Sam Altman's Board Drama
Emerging AI Interfaces Beyond Chatbots
Leveraging AI for Company Efficiency
Practical AI Use Cases and GPT Examples
Understanding Prompt Engineering
Tips for Effective Prompt Writing
OpenAI's GPTs and the GPT Store
OpenAI's Hiring Philosophy: High Agency, Urgency
OpenAI's Planning and Prioritization Process
Role of Slack in OpenAI's Fast-Paced Culture
OpenAI Team Growth and Research Constraints
Future of OpenAI: New Modalities and Agents
Building for a GPT-5 Future
OpenAI's B2B and Enterprise Offerings
Latest OpenAI Updates: Models and Embeddings
Opportunities for Leveraging OpenAI Tech in Products
6 Key Concepts
High Agency
This characteristic describes individuals who take initiative and solve problems without needing extensive consensus from many people. They proactively push for solutions to customer challenges rather than waiting for traditional, multi-departmental approval processes.
High Urgency
Complementary to high agency, this refers to the ability to act quickly and implement solutions without delay. It's about seeing a problem and tackling it immediately, rather than postponing action.
Prompt Engineering
This is the skill of effectively communicating with an AI model to elicit the best possible output, similar to how one would communicate with a human. It emphasizes providing sufficient context to the model, as it lacks inherent understanding of the user's background, goals, or specific needs.
GPTs (Generative Pre-trained Transformers)
These are custom versions of ChatGPT that allow users to embed specific context, upload files, provide custom instructions, and connect to external APIs. They empower non-developers to solve challenging, domain-specific problems by tailoring the model's behavior and knowledge.
Embeddings
This technology enables question-answering systems to operate on custom documentation or knowledge bases. It involves converting a corpus of knowledge into numerical representations, then finding similarities between user questions and this embedded knowledge to generate grounded, factual answers.
Laziness Phenomenon
A behavior observed in AI models where they provide less detailed or complete answers, often due to insufficient prompting or a tendency to give a basic response to a basic question. OpenAI has worked to improve this in their updated models.
12 Questions Answered
It was a very stressful Thanksgiving week, surprising due to the deep trust in leadership and the company's usual transparency. Despite the shock, the team quickly returned to work, demonstrating their focus on the mission, which Logan attributes to the caliber of the team.
The most exciting developments are new interfaces around AI, like the Rabbit R1 and TLDraw, which push beyond chat as the primary interaction method. These new UX paradigms, such as infinite canvas experiences, offer more intuitive ways for humans to interact with AI.
AI tools, especially for engineering tasks, can provide significant efficiency gains, potentially 50% or more for lower-hanging fruit. GPTs allow companies to build custom, domain-specific solutions that incorporate their unique voice and nuance, solving tactical problems more effectively than general ChatGPT.
OpenAI focuses on general use cases like reasoning, coding, and writing, while vertical applications (e.g., legal AI like Harvey) with domain-specific knowledge are less likely to be disrupted. Companies building general-purpose assistants will directly compete with OpenAI, but those with radical differentiation can still succeed.
OpenAI prioritizes hiring individuals with "high agency" and "high urgency." High agency means people who take initiative and solve problems without needing extensive consensus, while high urgency means they act quickly to implement solutions.
Planning is guided by core principles like advancing towards AGI and ensuring API reliability. While there are H1/Q1 goals, the process is adaptable to the rapid changes in the AI space, focusing on mechanisms to update understanding and goals as the ground shifts.
Success is measured through metrics like adoption, revenue, and the number of developers on the platform. Revenue, for instance, is seen as a proxy for acquiring more compute (GPUs), which directly helps train better models and achieve the mission of AGI.
Future developments include new modalities for interacting with AI (beyond text-in/text-out), the evolution of GPTs towards an "agent future" where AI can take actions and complete tasks over time, and using GPTs to onboard hundreds of millions more people to AI through narrow, vertical tools.
People should build with the understanding that GPT-5 will be an extremely useful and effective tool, solving a new echelon of problems, but it will also become "very normal very quickly." The focus should be on solving specific use cases more effectively, rather than expecting unrealistic, generalized capabilities.
ChatGPT Enterprise offers features like SSO, higher limits, and enhanced security controls, but its coolest aspect is the ability to share custom prompt templates and GPTs internally. This allows companies to collaborate on domain-specific AI applications tailored to their internal information and problems.
Recent updates include an improved GPT-4 Turbo model that addresses the "laziness phenomenon," and a third-generation embeddings model (V3). The V3 embeddings offer state-of-the-art performance, significantly improved non-English language support, and are five times cheaper, enabling more widespread use for grounding models in custom knowledge bases.
The biggest opportunities lie in creating new, intuitive AI experiences that move beyond the traditional chatbot interface. Products that allow users to ask complex questions and receive data-grounded summaries, or that build core experiences from the ground up with AI, will have a significant advantage.
31 Actionable Insights
1. Hire for Agency & Urgency
When hiring, prioritize candidates who demonstrate high agency and urgency. High agency individuals proactively identify and solve problems without needing extensive consensus, enabling faster progress and greater impact.
2. Build Vertical AI Products
When building AI products, focus on specific, vertical use cases rather than general-purpose assistants. This strategy helps avoid direct competition with foundational AI providers and allows for deeper domain-specific value creation.
3. Provide AI Prompt Context
Provide ample context to AI models in your prompts, including relevant background information, links to external resources, or specific details, to ensure more accurate and valuable responses.
4. Leverage AI for Engineering
Utilize AI tools like ChatGPT for software engineering tasks, especially lower-hanging fruit, to achieve significant efficiency gains (potentially 50% or more) in development.
5. Utilize Custom Internal GPTs
Develop and use custom GPTs to address tactical, company-specific problems by incorporating internal nuance and domain knowledge, making AI solutions highly relevant and effective for your organization.
