OpenAI is orchestrating a profound consolidation of its product ecosystem—not merely merging two applications, but aiming to dissolve the very concepts of "ChatGPT" and "Codex" into a single, persistent personal AGI that comprehends users over the long term and executes tasks autonomously.

In a recent interview with tech commentator Matthew Berman, Tibo Sottiaux, who leads the Codex division at OpenAI, confirmed that Codex has been deeply woven into the ChatGPT product suite, progressively opening its developer-oriented coding capabilities to the broader user base. However, he underscored that this integration represents only the initial phase. The endgame Sottiaux envisions is striking: all intricate backend architecture will be entirely concealed, leaving a singular, ultra-minimalist "Personal AGI" that understands you profoundly and can instantly marshal vast cloud computing resources.

During the conversation, Sottiaux also disclosed several data points and projections with market relevance: Codex has reached approximately 20 million users; OpenAI's standard inference speed has improved by roughly 60% over the past three months; the Ultra Fast mode can currently achieve up to a 14x generation speedup, which he predicts will approach the industry default within one to two years; and OpenAI has leveraged powerful models to optimize its underlying inference architecture, reducing the operational cost of its Luna model by approximately 80%—a process Sottiaux explicitly characterized as an early manifestation of Recursive Self-Improvement.

Dissolving Boundaries Toward a Unified AI

Sottiaux first clarified a common misconception: the integration of Codex and ChatGPT is not a future plan but an accomplished reality. He attributed the inevitability of this merger to the evolutionary logic of model capabilities themselves—future models will natively possess skills like coding, search, research, tool invocation, and voice and vision comprehension, all built upon a single agent harness. In this context, artificially distinguishing "programmers using Codex" from "regular users using ChatGPT" becomes meaningless.

The end state Sottiaux describes is a Personal AGI: programmers, designers, product managers, salespeople, and even non-technical users all tap into the same underlying AI. The system automatically presents differentiated interfaces based on user identity, connected work tools, and permissions—programmers see deeply customized development environments, sales teams see email and customer data dashboards, and everyday users see a clean chat interface. His remarks clearly point to OpenAI's product strategy: not building multiple vertical AI applications, but crafting one AI that adapts itself based on who the user is.

Agent Architecture Overhaul: Complexity Sinks, Interfaces Simplify

Sottiaux's assessment of next-generation agent systems was among the most information-dense segments of the interview. He noted that many advanced Codex users currently maintain skill configuration files, memory contexts, and sub-agent orchestration networks manually—practices OpenAI views as extremely early transitional forms. Skill files accumulate and become hard to maintain, memory frequently loses context, and scheduling issues with multiple sub-agents often disrupt the user experience.

His central thesis: the more powerful the model, the less users should need to directly manage the agent itself. An ideal agent should understand the user deeply over time—knowing their goals, daily workflows, personal habits, and team progress—and autonomously decide which skills to invoke, what critical information to retain, and whether to launch other agents in the background. The essence of this evolution: backend architecture can be immensely complex, but the user-facing interface must remain consistently simple.

Compute Bottleneck: Laptops Yield to Cloud Clusters

Sottiaux put forward a counterintuitive judgment: the future bottleneck constraining agent capabilities will be the user's personal computer. His reasoning: existing PC hardware is designed around human working speeds—people can only operate a limited number of windows and applications at once, with clear physiological limits to processing speed. Models, however, are not bound by these constraints. He cited examples where a cloud-based model might simultaneously handle 100 applications, exploring multiple solutions, writing test cases, compiling code, validating hypotheses, and dispatching sub-agents in parallel.

At this level of concurrency, even a top-spec MacBook cannot carry the corresponding workload. Sottiaux therefore stated plainly that cloud agents will become the dominant paradigm: user terminals serve merely as lightweight interaction entry points, while the actual execution happens on backend cloud computing clusters. This implies the endgame for agents is not "more powerful local software," but "a cloud-based compute team on standby at all times."

