Key Moments
AI Dev 26 x SF | A Fireside Chat with OpenAI's Marc Manara
DeepLearning.AIWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
Software engineering isn't dying, but engineers will manage AI agents instead of writing basic code. This expands the definition of an "engineer" to anyone producing software, leading to an explosion of new creators.
Key Insights
OpenAI partners with startups like Cursor and Augment to help them optimize their use of new models, working with them on 'harnesses' that are trained alongside the models for better performance.
GPT-5.5 demonstrated a significant step-change in intelligence and token efficiency, reducing "thinking tokens" and output lengths to lower latency and cost without sacrificing quality.
The ability of AI to write code and operate over long trajectories (e.g., 30 minutes to hours) is becoming very good, but inferring ambiguous user intent and selecting the correct tools from a large pool remain more brittle areas.
The next big unlock is achieving greater trust in AI agents to run end-to-end for hours without stopping or making wrong decisions, enabling a "set it and forget it" workflow for developers.
Startup teams are becoming smaller, with an emphasis on hiring exceptional, cross-functional individuals who combine strong product sense with deep technical skills, and who embrace rapid iteration.
Software companies (like DoorDash, Canva, Figma) are moving fastest in AI adoption, but surprisingly, the legal and healthcare sectors are also showing rapid adoption due to unstructured data and reasoning needs.
OpenAI's collaborative approach with AI startups
OpenAI actively collaborates with startups building on their AI models, such as Cursor, Augment, and Cognition. Marc Manara's team focuses on helping these companies move beyond naive model usage to optimal performance tuning. This involves working closely with them, sometimes before a new model launches, to test and refine its capabilities. They help tailor the "harness" – the software layer that interacts with the model – to work synergistically with the model's training, creating a virtuous cycle where both improve. This partnership extends to solving practical challenges for startups, like reducing latency without sacrificing performance or handling ambiguous user instructions more effectively, ensuring end-users get the best possible experience.
Optimizing AI models for coding and user experience
OpenAI dedicates significant effort to making its models excel at coding, using benchmarks like Sweetbench Pro to measure raw intelligence. However, they recognize that end-user experience involves more than just accuracy. For instance, with GPT-5.5, a key focus was on "preambles" or "thinking tokens" – enabling agents to communicate their reasoning and trajectory to the user during complex, multi-tool operations. This enhances user understanding and trust. Another major optimization is token efficiency. GPT-5.5 achieved better intelligence with fewer tokens, leading to lower latency and cost without compromising output quality. This focus on efficiency is crucial for making AI tools responsive and affordable for widespread adoption.
Progress and brittleness in agentic AI workflows
AI's ability to write code and execute tasks over extended periods (minutes to hours) is rapidly improving, with capabilities now considered "very good." However, certain areas remain more "brittle." One such challenge is inferring ambiguous user intent, especially from non-technical users who provide simpler prompts. The AI must accurately guess the user's true goal and act on it without overstepping or underdelivering. Another brittle area is tool selection, where models need to intelligently choose the right tools from a vast pool, a problem OpenAI is addressing with features like "tool search." Successfully navigating these nuances is critical for building reliable AI agents that can handle complex operational tasks.
The quest for greater AI trust and autonomy
The "next big unlock" in AI, according to Manara, is achieving a higher level of trust in AI agents to operate autonomously over much longer durations, potentially for hours. Currently, developers might not fully trust an agent to complete an entire end-to-end task without intervention, stopping early, or making incorrect decisions. As AI agents become more capable of extended operation, the bottleneck is shifting from AI execution to human cognition – understanding and directing the numerous simultaneous tasks being performed. This "set it and forget it" capability is key to unlocking new levels of productivity and enabling developers to focus on higher-level strategy.
The shift from 'how to build' to 'what to build'
The landscape for startups has fundamentally changed from struggling with 'how to build' to focusing on 'what to build.' Manara, a former founder, emphasizes that the speed of iteration is now paramount. Unlike a decade ago, where building software was a slow slog, today's tools allow for extremely rapid cycles of development, market testing, and adaptation. This rapid iteration is no longer just an advantage but a necessity in a competitive environment. The value of "updating quickly," even if imperfect initially, is preferred over lengthy deliberation, enabling founders to learn from real-world usage and pivot swiftly.
The rise of smaller, cross-functional startup teams
Modern AI-native startups are observed to have remarkably small teams, often comprising only five to ten people, yet achieving significant revenue (tens of millions ARR). This lean structure aims to avoid the bureaucracy that hinders larger organizations. Hiring focuses on exceptional, "swarm" tactics to recruit individuals with a potent blend of strong product sense and deep technical skills. Roles are less siloed, with greater blending of product, engineering, and design. The ideal candidate can leverage AI tools for heavy lifting while possessing the technical understanding to guide and rein them in, spending more time on conceptualizing what to build and less on the granular coding aspect.
