OpenAI Executive: Codex and ChatGPT Will Eventually Merge Into a Single Personal AGI
Nashnova编辑部
OpenAI's Codex lead Tibo Sottiaux says all product boundaries will eventually disappear, converging into a "Personal AGI" that deeply understands each user; the bet is not on multiple vertical AI apps but on one adaptive system tied to your identity.
Codex merging with ChatGPT — what's the endgame?
Tibo's vision is explicit: Codex, ChatGPT, and every future product line ultimately become one thing — what he calls "Personal AGI."
This means → OpenAI's strategy is not to build a suite of vertical AI tools. It is to build a single entry point that recognizes who you are and summons whatever capability you need.
In plain terms = you will never have to decide "should I ask ChatGPT or use Codex for this?" There is just one AI, and it does everything.
20 million users — what's driving Codex's growth?
Codex has reached roughly 20 million users. Tibo attributes the recent acceleration to its deep integration with ChatGPT.
The key shift: coding ability has become invisible infrastructure. A user asks to process 5,000 data records; the system writes and runs a Python script in the background — no coding knowledge required.
This means → programmers were simply the earliest adopters. Product managers, designers, and finance staff are the real next wave of growth.
What should the ideal Agent look like?
Tibo's take is counterintuitive: the stronger the model, the less users should manage the Agent themselves. Power users who manually maintain skill configs, memory contexts, and sub-agent orchestration are, in OpenAI's view, in an extremely early transitional phase.
The ideal Agent understands you deeply over time — it knows your goals, workflows, and habits, and decides on its own which skills to invoke, what information to retain, and whether to spin up other Agents in the background.
In plain terms = all the time you spend "training" your AI today is work OpenAI believes the AI should eventually handle itself. The underlying architecture can be enormously complex, but the interface you see should only get simpler.
Why will cloud-based Agents become the norm?
Tibo identifies a bottleneck most people overlook: what will constrain future Agents is not the model itself but the user's local hardware.
His example: a cloud model might concurrently handle 100 applications, dispatching multiple sub-Agents to explore parallel solution paths — a workload no top-spec laptop can support.
This reflects OpenAI's product-form bet: cloud-heavy, local-light. The real compute battlefield for Agents is in the cloud.
What structural shifts are happening in inference speed and cost?
Ultra Fast mode currently delivers up to a ~14× speed increase in generation. Tibo expects this to approach the industry-default experience within one to two years.
Even without Ultra Fast, OpenAI's standard inference speed has improved by roughly 60% over the past three months.
The cost side is even more striking: after OpenAI used the Sol model to optimize Luna's inference architecture, Luna's operating cost dropped by approximately 80%.
"Recursive self-improvement" — what does this loop mean?
Asked directly whether Sol optimizing Luna qualifies as "recursive self-improvement," Tibo answered unequivocally: "Yes — this is the early form."
This means → the loop works like this: model gets stronger → helps engineers optimize the underlying system → system gets faster and cheaper → model can take on far more work → continues optimizing even more complex architecture.
In plain terms = AI is now helping improve the very system that runs AI. Once improved, the AI becomes stronger and cheaper, then keeps improving — if this cycle continues to accelerate, it is exactly the path OpenAI describes toward AGI. Whether this loop truly sustains its acceleration is the core proof point for tracking OpenAI's strategic delivery.
Content is for reference only, not financial advice.