Speed Leap: Ultra Fast Poised to Become Industry Benchmark Within Two Years

On inference speed, Sottiaux offered a notably aggressive forecast. Ultra Fast mode currently delivers up to approximately 14x generation speed improvements. He expects that within one to two years, this speed will gradually approach the industry default experience. He also added an important applicability caveat: for pure code or text generation tasks, the acceleration is significant; but for workflows involving extensive external tool calls and network I/O, constrained by network and architecture latency, the actual perceived speedup is around 3 to 4 times.

More noteworthy is what drives the speed gains. Sottiaux indicated that OpenAI's acceleration is not solely from stacking GPUs; model efficiency is continuously improving. The Sol model is already substantially more efficient than the previous Terra model, and next-generation models will leap further in token utilization efficiency. Even without enabling Ultra Fast, OpenAI's standard baseline inference speed has improved by about 60% in the last three months.

Work Paradigm Shift: Speed Equalization Sparks Collaboration Upgrades

Sottiaux pointed out that when AI response speeds approach or surpass human thinking and expression speeds, the human-machine relationship undergoes a qualitative change. He observed that many heavy agent users currently operate 10 to 15 agent windows simultaneously, feeding tasks and rotating through progress checks—effectively becoming harried "AI project managers." The core problem is frequent attention switching.

Sottiaux said OpenAI places high priority on protecting and managing user attention. He described the ideal scenario combining Ultra Fast with voice interaction: a user casually voices an idea, the AI instantly generates a prototype; a quick glance and a voice command adjusts it, with modifications completed in milliseconds; the entire process maintains a state of flow without any context switching. Once AI speed aligns with human thinking speed, human-machine collaboration will upgrade from one-way task delegation to genuine real-time synergy.

Recursive Self-Improvement: Already Underway in Underlying Code

The most technically profound discussion in the interview centered on AI systems' self-optimization capabilities. Sottiaux revealed that OpenAI is already using advanced models to analyze and optimize the service architecture of existing models. He specifically mentioned that the Sol model contributed to optimizing Luna's inference architecture, reducing Luna's operational costs by about 80%. Models are deeply involved in rewriting CUDA kernels, optimizing inference stacks, and redesigning system architectures.

When pressed by the host on whether this constitutes "Recursive Self-Improvement," Sottiaux answered affirmatively—"Yes, this is the early form." He also corrected a common misconception: the public often imagines recursive self-improvement as a single model generation suddenly writing the next one autonomously, but reality is more pragmatic. As models become stronger, they first assist engineers in optimizing underlying runtime systems; systems thus become faster and cheaper; reduced compute costs enable models to undertake more massive workloads, continuing to optimize increasingly complex architectures. Sottiaux summarized it as "a complete closed-loop system."

This means the long-awaited moment of "AI starting to iterate autonomously" may never arrive in dramatic fashion—it is already happening continuously within the unglamorous work of infrastructure optimization.

Behind 20 Million Users: Coding Becomes a Foundational Universal Capability

Sottiaux disclosed that Codex has reached approximately 20 million users, with deep integration into ChatGPT being a primary driver of recent growth acceleration. This data points to a structural trend: once Codex's coding capabilities are embedded in ChatGPT, product managers, designers, finance and marketing personnel can invoke them seamlessly. Using data analysis as an example, a user asks ChatGPT to process 5,000 data records; the system may automatically generate and run Python scripts in the background, ultimately presenting chart results to the user without requiring any programming knowledge.

Sottiaux's assessment is that code will become the underlying machine language through which AI operates the digital world. What is called a "Coding Agent" ultimately serves far more than just programming scenarios. Programmers are merely the earliest group to experience this foundational capability revolution.

Insights from the Extended Conversation

Sottiaux reflected on his time at Google DeepMind, noting that the team had built an internal tool similar to what became ChatGPT roughly a year before its public release, but DeepMind's structure was not designed to ship consumer products. In contrast, he praised OpenAI's culture of tight research-product collaboration, strong bias for shipping, and bottom-up empowerment, balanced by a commitment to minimalism and product quality.

