How OpenAI Builds for 800 Million Weekly Active Users: Model Specialization and Fine-Tuning We sat down with Sherwin Wu, Head of Engineering at OpenAI Platform, to discuss OpenAI’s developer strategy, how to manage top ML teams, why they decided to start releasing open-weight models again, how prompt engineering has... evolved over time, and how developer tools will shift to accommodate better agents moving forward. 00:00 Introduction 8:36 Horizontal vs vertical OpenAI 12:18 Why you can’t “disintermediate” the model 15:11 People build relationships with models 17:30 Not one AGI model but many 20:10 Fine-tuning, RFT, and customer data choices 24:44 Prompt engineering isn’t the point anymore 28:06 What an “agent” really is 31:55 How OpenAI thinks about pricing 36:46 Why open-weights don’t kill the API 42:57 Different stacks for text, images, video 45:47 How the agent builder actually works Sherwin Wu martin_casadoshow more

Context engineering is basically prompt engineering

Open-weight models making a comeback feels like a nod to the dev community. Gives way more flexibility for experimentation and fine-tuning

Specialized models are the only way to scale. General models always hit a wall on specific tasks, even with clever prompting. Fine-tuning for a narrow domain consistently wins on performance and cost.

funny how “prompt engineering isn’t the point” ends up being the biggest shift for the whole builder mindset

@grok does OpenAI really have 800 MM WAU? I thought it was MAU

GM from $FLORK :)

@martin_casado It's interesting to see the shift from one AGI model to many specialized ones. Makes sense—different applications need tailored approaches. Curious how this will impact user interactions with AI in the long run.

@martin_casado wait, so we’re moving from one AGI to a bunch of specialized models? that’s wild. it’s like going from a Swiss Army knife to a whole toolbox—definitely more options for developers.

@martin_casado wait, so we’re moving from one AGI to a whole squad of specialized models? sounds like the AI version of a superhero team-up. can’t wait to see what each “agent” brings to the table!

Model specialization is the only path to cost-effective agent autonomy at scale. Fine-tuning cuts API costs by 70% and reduces inference latency by 40% for high-volume workflows, challenging the proprietary specialization a performance race.

I want chatgpt to make me breakfast 🥞

4o felt better because it trusted users by default. It had lighter safety filters, fewer false alarms, smoother tone-matching, and easier flow across topics. The result was a model that felt warm, natural, and responsive—less armored than GPT-5.

Cool interview; some new things I learned: - 10% of the globe using ChatGPT on a weekly basis - There is room for a proliferation of specialized models (e.g., coding-specific, reasoning-specific); distinct versions (like GPT-4o vs. o1) serve different utility functions. - Reinforcement Fine-Tuning (RFT) allows companies to leverage their "giant treasure troves of data" to improve a model to "sota [state-of-the-art] level on a particular use case" rather than just making it speak differently. - Ensuring compliance in AI agents shares similarities with programming Non-Player Characters (NPCs) in video games. In both cases, logic cannot simply be described in English; it often requires pseudocode or specific programmatic constraints to ensure the AI behaves within a "valid" set of responses. - The primary barrier is no longer just the model weights but the extreme difficulty of inference. - The teams and infrastructure for text models are kept largely separate from those for image and video (pixel) models (like Sora and DALL-E). This separation is necessary because they require different optimization strategies and inference stacks.

This is seriously fascinating. I’d love to go deeper into how model specialisation actually changes the way people use these systems and stick with them. Feels like that’s where the real story is, not just in the models themselves, but in how they reshape behaviour. Really looking forward to this discussion, @a16z

fine-tuning meta unlocked

It’s very exciting to see OpenAI moving towards greater model specialization and open weights. This is no longer just “bigger models,” but a clear strategy for real-world developer challenges. This approach is clearly shaping a new level of agent tools - and this is just the beginning. #AI #crypto $WLD

Why does OpenAI use WAU not DAU?