OpenAI DevDay: Beyond the Headlines with Logan Kilpatrick, OpenAI's Dev Relations Lead

Episode Summary

Executive Summary: OpenAI’s Logan Kilpatrick discusses the company’s major developer-day releases and argues they expand, rather than eliminate, startup opportunities by lowering the cost and effort of building AI products. He frames ChatGPT and the API as ways to expose society to AI iteratively on the path to beneficial AGI, while detailing new models, modalities, GPTs and Assistants, safety constraints, and the remaining reliability and autonomy challenges for agents.

Main Topics: Startup opportunity and defensibility: Kilpatrick rejects the idea that the releases simply kill startups: thin API wrappers may be vulnerable, but lower barriers, lower prices, and new capabilities enable many more products. He presents custom models trained on proprietary data as one route for well-funded companies to build defensible businesses. OpenAI’s AGI mission and gradual deployment: OpenAI’s products and APIs are described as mechanisms—not ends in themselves—for developing beneficial AGI and letting people experience, adapt to, and shape AI before more powerful systems arrive. GPTs are characterized as an early step toward agents rather than fully autonomous systems. Agents: capability, reliability, and readiness: Current GPTs do not independently plan or run long-lived tasks in the background. Kilpatrick says agents remain difficult to trust and often require as much effort as doing tasks manually; product reliability, user consent, and the readiness of websites and services for large numbers of agents remain obstacles. Models, fine-tuning, and custom-model access: The discussion covers GPT-4 Turbo in preview, 3.5 Turbo fine-tuning with a 16K context window, planned GPT-4 fine-tuning access, and high-cost custom models. OpenAI is seeking feedback before general release and advises developers to establish fine-tuning experience before applying for GPT-4 access. GPTs, Assistants API, and platform expansion: GPTs build on plugins by combining instructions, uploaded knowledge, and actions, while Assistants provide API access to tools such as retrieval, code execution, and function calling. The conversation highlights evolving controls for authentication and user approval, alongside current limitations in integrations and retrieval customization. Multimodal APIs, pricing, and reproducibility: The releases include vision, Whisper Large V3, text-to-speech, reproducible outputs using seeds and system fingerprints, and upcoming log probabilities. The hosts emphasize lower API prices and possible applications, while noting tradeoffs such as image resolution, voice-cloning features not offered, and the need for a more integrated speech-to-speech API. Safety, privacy, and responsible use: Kilpatrick points to existing safety work on individual capabilities, including a GPT-4 with Vision safety report. He says OpenAI does not train on API or Assistants data by default, describes copyright protection as a way to ease customer concerns, and notes that generated voices must be disclosed as AI voices.

Key Arguments: The releases are more likely to create opportunities for startups than to eliminate them: businesses that are only thin wrappers around an API are fragile, while companies that build distinctive products and defensible capabilities can use the improved platform to compete. OpenAI sees public products and APIs as part of a gradual path to AGI: giving people experience with AI and influence over its behavior may reduce the disruption of a future powerful system arriving without prior public exposure. Agentic systems should not be scaled indiscriminately before they are reliable and the surrounding internet is prepared. Users need visibility and control over consequential actions, and current agents may add risk without saving effort. Fine-tuning and custom models can help companies adapt models to specialized needs. More context and better instruction following may narrow the performance gap between GPT-3.5 and GPT-4, while costly custom training may offer a proprietary-data advantage to organizations able to invest. GPTs are positioned as a more accessible evolution of plugins: users can combine custom instructions and knowledge with actions, without building a full developer-facing plugin experience. Broader, easy-to-enable integrations remain a desired future improvement. AI adoption depends on trust as well as raw capability. Data-use assurances, copyright protections, safety testing, and clear disclosure of AI-generated voices can reduce barriers for individuals and enterprises. OpenAI’s research roadmap includes planned bets, but the products that follow depend on which research efforts yield breakthroughs; Kilpatrick says both the direction and the outcomes are not fully predictable.

Data Points: GPT-4 Turbo preview request limit: 100 requests per day - The hosts mention this limit and Kilpatrick explains that the preview is an early release candidate intended to gather feedback before general availability. Internal 3.5 Turbo evaluation improvement: 38% - A cited example describes improvement on format-following tasks such as generating JSON, XML, and YAML; Kilpatrick says more benchmark information is desired but does not promise a specific release plan. ChatGPT weekly active users: 100 million - Kilpatrick uses this figure to illustrate why deploying agents at ChatGPT scale could overwhelm websites and services that are not ready for large numbers of automated users. OpenAI employee count when Kilpatrick joined: Around 300 - Kilpatrick estimates the company had about 300 employees when he joined and says it has since grown significantly, without giving a current headcount. 3.5 Turbo fine-tuning context window: 16K tokens - The newly announced fine-tuning option supports 16K context; the hosts discuss how this may allow longer inputs and more detailed training examples. Custom-model investment: Approximately $2–3 million - The hosts discuss this as the anticipated entry-level price for custom models; Kilpatrick frames the offering as an investment for companies with substantial resources and proprietary data. Custom-model training data: On the order of billions of tokens - Kilpatrick says custom models may be trained on a scale of data that other companies do not have access to, potentially helping create a defensible business. Low-resolution vision input: 512 × 512 pixels - Kilpatrick says low-resolution processing reformats images to this size and is suitable when the relevant detail remains visible at that resolution. Assistant file-retrieval storage price: $0.20 per gigabyte per assistant per day - The host cites this price while asking whether billing is prorated for smaller amounts of storage; Kilpatrick says he is unsure and would check. Text-to-speech API pricing: $0.015 per 1,000 characters for standard audio; $0.03 per 1,000 characters for HD audio - The host compares these announced rates with higher prices he observed from a competing provider; the comparison is the host’s estimate.

Pivotal Quotes: "The barrier to entry for new startups and for people to build extremely compelling products has never been lower, and there's never been more amazing technology to build with." — Logan Kilpatrick: His central response to the claim that OpenAI’s releases will kill startups: he argues they also enable new companies and products. "OpenAI is not, we don't exist to make the best chat platform. We don't exist to make the best API. Like, these are mechanisms in which help get us to the point of AGI that benefits humanity." — Logan Kilpatrick: He explains how ChatGPT and the API fit into OpenAI’s stated AGI mission. "I've tried a bunch of the agents' products, and I'm just like, this basically takes as much work as if I was going to do it myself." — Logan Kilpatrick: He describes the current gap between agent demos and dependable practical value, alongside concerns about risk and reliability.

Implications: AI platforms are lowering the cost of experimentation, but durable companies will need more than a thin interface around a model. Enterprises may invest in proprietary data and custom models, while developers should design for privacy, user control, and reliable tool use. Broad agent deployment will depend on both technical maturity and internet infrastructure.