Overcoming speed bumps in AI adoption
While speed is a startup's primary advantage, challenges like code reviews, security, governance, and compliance can slow down adoption. Many startups address these "just-in-time" as needs arise. Larger, digital-native companies are surprisingly aggressive in adopting AI, often driven by top-down leadership buy-in and a culture of experimentation. They provide tools and freedom, encouraging workforce-wide AI integration to maintain pace with startups. While overcoming ingrained bureaucracy and legacy systems is tough for enterprises, innovative strategies like automated code reviews and human-in-the-loop processes can enable faster shipping, preventing them from losing ground to agile startups.
The future of the software engineer role
Manara firmly believes software engineering is not dying but is undergoing a significant transformation. The abstraction layer is moving upwards, shifting engineers from managing programming languages to managing multiple autonomous AI agents. This evolution expands the definition of an "engineer" to encompass anyone capable of producing software using these advanced tools. Consequently, instead of a decrease in demand, the author predicts an "explosion" of individuals creating and maintaining software. While traditional computer science education may need to adapt, the foundational understanding of software development will remain crucial for this burgeoning field.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Studies Cited
●Concepts
●People Referenced
AI Native Startup Strategy & Team Building
Practical takeaways from this episode
Do This
Avoid This
Common Questions
OpenAI partners closely with companies to help them optimize the use of their AI models. This involves working with them before model launches, testing candidate models, and jointly developing parts of their technical 'harness' to ensure the best performance and user experience.
Topics
Mentioned in this video
A company experiencing record-breaking growth in the legal space due to strong demand for AI solutions.
A health system product designed to assist clinicians by saving them time, indicating rapid AI adoption in the healthcare sector.
A company noted for being on top of AI adoption and moving exceptionally fast, characteristic of digital-native scale-ups.
A company identified as a fast-moving, digital-native scale-up that is aggressively adopting AI technologies.
A coding product that allows users, including non-technical ones, to interact with AI for coding tasks. It's mentioned in the context of inferring user intent.
A product used within health systems to give clinicians back time, showing the fast pace of AI adoption in healthcare.
Amazon Web Services, where Marc Manara previously led startup teams.
A company that works with OpenAI, utilizing their models. It's an example of an AI-native company building on OpenAI's technology.
A product used in health systems to help clinicians save time, reflecting the quick adoption of AI in the healthcare industry.
A company that works with OpenAI, using their models. It's mentioned as an example of an AI-native company building on OpenAI's technology.
A recently shipped model from OpenAI that demonstrates improvements in intelligence and token efficiency, focusing on fewer thinking tokens and shorter solution lengths.
A communication platform that Codeex can be used to navigate, as part of its broader knowledge work capabilities.
A company noted for its record-breaking growth in the legal sector, driven by the strong adoption of AI tools.
A coding product mentioned in the context of inferring user intent and handling ambiguous instructions, especially for non-technical users.
A company mentioned as a prime example of a fast-moving, digital-native scale-up that is aggressively adopting AI technologies.
OpenAI's product that is used for coding tasks and broader knowledge work. It is described as an agentic coding product that can be trusted for longer trajectories.
A financial firm whose research suggests a rise in demand for software engineering roles, contrasting with claims of the role's demise.
An AI research and deployment company that builds models for various applications. The company partners with startups to help them optimize the use of their AI models.
A company that partners with OpenAI, utilizing their models. Mentioned as an example of an AI-native company leveraging OpenAI's technology.
A company mentioned as an example of a fast-moving, digital-native scale-up that is aggressively adopting AI technologies.
A benchmark used to measure progress in raw intelligence for coding tasks, assessing capabilities relevant to software engineers.
A benchmark used to assess the raw intelligence of AI models, particularly for software engineering tasks like writing, reviewing, and refactoring code.
More from DeepLearningAI
View all 80 summaries
33 minAI Dev 26 x SF | Carter Rabasa: File Systems Are the New Primitive for AI Agents
28 minAI Dev 26 x SF | Melissa Herrera: Your Agents Should Be Durable
31 minAI Dev 26 x SF | Vlad Luzin: Herding Cats—The Hidden Challenges of Multi-Agent Autonomy
52 minAI Dev 26 x SF | Eli Schilling: Hands On Agent Context & Memory Engineering with Oracle AI Database
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free