6. Prioritize Product Reliability
Prioritize providing a robust and reliable user experience for your products and APIs. Even the most innovative features are useless if customers cannot dependably use them.
7. Adopt “Measure in Hundreds”
Adopt a “measure in hundreds” mindset for new endeavors, viewing initial failures as early attempts in a long series, to encourage persistence and acknowledge that success often requires numerous tries and compounding effort.
8. Prioritize Quality Sleep
Prioritize and improve your sleep habits, as sleep science suggests it’s one of the most significant ways to enhance personal well-being and performance.
9. Innovate Beyond Chat Interfaces
Develop AI products that go beyond traditional chat interfaces, offering novel user experiences to gain a significant competitive advantage in the market.
10. Plan for AI Normalization
Plan for rapid user adoption and normalization of new AI tools; assuming immediate, radical, and disruptive change can be a disadvantage, while anticipating quick integration into daily routines provides an edge.
11. Augment Human Capabilities with AI
Actively learn and use AI tools to augment your capabilities, as humans augmented by AI will be more competitive in the job market than those who don’t adopt these technologies.
12. Use New Embeddings Models
Leverage the new V3 embeddings models for question-answering systems based on your own documentation, especially for non-English languages, due to state-of-the-art performance, improved multilingual capabilities, and significantly lower cost.
13. Adopt ChatGPT Business Accounts
For businesses, consider using ChatGPT Teams or Enterprise to manage multiple subscriptions, enable SSO, and securely share internal prompt templates and custom GPTs, enhancing collaboration and control.
14. Report AI Bugs with Examples
When reporting bugs or issues with AI models, provide shared chats or tangible, reproducible examples to enable developers to effectively diagnose and fix the problems.
15. Expense AI Tool Subscriptions
Request your company to expense your ChatGPT Plus subscription (or similar AI tools) to ensure you’re using the latest versions and maximizing your productivity with AI.
16. Consult Prompt Engineering Guides
Consult official prompt engineering guides (e.g., OpenAI’s) to learn best practices and nuanced techniques for improving AI model performance.
17. Integrate Zapier with GPTs
For non-developers, explore integrating Zapier with your GPTs to connect with thousands of applications and enable your custom AI to perform a wide range of automated actions without coding.
18. Build Socratic Learning AI
Consider building AI tools that employ a Socratic method or interactive learning approach to help users learn new concepts more effectively.
19. Foster Slack-Heavy Culture
Foster a Slack-heavy culture for instantaneous, real-time communication and cross-functional coordination, as it can be more efficient than in-person interactions for quick alignment.
20. Maintain Small Research Teams
In resource-constrained research environments (e.g., GPU capacity), maintain an intentionally small research team, as adding new researchers can decrease overall productivity unless they significantly up-level existing team members.
21. Prioritize Mission Alignment
When making product decisions, consistently evaluate whether a proposed initiative aligns with and contributes to the core mission (e.g., achieving AGI), using it as a primary filter for prioritization.
22. Adapt Planning to Change
Implement mechanisms for continuously updating your understanding of the market and adapting goals, especially in rapidly evolving fields like AI, to remain agile and responsive to change.
23. Build AI Solutions Now
If you have innovative ideas for AI solutions, now is the opportune moment to build them, as the world needs more practical applications of these tools.
24. Try Weighted Sleep Mask
Consider using a weighted sleep mask (e.g., from “Manta Sleep”) to improve sleep quality, as the added pressure can be beneficial.
25. Try OAWAAWAAW Sleep Mask
Try the OAWAAWAAW sleep mask, which is designed with ample eye space to prevent pressure on eyelids and eyelashes, potentially improving comfort and sleep quality.
26. Build Radically Different General AI
If you aim to build a general-purpose AI assistant, ensure it offers a radically different and superior solution to multiple existing problems to effectively compete with established platforms.
27. Create Private Planning GPT
Build a private GPT for personal or team planning (e.g., OKR planning) to ensure consistency in framing, focus on individual metrics, and avoid common planning pitfalls.
28. Prepare for GPT Monetization
Prepare for future monetization opportunities within the GPT store by building valuable custom GPTs, as creators will soon be able to earn revenue based on usage.
29. Use Narrow Vertical GPTs
When introducing AI to new users, focus on packaging GPTs as narrow, vertical tools that solve one very specific problem well, making the value immediately clear and encouraging broader adoption.
30. Design Data-Grounded AI Answers
Design AI products that allow users to ask natural language questions and receive data-grounded summaries or insights, moving beyond dashboard filters to provide direct answers to complex queries.
31. Experiment with Prompt Cues
Experiment with subtle cues like adding a smiley face or instructing the model to “take a break” before answering, as these human-like interactions can sometimes yield small performance improvements due to training data patterns.
5 Key Quotes
You can take on the world if you have people who have high agency and not needing to get 50 people's different consensus.
Logan Kilpatrick
Context is the only thing that matters.
Logan Kilpatrick
Nobody cares if you have something great, if they can't use it robust and reliably.
Logan Kilpatrick
I measure in hundreds, the five times that you failed at something you failed and tried zero times.
Logan Kilpatrick
It's not AI that's going to replace humans. It's like other humans that are being augmented and like using AI tools that are like going to be more competitive in a job market and stuff like that.
Logan Kilpatrick
1 Protocols
Personal Planning GPT
Logan Kilpatrick- Take suggestions from an article on best ways to set up for planning success.
- Put those suggestions into a private GPT.
- When doing any planning (e.g., for a new project), feed the plan into the GPT.
- Have the GPT generate a timeline, specific metrics and success criteria, and identify important cross-functional stakeholders.
- Use the GPT to ensure consistency in framing and to drive back to individual metrics, addressing common planning misses.