OpenAI is driving a deep product integration\u2014not merely merging two applications into one, but making the very concepts of “ChatGPT” and “Codex” disappear entirely, ultimately converging into a Personal AGI that understands users over the long term and autonomously completes tasks.<\/p>\n
Tibo Sottiaux, head of Codex at OpenAI, made clear in a recent interview with tech blogger Matthew Berman that Codex has been deeply integrated into the ChatGPT product ecosystem, with programming capabilities originally aimed at developers now gradually opening up to all users. But he stressed that this is only the first step. The endgame he described is striking: all complex underlying architecture will eventually be completely hidden, leaving only a single, radically minimalist “Personal AGI” that deeply understands the user and can instantly mobilize massive cloud computing power.<\/p>\n
During the interview, Tibo disclosed a series of data points and judgments with market relevance: Codex has reached roughly 20 million users; OpenAI’s standard inference speed has improved about 60% over the past three months; the Ultra Fast mode currently delivers up to a 14x generation speedup, which he predicts will approach industry-default levels within one to two years; additionally, OpenAI has used a powerful model to optimize the underlying inference architecture, cutting operating costs for its Luna model by roughly 80%\u2014a process he explicitly characterized as an early form of Recursive Self-Improvement.<\/p>\n
Boundaries dissolving, moving toward a single AI<\/p>\n
Tibo first clarified a common misconception in the interview\u2014the integration of Codex and ChatGPT is not a future plan but an already-completed reality. He attributed the inevitability of this convergence to the evolutionary logic of model capabilities themselves: future models will natively possess programming, search, research, tool-calling, voice, and visual understanding capabilities, all of which will ultimately be built on the same Agent Harness framework. In this context, artificially distinguishing between “Codex for programmers” and “ChatGPT for regular users” will lose all meaning.<\/p>\n
The endgame he described is a Personal AGI: programmers, designers, product managers, salespeople, and even ordinary users with zero programming knowledge will all tap into the same underlying AI. The system will automatically present differentiated interfaces based on user identity, connected work tools, and permissions\u2014programmers see a deeply customized development environment, salespeople see email and customer data dashboards, and regular users still see a simple chat box.<\/p>\n
His remarks clearly point to OpenAI’s product strategy: not building multiple vertical AI applications, but building one AI that automatically adapts based on who the user is.<\/p>\n
Agent architecture restructuring: complexity sinks downward, interfaces trend simpler<\/p>\n
Tibo’s assessment of next-generation agent systems was among the most information-dense parts of the interview. He noted that many advanced Codex users currently maintain skill configuration files (Skills), memory contexts (Memory), and sub-agent orchestration networks manually\u2014which OpenAI views as an early transitional form. Skill files accumulate and become difficult to maintain over time, memory frequently loses context, and scheduling issues when multiple sub-agents work in coordination can easily disrupt the user experience.<\/p>\n
His core judgment: the stronger the model becomes, the less users should be directly involved in managing the Agent itself. An ideal Agent should understand the user deeply and over the long term\u2014knowing their goals, daily workflows, personal habits, and even team progress\u2014and autonomously decide which skills to invoke, which key information to retain, and whether to launch other Agents in the background. The essence of this evolution: the underlying architecture can be extremely complex, but the interface exposed to users must continuously trend toward simplicity.<\/p>\n
He offered a counterintuitive judgment: the future bottleneck constraining Agent capabilities will be the personal computer sitting at the user’s desk. Existing PC hardware is designed entirely around human work speed\u2014a person can only operate a limited number of windows and applications at once, with clear physiological limits on processing speed. But models are not bound by these constraints. He gave an example: a future cloud model might concurrently process 100 applications at once, exploring multiple solution paths simultaneously, writing test cases, compiling code, validating hypotheses, and orchestrating multiple sub-Agents in parallel.<\/p>\n
At that level of concurrency, even a top-spec MacBook cannot handle the corresponding workload. Tibo therefore stated explicitly that Cloud Agents will become the dominant paradigm: the user terminal is merely a lightweight interaction entry point, while the actual task execution happens on cloud computing clusters behind the scenes. This means the endgame for Agents is not “more powerful local software,” but rather “a cloud-based compute team on standby at all times.”<\/p>\n
Speed leaps and recursive self-improvement<\/p>\n
On inference speed, Tibo offered a notably aggressive prediction. The Ultra Fast mode currently delivers up to roughly a 14x generation speedup, and he expects this speed to gradually approach industry-default experience within about one to two years. He also added an important caveat on applicability: for pure code or text generation tasks, the acceleration is significant; but if the workflow involves heavy external tool calls and network I\/O operations, the actual perceived speedup is roughly 3 to 4x due to network and architectural latency.<\/p>\n
More noteworthy is what’s driving the speed gains. He said OpenAI’s acceleration is not simply a matter of stacking more GPUs\u2014the models themselves are becoming more efficient. The Sol model is already far more efficient than the previous Terra model, and the next generation will make further leaps in token utilization efficiency. Even without enabling Ultra Fast, OpenAI’s standard baseline inference speed has improved about 60% over the past three months.<\/p>\n
The most technically profound discussion in the interview pointed toward AI systems’ self-optimization capabilities. Tibo revealed that OpenAI is already using advanced models to analyze and optimize the serving architecture of existing models. He specifically mentioned that after the Sol model participated in optimizing the Luna model’s inference architecture, Luna’s operating costs dropped by roughly 80%. Models are now deeply involved in rewriting CUDA kernels, optimizing inference stacks, and redesigning system architectures\u2014fundamental infrastructure work.<\/p>\n
