Tibo 访谈：极速模式、重置机制与 OpenAI 的 Agent 路线

Brief Description

In this conversation, Tibo discusses lessons from DeepMind and OpenAI, the product direction behind ChatGPT and Codex, and why faster models change the way people work with agents. He describes a future in which a personal AI adapts to an individual rather than asking the individual to adapt to software, while also addressing model efficiency, safety, broad access, and OpenAI's practice of compensating users when a product experience falls short.

Table of Contents

• From DeepMind research to a bias for shipping
• Building an empowered product culture
• Personal agents, harnesses, and cloud-scale work
• Managing attention with faster agents
• ChatGPT and Codex as one adaptive product
• Natural and ambient interaction
• Competition, community, and usage resets
• Compute capacity and recursive improvement
• Safety pauses and operational judgment
• What ultrafast mode changes
• Efficiency, access, and the case for trying AI

From DeepMind research to a bias for shipping

Matthew: Before ChatGPT changed the public conversation, Google had an internal language-model chat product. What was it like to work on products that seemed to anticipate that moment?

Tibo: DeepMind was an exceptionally creative place. My work was largely on infrastructure and products that accelerated research, while another group was advancing language models and scaling them. Once the results became strong enough, it was natural to ask whether the models could become something people could chat with and use for many tasks. An internal product resembling LM Chat emerged, and there was an ambition to make it public roughly a year before ChatGPT.

DeepMind was not organized to ship products in the way OpenAI is. At OpenAI, research and product work closely together: they ideate together, co-design, and have a strong bias toward getting useful things into people's hands. That combination of mission, people, talent density, and availability to users was a major reason to join.

The early models felt special. They produced coherent text, and at first the output was more amusing than helpful. Over time, it steadily became more useful. That experience remains a useful reference point for thinking about what parts of a culture to preserve and what failures to avoid.

Building an empowered product culture

Matthew: What lessons from that period shape the culture you want at OpenAI, and what would you tell a founder trying to build a similar culture?

Tibo: OpenAI is bottom-up and empowering. People can bring forward ideas, form a group, and ship something quickly with relatively little stopping energy around new product ideas. That is exciting because it is aimed at positive impact, but empowerment cannot become a hodgepodge of unrelated features. It has to be balanced by simplicity, a coherent direction, and pride in product quality.

The organization invests in delight, performance, efficiency, and simplicity. A product can move quickly without abandoning those principles. For founders, conviction matters, as does finding users early and iterating quickly from feedback. Companies also need the willingness to disrupt themselves: new research and new ideas can deserve investment even when that means reallocating resources away from the established main business.

As companies mature, that can become harder. The future of AI does not wait for the next quarter's established product. Teams need to remain open-eyed about where the wave is going, experiment directly with models rather than relying only on benchmarks, and adjust product thinking when capabilities reveal new ways to help people.

Personal agents, harnesses, and cloud-scale work

Matthew: What parts of the agent harness are still open to innovation as models improve?

Tibo: There is still a great deal to improve. Sophisticated users of coding agents have learned to live with clunky constructs: maintaining skill files, dealing with imperfect memory, and explicitly managing subagents. These mechanisms can be useful, but they also break the feeling that the agent is a seamless partner.

The goal is an agent that deeply understands a person's goals, daily work, knowledge, and team, reacts appropriately, and can be proactive. The interface should preserve that sense of partnership. Voice already points in this direction: natural voice interaction and strong dictation change how people use a model, including by making it easier to give a phone-based agent a set of tasks that can use connected tools.

More capable models also expose the limits of the laptop as the primary work environment. Laptops were designed around human limits: how fast a person types, thinks, and consumes work, and how many applications a person can manage. A model need not share those limits. It could work across many applications at once, which makes cloud resources, parallel work, and a new kind of harness increasingly important.

Managing attention with faster agents

Matthew: If token speeds become dramatically faster, how does a solo developer's workflow change?

Tibo: As models get faster, tool calls, network work, and other stack overheads can become the bottleneck. One response is concurrency: exploration, test writing, compilation, and testing a hypothesis can happen at the same time. But giving people more parallel work is not automatically better. Launching ten or fifteen agents creates cognitive overhead and forces context switching while the person waits for results.

The product should be friendly to attention. It should understand whether an interruption is useful now or should wait, and it should let the person remain in flow while ideas, prototypes, and small reports appear in real time. The technology should adapt to the user instead of requiring the user to adapt to the technology.

There are two broad categories of agent problems. One is the personal agent: an agent in a person's flow, tailored to that person, able to surface useful ideas and handle technical work, research, or advice. The other is full automation: systems that can take on complex processes such as reading production logs, performing performance optimizations, identifying regressions, or helping reduce the time a vulnerability remains open. In those systems, the human may only need to approve high-risk actions rather than remain in every loop.

ChatGPT and Codex as one adaptive product

Matthew: How has the convergence of ChatGPT and Codex been received, and what is the intended end state?

Tibo: There was initial concern about merging them, but the future models call for a unified product. The aim is a highly capable personal agent using the same underlying technology and harness regardless of whether someone is coding. It should be multimodal, voice-first, efficient, and capable across tasks.

People should not need to choose between a coder interface and a nontechnical interface. Labels such as software engineer or designer are useful abstractions, but individuals sit on a spectrum and have their own goals, tools, and ways of working. The same personal AGI could serve a technically sophisticated user and that user's parent, while adapting to different tasks, connected tools, and preferences.

