Overview of OpenAI's Plan To Merge Chat And Agents — With Greg Brockman
This conversation with OpenAI cofounder Greg Brockman focuses on the company’s long-term product vision: moving from a chat-based AI experience to a unified, agentic system that can understand your goals, access context, and take actions on your behalf. Brockman argues that the future of AI is not just better conversations, but a “personal AGI” that can work across tools, devices, and workflows with minimal interface friction. The discussion also covers model scaling, compute scarcity, pricing, enterprise adoption, voice, and the biggest near-term impact areas such as software, business operations, science, and health.
The Big Idea: Chat Is Evolving Into Agents
Brockman’s core point is that ChatGPT is only the beginning. The real destination is an AI that can:
- Understand a user’s goal or intent
- Pull in context from tools like email, calendar, Slack, and file systems
- Decide what to do next
- Carry out actions with oversight and trust
He describes this as a shift from conversational intelligence to agentic intelligence. In that world, the interface “melts away” because users will mostly just talk to a persistent assistant that can get work done.
What this looks like in practice
Examples he gave include:
- Organizing an inbox automatically
- Helping schedule meetings or appointments
- Drafting and sending emails based on context
- Managing event logistics, like dietary preferences and seating charts
- Helping with health decisions, including specialist referrals or treatment research
Trust, Context, and Control Are the Key Product Problems
Brockman emphasizes that the hardest part of building agents is not raw intelligence, but trust.
Why trust matters
Users need to know:
- What the AI can access
- What it can do autonomously
- When it should ask for confirmation
- How much responsibility they want to delegate
He argues that trust must be earned, not assumed. OpenAI’s approach is to provide:
- Lots of tools
- Clear permissions
- Oversight and supervision
- Strong control boundaries
That, in his view, is what will make people comfortable letting AI take real actions.
OpenAI’s Product Vision: Unification Over Fragmentation
Brockman says OpenAI is steadily merging its products and capabilities:
- ChatGPT
- Codex
- Browser/computer interaction
- Tool use and connectors
The long-term goal is a single, unified assistant that can work across many tasks instead of making users switch modes or apps. He sees this as a simplification of the entire software experience.
Not an OS in the traditional sense
He pushed back on the idea that OpenAI is simply building a new operating system like iOS. Instead, he framed it as a new layer of interaction between humans and technology:
- More like an always-available assistant
- Less like a folder/app-based computer interface
- Closer to how humans work with coworkers or executive assistants
Voice Is a Major Part of the Future
Brockman sees voice as central to making AI feel natural.
Current limitations
Today’s voice systems still feel awkward because:
- They often rely on stitched-together components
- Turn-taking is unnatural
- Interruptions and overlapping speech are hard to manage
Where it’s heading
He expects future voice AI to:
- Process input and output simultaneously
- Allow interruptions and follow-ups naturally
- Feel more like a human conversation
- Be useful both personally and professionally
He also noted that voice is especially powerful for work, where speaking feedback is often much easier than typing long explanations.
Model Improvement: No Wall in Sight
Brockman strongly rejected the idea that scaling laws are running out of steam.
His view on scaling
He argued that progress has continued because:
- More compute still helps
- Better data and architectures still matter
- Apparent “walls” often turn out to be implementation issues, bugs, or mismatches between theory and practice
He pointed to decades of progress in neural networks and said the field has repeatedly shown smooth improvement over time.
The practical constraint is compute
The real challenge, according to Brockman, is not whether models can improve, but whether enough compute, energy, and infrastructure can be built to satisfy demand.
Compute Will Be the Scarce Resource
A major theme of the conversation was that AI demand is growing faster than supply.
Brockman’s view
He believes:
- Compute is becoming the bottleneck
- The market will keep absorbing available capacity
- There is room for more players, but the demand is enormous
- OpenAI is investing aggressively in infrastructure and even chips
He described compute as the critical scarce resource in an AI-powered economy.
Pricing: Cheaper at Today’s Intelligence, Better at the Frontier
Brockman said OpenAI’s pattern has been to:
- Increase intelligence
- Lower price for equivalent capability
His expectation is that:
- Today’s frontier models will become much cheaper over time
- A newer, better frontier model will appear
- Users will naturally move to the more capable version
He also acknowledged that enterprise buyers are now asking more practical questions:
- What is the ROI?
- How do we control spend?
- How do we monitor usage?
- How do we manage observability?
OpenAI, he said, is responding with features like spend controls and more enterprise-focused tooling.
Competition and Differentiation
The discussion touched on competition from Microsoft, Apple, and other model makers.
Brockman’s stance
He does not think the stack will collapse into a single commoditized layer. Instead:
- Compute remains valuable
- Models remain important
- Enterprise workflows and domain expertise matter
- Different industries require different kinds of intelligence and orchestration
He also argued that intelligence is not one-dimensional. Being broadly smart is helpful, but real differentiation comes from:
- Domain expertise
- Workflow integration
- Context awareness
- Customization for specific jobs
The Impact on Science and Medicine
Brockman was especially enthusiastic about AI’s role in health and scientific discovery.
Health use cases
He highlighted examples of:
- Patients using ChatGPT for health questions
- Doctors using reasoning models to solve difficult diagnoses
- People combining diagnostics, model outputs, and domain knowledge to uncover answers
- AI helping with chronic conditions, treatment decisions, and medical research
He said this is already happening and will become standard.
Broader scientific upside
He also mentioned AI use in:
- Chemistry
- Physics
- Drug discovery
- Scientific problem-solving
His view is that every solved mystery reveals many more, meaning the work of science will expand rather than end.
Key Takeaways
- OpenAI’s long-term goal is agentic AI, not just chat.
- The future interface is a persistent assistant that can act, not just answer.
- Trust and control are central to adoption.
- Voice will become much more natural and fluid.
- Model progress is still strong; no scaling wall is visible.
- Compute is the scarce resource shaping the AI economy.
- Enterprise customers are shifting from hype to ROI, controls, and observability.
- AI’s biggest near-term societal impact may be in health, science, and workflow automation.
Notable Insight
Brockman’s clearest framing of the future was that AI should feel less like software you use and more like an intelligent entity you collaborate with:
“You want almost no interface.”
That idea captures the entire direction of the conversation: AI that becomes more capable, more connected, and less visible as a tool—until it simply gets things done for you.