When the host pressed him on whether this constitutes “recursive self-improvement,” his answer was affirmative\u2014”Yes, this is the early form.” He also corrected a common public misconception about the concept: the popular imagination of recursive self-improvement involves one generation of models suddenly and autonomously writing the next generation, but the reality is far more pragmatic. Models get stronger, then first assist engineers in optimizing the underlying operating systems; systems speed up and costs drop; with lower compute costs, models can take on vastly more work, and then continue optimizing more complex architectures. He summarized it as “a complete closed-loop mega-system.”<\/p>\n
This means the long-awaited moment when “AI begins iterating autonomously” may never arrive in dramatic fashion\u2014it is already happening continuously in the unglamorous work of underlying architecture optimization.<\/p>\n
Structural trends behind 20 million users<\/p>\n
Tibo disclosed in the interview that Codex has reached roughly 20 million users, with deep integration into ChatGPT being one of the primary drivers of recent growth acceleration. This figure points to a structural trend: once Codex’s programming capabilities were folded into ChatGPT, product managers, designers, finance, and marketing personnel can all invoke them seamlessly. Take data analysis as an example: a user asks ChatGPT to process 5,000 data records, and the system may automatically generate and run a Python script in the background, ultimately presenting chart results to the user\u2014all without the user needing any programming knowledge whatsoever.<\/p>\n
His assessment: code will become the underlying machine language through which AI operates the digital world. The so-called “Coding Agent” ultimately serves far more than programming scenarios alone. Programmers are simply the earliest group to experience this revolution in underlying capabilities.<\/p>\n
When discussing competition with Anthropic, Tibo said he generally doesn’t pay much attention to what competitors are doing, focusing instead on OpenAI’s unique strengths, values, and how to move toward the goal at maximum speed. He emphasized OpenAI’s differentiating narrative: not simply “stronger models,” but “putting the strongest capabilities into as many hands as possible.” This is also a key reason Codex was merged into ChatGPT\u2014leveraging ChatGPT’s massive existing user base for rapid distribution, enabling product managers, designers, sales, marketing, and PR professionals to use the technology with zero barriers.<\/p>\n
From DeepMind to OpenAI: cultural DNA migration<\/p>\n
Tibo reflected on his experience at Google DeepMind during the interview. He revealed that roughly a year before ChatGPT’s debut, Google had an internal project called “LM Chat,” but DeepMind at the time was “too cautious to release it,” constrained from launching products that might disrupt Google’s existing businesses. In contrast, OpenAI’s research and product teams collaborate extremely closely, with “a very strong bias toward shipping, very eager to get products into public use.”<\/p>\n
He characterized OpenAI’s culture as “bottom-up with full empowerment”: people can freely propose all kinds of ideas, come together to ship products quickly, and face almost no bureaucratic friction when pushing new product ideas forward. At the same time, the company balances this openness with a relentless pursuit of “radical simplicity” and pride in product quality, preventing products from becoming “feature grab-bags without overall direction or coherence.”<\/p>\n
He also mentioned that OpenAI has a physical “quota reset button” internally. The practice originally began as a way to compensate users for service outages or poor experiences, and later evolved into a beloved topic in the community. “There’s no complex bureaucratic review behind it\u2014no need for layer upon layer of approvals from marketing or finance. As long as I think it’s appropriate, I can press the button at any time.”<\/p>\n
Tibo’s full name is Thibault Sottiaux. He is from Belgium and studied applied mathematics as an undergraduate at UCLouvain. He joined Google in 2015, initially working on Google Maps, then moved to DeepMind where he was responsible for AI research infrastructure, supporting cutting-edge AI projects including AlphaGo. He joined OpenAI in 2024 and took charge of the Codex project. Because he frequently responds to user feedback personally and helps developers reset their Codex usage quotas, he has been affectionately nicknamed “Cyber Godfather” in the developer community.<\/p>\n
The signals from this interview carry profound implications for the AI industry. If OpenAI’s product integration path proceeds as planned, the current startup logic centered on “vertical AI applications” will face a fundamental challenge\u2014when underlying model capabilities become sufficiently powerful, users will no longer need to choose different tools for programming, writing, design, and other scenarios, but will simply need one general-purpose agent that can adaptively adjust its interface. Meanwhile, the rise of cloud agents means compute power will become a more critical strategic resource than terminal hardware, with lasting implications for the competitive landscape of cloud service providers and the chip supply chain.<\/p>\n","protected":false},"excerpt":{"rendered":"OpenAI is driving a deep product integration\u2014not merely merging two applications into one, but making the very concepts…\n","protected":false},"author":2,"featured_media":150369,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[53,580,2317,5044,52404,157,65946,15891,73554,73555],"class_list":["post-150368","post","type-post","status-publish","format-standard","has-post-thumbnail","category-openai","tag-anthropic","tag-chatgpt","tag-codex","tag-deepmind","tag-luna","tag-openai","tag-personal-agi","tag-sol","tag-tibo-sottiaux","tag-ultra-fast"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/150368","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=150368"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/150368\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/150369"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=150368"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=150368"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=150368"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}