The desired interaction is rooted in human communication. Natural language succeeded because people already know how to use it and can understand not only literal words but emotion and nuance. A good AI should reduce the experience of feeling misunderstood because it missed the intent or tone behind a request. It should feel like a natural extension of how people already act in the world.

Natural and ambient interaction

Matthew: Human communication includes gesture, facial expression, and shared context. How important is that for AI?

Tibo: The future is likely to be ambient and natural. An agent should be able to participate when someone writes an idea on a whiteboard, speaks through a problem in an office, or wants to have a natural conversation over voice. The growth of voice interaction after the launch of newer ChatGPT voice capabilities reinforces a simple lesson: when technology becomes more natural and lowers friction, people tend to choose it.

Typing into a small text box can be natural for some users but not for everyone. Voice, visual context, and other non-text interactions can make the experience easier and more capable. The direction is not merely adding modalities for their own sake; it is making the technology better at understanding what a person is trying to do.

Competition, community, and usage resets

Matthew: How do you view the competitive market, especially comparisons with Anthropic?

Tibo: The focus is on building the most capable and efficient models, then taking pride in products that can reach everyone. Merging ChatGPT and Codex was part of making powerful technology safer and easier to use across roles: product management, design, sales, marketing, communications, and beyond. A broad existing user base can help distribute those capabilities quickly.

Rather than watching competitors closely, the more useful question is what OpenAI can do uniquely well, what its values are, and how quickly it can move toward them. Community matters: taking ideas from users, being transparent, and building technology for the world rather than only for the company creates energy and a grounded sense of purpose.

The well-known usage-limit resets began as a direct response to product failures or a suboptimal experience. When the team was iterating, broke something, or misconfigured an experience, giving users additional usage was a way to acknowledge that users rely on the product and to thank them for trying an early system. The reset button grew out of that principle, not from a marketing or finance process. If the experience is not good enough, the intent is to make up for it.

Compute capacity and recursive improvement

Matthew: Does that generosity require substantial capacity planning, and how do you think about accelerating model progress?

Tibo: Compute planning matters, but so does using capacity well. Better models and more efficient models reinforce one another. Advanced models can help people understand systems, improve code, and work through parts of the research and engineering process. That is a practical form of recursive self-improvement: not necessarily a model independently creating the next model, but capable models helping improve the critical path used to develop future capabilities.

This also changes how products evolve. Teams can use the systems, learn where they help and where they fail, then apply those lessons to the next iteration. The organization needs enough capacity to serve users while still giving internal teams the ability to understand and harden the system.

Safety pauses and operational judgment

Matthew: How should people interpret a pause in training or development when safety questions arise?

Tibo: When a pause is necessary, it gives teams and individuals time to understand and harden all parts of the system. The purpose is to restart only with full command of the situation. OpenAI should be able to make that kind of decision efficiently when it is needed.

The decision is not reducible to a single mechanical threshold. It belongs with the safety team and involves debate and discovery as evidence develops. At the same time, the team can establish clear principles that define when it is in a good position to resume. Capabilities and safety have to be considered together rather than as independent concerns.

What ultrafast mode changes

Matthew: What use cases become possible when token generation is much faster?

Tibo: Ultrafast access is especially valuable in high-stakes moments. During an outage, for example, incident commanders and response teams benefit because seconds matter. Teams working on something critical, or facing a near-term decision about whether to include an idea in an important release, can also benefit from immediate iteration.

The advantage depends on the task and on personal working style. People who prefer to stay focused on one thing can remain in flow when the system responds at the pace of thought. Others thrive on context switching and making many small decisions. Ultrafast generation is most striking when a task needs substantial generation with relatively few tool calls, such as rapidly prototyping a website or game. When the agent trajectory contains many network or tool operations, those external costs limit the end-to-end speedup.

OpenAI does not simply give every employee unlimited ultrafast capacity. Internal use is valuable for understanding and improving the product and for recursive improvement, but most capacity is reserved for customers. Beyond text, fast interaction may enable shared canvases, real-time creation, rapid image exploration, and prototypes that a person can steer through voice or text. Instead of spending a long time only reasoning about a system's requirements and trade-offs, an engineer can sometimes create a version quickly, see how it behaves, and iterate while still in the creative flow.

Efficiency, access, and the case for trying AI

Matthew: Will ultrafast capability remain a premium feature, and what would you say to people nervous about AI?

Tibo: Technology generally becomes more broadly accessible over time. Progress comes not only from raw inference speed but also from models becoming more token-efficient and from continued improvements in hardware. In a year or two, current ultrafast speeds may be close to normal, even if there will always be a higher tier that uses more hardware for extra capability.

Broad access depends on efficiency. A smaller model such as Luna can be remarkably cheap while delivering capability that would have looked frontier-level only months earlier. As serving costs fall, powerful intelligence can reach more people and provide more direct daily utility. The expectation is that today's frontier becomes much cheaper to run over time.

For people hesitant to try AI, the most useful starting point is close at hand. People already use ChatGPT for writing, personal advice, and to become better informed before a doctor's visit. New health and finance experiences can be useful, but they do not replace professional judgment. Seeing how others use the tools and drawing inspiration from their practical benefits is a reasonable way to discover where AI may help in one's own life.