Google had something like ChatGPT. Why didn't it ship?

Tibo recalled that about a year before ChatGPT appeared, DeepMind already had an internal chat product called LMChat. Early responses were more funny than useful, but gradually became practical, and the team considered turning it into a public tool.

LMChat never made it to the general public, and the problem was mostly organizational:

• Google worried a new product would cannibalize its search business.
• DeepMind excelled at research but not at actually shipping products.
• Good ideas stayed inside the company for too long.

Tibo believes OpenAI's biggest difference is the tight connection between research and product teams: they brainstorm and design together, and there's a strong bias toward shipping. Products reach users first, then iterate quickly based on real feedback. Meanwhile, individual contributors have more autonomy to form teams, experiment, and launch.

That can be compressed into one sentence:

Big companies get disrupted not because they fail to see new technology, but because they refuse to let it cannibalize their existing business.

This comparison doesn't prove OpenAI will always be faster. It only shows that whether technology reaches the world is first an organizational question.

But a bias toward shipping doesn't mean piling on features. Tibo quickly added that rapid experimentation must be balanced by another set of principles: the product keeps a consistent direction, the interface stays simple, and quality, performance, and efficiency aren't sacrificed for speed. The hard part is doing both at once—giving frontline teams autonomy without turning the product into a pile of disconnected features.

Benchmarks only capture part of a model's capabilities. Tibo said that after training, teams still need to use the model themselves to discover capabilities that benchmarks don't reveal, then adjust the product direction accordingly. The new voice model is a case in point: once natural conversation, tool calling, and high-quality spoken input came together, he started dictating tasks directly into his phone, and the product—originally designed around a keyboard—had to be rethought.

The seemingly casual "credit reset button" mentioned in the interview is a concrete example of this culture. When Codex malfunctions, has a configuration error, or delivers a poor experience, Tibo can directly reset a user's credits. He explained this as a simple principle: if the product doesn't deliver what it promised, compensate the user—no need to wait for marketing or finance to design a growth campaign.

This eventually became a physical button. It isn't only for failures: the team also grants credits at milestone moments worth celebrating, or gives people who haven't tried Ultra a chance to experience it. What started as an internal compensation gesture gradually became a product experience and community ritual. As for how much growth it actually drove, the interview offered no independent data.

But the button can be pressed repeatedly only because of a physical prerequisite: enough compute. That's where the conversation turned from organizational culture to infrastructure, which we'll get to below.

What this physical button really shows is how close the person in charge is to user problems. Whether a problem can be fixed on the spot determines if "user first" is just a slogan or an institutional mechanism that actually triggers action.

Current Codex is still a transitional form

OpenAI ships fast, but Tibo doesn't think today's Codex is anywhere near the end state. Right now, using a coding agent still requires users to maintain:

• Skill files and working guidelines
• Long-term memory
• Sub-agents and their contexts
• Tools, permissions, and workflows
• Multiple parallel tasks and their status

These capabilities are powerful, but they're still complex to use. When memory misses something, a sub-agent loses context, or a tool call gets stuck, the feeling that "it really understands me" falls apart. A truly ideal AI shouldn't require users to build and maintain an "agent management system" on an ongoing basis.

A simple concept helps here: the Agent Harness is the working system wrapped around the model. The model handles reasoning, while the harness connects context, memory, tools, permissions, sandboxing, and execution flow. The model is like an engine; the harness determines whether the car can actually stay on the road.

Today's power users are essentially still managing these seams on behalf of the agent.

When the product absorbs these seams internally, people can step back from being system administrators and return to being goal-setters and risk approvers. The future AI should automatically understand:

• Who you are
• What your goals are
• What you do every day
• What your team is doing
• What's worth surfacing now
• What can happen in the background

It wouldn't just wait for commands. It should proactively identify problems, make suggestions, and execute tasks within permission boundaries.

This doesn't mean Skills, memory, or multi-agent setups will disappear. They're more likely to shift from "explicit parts" that users manually maintain every day into internal capabilities the system manages automatically. OpenAI has already put multi-agent, cloud environments, Skills, and background tasks into the Codex product, but that's still a long way from the end state where the system understands you over time and coordinates everything on its own.

Related readingOpenAI Open-Sources Codex's Core Agent Harness Framework: Easily Bring Agents into Your Own Business SystemsOpenAI is releasing the execution backbone that manages context, tool calls, sandboxes, and approvals for Codex. Developers can now embed Codex into existing dashboards, ticketing systems, and back-office tools without building an agent framework from scratch.→

So when Tibo says today's Codex will soon look primitive, the key reason is that people still have to manually organize these capabilities into a working system.

ChatGPT and Codex are converging on a single entry point

Over the past few months, ChatGPT and Codex have kept moving closer together. Some users ask: why merge them at all? Tibo's answer is that future models will require it.

There's also a very practical distribution logic. Tibo said Codex can leverage ChatGPT's existing user base to quickly reach non-technical roles like product, design, sales, marketing, and communications—which he believes drove recent growth. The host also noted that Tibo posted online that day claiming Codex reached 20 million users. The interview didn't show the counting methodology or third-party data, so this number should be treated as a company growth metric, not an independently audited figure.

