Key Takeaways

  • OpenAI logs over one billion weekly active users in standard chat, but user growth for autonomous tasks on ChatGPT Work is climbing vertically.
  • Brockman argues that forcing users to pick model versions, tune context windows, and adjust thinking strengths represents an interim failure of interface design.
  • The long-term product vision merges reactive chat and persistent autonomous agents into a single system that handles execution in the background.
  • The primary bottleneck for agentic software is discovery: users still treat text inputs as search bars rather than delegating complex, multi-step workflows.

The Death of the Model Picker

For the last three years, building with large language models looked like tuning an engine while driving it. Users had to choose between fast models and reasoning models. They had to count tokens, format context windows, and babysit every step of an output.

Greg Brockman sees that as a temporary tax on early adopters. As he put it during the conversation:

“What we were promised if you were to rewind five, 10 years ago was never a low-level language model. You have to think about context windows and you have to select thinking strength and you have to select different models. Like none of that, none of that is the future.”

Right now, OpenAI supports two parallel modes. Standard chat attracts more than one billion active users every week. At the same time, ChatGPT Work is seeing a steep rise in autonomous usage, where the system executes tasks independently instead of waiting for line-by-line prompts. Brockman notes that having two separate modes is merely “a point in time that is something that we're continuing to unify and merge.” The end state is not a dashboard of technical dials, but software that takes high-level intent and chooses the right execution path on its own.

The Onboarding Problem in Agentic Software

When software moves from reactive to proactive, product design breaks down. When you open a blank chat prompt, you treat it like Google: you type a question, get an answer, and close the tab. You do not ask it to research twenty target accounts, draft personalized memos, and run data validation overnight, because standard interfaces do not signal that capability.

Brockman points out that the real objective of agentic tools is simple: “the whole point is to get your time back, not for you to have to go and become an expert in all these internal details.” The software must understand user intent, execute the low-level steps behind the scenes, and return finished work.

For builders, this requires a total reset on user onboarding. If your application requires a customer to understand prompting syntax, token limits, or reasoning depth to get value, your product is unfinished. Users should set goals, define constraints, and review completed outcomes. Everything in the middle belongs to the model.

What to Do With This

Audit your product's user interface this week. Find every place where you force a customer to pick a model tier, set a temperature, or write a structured system prompt. Strip out those settings and replace them with a single goal-oriented workflow that handles the model orchestration behind the curtain.