On self-disruption, he emphasized that AI's future trajectory does not wait for any company's quarterly results, so OpenAI maintains a forward-looking posture, willing to divert resources from even its most profitable current businesses to pursue next-generation breakthroughs. He cited the newly released advanced voice mode as an example—its natural conversational quality and direct tool-calling ability have fundamentally changed how he personally uses the product, with voice dictation now a daily staple.

Regarding agent systems, Sottiaux stressed that the goal is protecting user attention through thoughtful design. Whether a task warrants immediate notification or a 30-minute delay should be intelligently determined. When combined with Ultra Fast speed and voice input, AI can respond faster than human thought, keeping users in a continuous flow state—an experience that makes reverting to manual multi-agent management unthinkable.

On the spectrum of agentic systems, he distinguished between two problem types: crafting the ultimate personal AI that stays in flow with an individual user, and end-to-end full automation for complex business processes, such as autonomous log analysis and performance tuning, or cybersecurity systems that self-remediate vulnerabilities, potentially operating with minimal human oversight except for the highest-risk actions.

When asked about the user experience in this ultimate form, Sottiaux emphasized deep grounding in human communication. Natural language is a purely human concept, and the technology must understand not just words but the emotional nuances and contextual subtleties behind them. He also highlighted the growing importance of visual perception—AI should understand gestures and facial expressions, and interact naturally with the physical environment, such as observing a whiteboard sketch and discussing it verbally. He noted that pure voice interaction with ChatGPT is exploding in usage since the new voice features launched, confirming that people naturally choose the path of least resistance.

Addressing competition with Anthropic, Sottiaux said OpenAI's focus is on building the most capable, efficient models and democratizing access. He framed the Codex-ChatGPT merger as part of this mission, making the technology accessible to product managers, designers, sales, and marketing professionals. He expressed a commitment to community, transparency, and a philosophy of directly benefiting users through cost reductions and credit resets—a practice that began as compensation for service disruptions and evolved into a goodwill gesture that builds trust, exemplified by an actual physical reset button.

On compute strategy, Sottiaux defended OpenAI's massive upfront investment in capacity, which was once questioned but has proven prescient. This capacity fuels frontier research and enables significant efficiency gains. He detailed how the strongest models are now used to redesign inference stacks and system architectures, yielding both cost reductions and speed improvements. Notably, even without Ultra Fast, baseline speed rose 60% in three months. He emphasized that OpenAI passes these savings to users rather than pocketing profits, as demonstrated with the Luna model's dramatic price drop.

When asked about the balance between allocating compute to new model R&D versus optimizing existing models, Sottiaux described a first-principles approach with dedicated research and product quotas. The Luna optimization, he noted, was not even a difficult tradeoff given the substantial efficiency gains achieved under the same compute budget.

On Sam Altman's reported pause of frontier reinforcement learning training, Sottiaux explained it was a necessary step driven by the safety team to thoroughly understand and harden system components before resuming. He confirmed that clear safety principles were established, and the pause allowed the team to ensure full control before moving forward.

Regarding Ultra Fast usage patterns, he noted its critical application in high-stakes scenarios like incident response, where emergency teams receive immediate access. He also acknowledged the lighter side, including desktop pets that employees enjoy—a fun, stress-relieving element visible on screens across the office.

On pricing, Sottiaux predicted that Ultra Fast would become increasingly ubiquitous over time, following historical patterns of technology commoditization. He cited improvements in token efficiency, hardware, and architecture as compounding factors, projecting that within one to two years, today's fast mode will be the default or near-default, with even higher-tier options reserved for the most complex tasks.

For the broader public, Sottiaux offered reassurance: ChatGPT was built to benefit the world, and efficiency investments directly serve accessibility. He pointed to Luna, now free in the standard tier through partnerships like Replit, noting its capability equals what was frontier-level just six months ago. He encouraged hesitant users to start by talking to people already using AI, drawing inspiration from their experiences, and to explore practical applications like health and personal finance guidance, which he personally uses to prepare for medical appointments.