Tibo on Coding Agents, Faster AI and OpenAI’s Product Direction

Berman asks Tibo about his time at DeepMind and an internal language-model chat product that predated ChatGPT. Tibo says he worked on infrastructure and research products; early language-model results naturally led to the idea of a conversational tool.

Tibo recalls a creative research environment, but says DeepMind was not set up to ship products at the time. He contrasts that experience with OpenAI, where he says research and product teams collaborate closely and have a bias toward shipping.

Asked whether the early chatbot felt special, Tibo says the first coherent responses initially seemed more amusing than useful, then grew more useful. He learned to think about which elements of a company’s culture to preserve and which obstacles to avoid.

Tibo describes OpenAI as a bottom-up place where people can assemble around an idea and ship quickly. He also wants coherent direction, simplicity and product quality, rather than a collection of disconnected features.

For founders, Tibo recommends conviction, access to users and rapid iteration on feedback. For established companies he adds a harder requirement: be willing to redirect resources toward new ideas even if they could disrupt the existing business.

Berman asks whether that culture gets harder to preserve as a company matures. Tibo says organizations must stay forward-looking because the direction of AI does not wait for their existing arrangements.

Even after training a model, Tibo says, benchmarks do not reveal everything it can do. Teams discover capabilities by experimenting, and those discoveries can change how a product is designed.

As an example, Tibo describes using improved voice and dictation to speak tasks into his phone and have the assistant work with connected tools. He sees the interface improvement as changing the kinds of interactions people can have.

Berman asks which parts of the coding-agent harness still need innovation. Tibo points to the burden of maintaining skill files, imperfect memory and manually managing subagents; in his view these mechanisms interrupt the sense of a seamless assistant.

Tibo’s goal is an assistant that understands a person’s goals, daily work and team context, reacts appropriately and can raise useful ideas proactively. He presents this as a product direction, not as a claim that current systems already do it perfectly.

He argues that a laptop is sized around human constraints, while a future model might work across many applications at once. More capable agents could therefore need computing resources beyond a single laptop.

Faster inference can shift the bottleneck to network latency, tool calls or other parts of the stack, Tibo says. Concurrent exploration, testing and compilation can keep the system productive even when individual operations must wait.

Berman describes the cognitive overhead of running ten or fifteen agents in parallel. Faster responses could let him work with fewer at once; Tibo says managing human attention and choosing when to interrupt are product-design problems.

Tibo imagines fast voice interaction, prototypes and reports that keep a person in the flow. His aspiration is for technology to adapt to the user rather than making the user continually adapt to its mechanics.

Tibo distinguishes two problems: a personal agent tailored to an individual, and automation of a complex process. The first emphasizes a person’s context and collaboration; the second might monitor production logs or help fix security vulnerabilities.

For automated systems, Tibo envisions little human involvement in routine steps while retaining approval for high-risk actions. That is a proposed division of responsibility, not a demonstration that an autonomous system is safe in every setting.

Berman asks about the merging of ChatGPT and Codex. Tibo says some users initially questioned the move, but OpenAI’s intended direction is a shared underlying agent and harness that can handle coding and other work.

Tibo wants the interface to adapt to the individual rather than force a choice between fixed technical and nontechnical identities. He envisions the same broad product serving people with different tasks, connected tools and needs.

When Berman asks about the ideal human interaction, Tibo emphasizes natural language and understanding nuance. His stated goal is to reduce the work users must do to make the system understand what they mean.

Berman asks whether nonverbal signals matter. Tibo envisions an ambient assistant that can follow a whiteboard idea and hold a natural voice conversation, and says voice use was growing at the time of the interview.

Berman mentions a reported 20 million Codex users and asks about competition with Anthropic. Tibo focuses instead on OpenAI’s stated priorities: capable, efficient models and products distributed broadly to different kinds of users.

Pressed on differences in positioning and branding, Tibo emphasizes community feedback and OpenAI’s mission. His answer describes his company’s own outlook rather than giving a detailed comparison of the two companies.

Berman asks whether usage-limit resets are a growth tactic or a gesture to developers. Tibo says they began as compensation when the team broke or misconfigured something and users received a suboptimal experience.

Tibo says he can trigger a reset when it feels appropriate and that the process is not coordinated with marketing or finance. He also describes occasional extra usage as a way to celebrate or let users try a feature.

Berman connects resets to compute planning and asks how much of Luna’s reported price reduction came from efficiency rather than commercial strategy. Tibo does not provide a numerical breakdown; he discusses advance capacity planning and engineering gains.

Tibo says advanced models can help engineers restructure the serving stack and improve both cost and speed. He reports that ordinary operation had become roughly 60% faster over three months, a company claim from the interview period.

He says OpenAI aims to share substantial efficiency gains with customers rather than retain all of them. Discussing compute allocation, he says the improvements enabled more throughput within a similar compute envelope.

Berman asks whether using one model to improve another is early recursive self-improvement. Tibo applies the term broadly to models helping improve inference infrastructure, kernels, products and the ways people extract utility from models.

Asked about a reported pause in frontier reinforcement-learning work, Tibo says safety and alignment become more important as capabilities increase. He describes a pause as giving teams time to understand and harden the system before restarting.

Berman asks whether there was a predefined restart target. Tibo says the safety team’s process involved discussion and discovery and reached a fairly clear set of principles; he does not provide a detailed threshold in the interview.

On ultra-fast inference, Tibo says OpenAI uses it for high-stakes situations such as incident response, where seconds matter. Teams may also request it when a time-sensitive product decision depends on trying an idea quickly.

Both speakers mention ADHD but describe different preferences: Tibo often enjoys context switching, while Berman prefers focusing on a few tasks. Tibo says fast responses can nevertheless help him stay in the flow when he wants to concentrate.

Tibo says faster token generation helps most when an agent produces lots of code with few tool calls. If network or tool overhead dominates, the end-to-end speedup can be much smaller than the generation-speed multiplier.

Asked whether employees always use the fastest mode, Tibo says OpenAI limits internal use and reserves most capacity for customers. This is his description of the company’s allocation at the time.

For external users, Tibo is excited by nontext collaboration: generating alternatives on a shared canvas, selecting an image or steering a prototype in real time. He expects rapid feedback to make creative iteration more fluid.

Tibo predicts that very fast speeds could become close to standard within a year or two as hardware and token efficiency improve, while a more expensive higher tier might remain. This is a forecast, not a commitment or guaranteed price.

Berman asks how to reassure people worried about employment, environmental effects or AI generally. Tibo responds mainly by pointing to improving efficiency, lower costs and wider access; he does not fully address each concern.

For someone hesitant to try AI, Tibo suggests learning from others’ practical uses. He describes using health-related tools to become better informed before seeing his doctor, presenting it as his own experience rather than a substitute for professional care.