His vision is to remove the artificial boundary between "chatting" and "writing code"—not just put two buttons on the same page. The same agent can write code, do research, give advice, call work tools, or discuss whiteboard ideas with you by voice. The interface no longer asks whether you're a programmer, designer, or regular user first; it adapts based on your goals, experience, and task.

This is also what he means by "personal AGI" as a product concept. Here, AGI refers to a unified entry point, not a claim that artificial general intelligence has been achieved: you and your family use the same product, but the tools it connects to, the questions you ask, the information density you see, and the help you get are completely different.

On a deeper level, this agent has to adapt not just to occupation and interface density, but to how people naturally express themselves. Tibo envisions it understanding tone and nuance in natural language, and entering environments where voice, vision, gestures, and whiteboards exist. People shouldn't have to learn how to accommodate the machine first; technology should gradually become a natural extension of human action. This is still a vision, not a capability the current product fully has.

The interview also drew a clear distinction between two types of agents:

Both can use similar models and tools, but the product requirements differ. A personal agent needs to understand tone, goals, and daily rhythm; a fully autonomous system cares more about process stability, permission boundaries, and exception escalation. Calling both "agents" doesn't mean they can be managed the same way.

There's also a practical boundary worth keeping in mind: OpenAI's current help page still presents Codex as a standalone desktop entry point, with history kept separate from ChatGPT. Voice can already enter Work and Codex, but "one personal agent that adapts to everyone" is still a direction of evolution, not a fully delivered reality.

Faster models expose human attention as the real bottleneck

A unified agent doesn't just need a better interface. It needs an execution environment different from a personal computer.

Laptops are designed for humans: a person can only watch a handful of windows at once, and typing and thinking speed are limited. Models don't have those constraints. In the future, a model could handle many applications at once, explore solutions in parallel, write tests, compile code, and verify hypotheses simultaneously. A computer built around human throughput quickly becomes a resource ceiling.

So the meaning of a cloud-based agent isn't just "moving the computer to a server." It's giving the model resources that match its parallel capabilities. But as resources grow, human attention becomes an even more obvious bottleneck.

The host said he now launches 10 to 15 agents at once, because each task can take 30 to 45 minutes. The cost is constant context switching, checking status, and trying to remember where each task left off. The model is doing the work, but he's become an agent dispatcher.

As speed increases, the bottleneck shifts from generation itself to the entire execution chain.

So what Ultra Fast really shortens is the time from an idea to a judgment-ready result.

The interview gave a concrete internal example: during a service outage, the incident response team gets Ultra Fast because every second counts. When prototyping a website or game, code generation is fast enough that people stay in the same thought flow and iterate repeatedly, rather than submitting a task and walking away.

But extreme speed doesn't deliver equal value to everyone. Tibo said people who habitually juggle many tasks and switch frequently may not gain as much from Ultra Fast; those who prefer single-threaded progress and don't want to leave their current train of thought are the ones most likely to feel the feedback loop dramatically shorten. Whether speed has value depends on whether it fits a person's attention habits.

Low latency ultimately changes more than just coding speed. Tibo's hope is non-text interaction: generating multiple options on a shared canvas, picking images, immediately creating prototypes, then guiding changes in real time with voice or text. At that point, AI isn't just emitting text into a chat box faster—it's letting people iterate continuously in a visible, editable creative space.

Tibo predicts that in a year or two, today's extreme speeds could become the default experience, though there will always be a pricier, faster tier above. This is a product leader's forecast, not a delivery commitment. The real takeaway: the future speed race should be measured by how much the end-to-end feedback loop shrinks, not just tokens generated per second.

Stronger models are starting to reshape the entire AI factory

Speed gains and price drops don't come just from buying more GPUs. Tibo described a more specific kind of "recursive self-improvement."

This also explains why the reset button can keep existing. Tibo said OpenAI began planning compute capacity very early and was once questioned for investing so heavily. Now, that compute supports both research and a product that can absorb heavy usage and user compensation. On the surface, a leader pressed a button. Behind it was an infrastructure bet made long in advance.

Stronger models are now being used to rebuild the entire infrastructure that runs them.

This is system-level self-improvement. The model isn't copying itself out of thin air; it's helping people rewrite the factory that supports it, faster.

That "factory" includes inference services, CUDA kernels, the tech stack architecture, and product interactions. Models help small teams find performance regressions and refactor critical paths, so the same compute handles more requests. Then lower costs and faster speeds expand usage, which brings in new tasks and feedback that feed the cycle again.

The conversation surfaced two figures: the host noted that Luna's price had dropped by roughly 80%, while Tibo claimed that standard-mode speed had improved by about 60% over the previous three months. The first number is public information cited by the host; the second is self-reported by OpenAI's product lead. Neither comes from an independent benchmark, so they can't be used to directly infer company profitability, energy use, or real-world speedups across all tasks.

Tibo also stated that the team doesn't intend to keep all efficiency gains internal, but rather pass them along to customers and users through better price-performance. That commitment aligns with Luna's price cut, but it's still a vendor's self-description. What matters is whether future efficiency gains actually show up in pricing, limits, and availability over time.

This fast loop isn't without brakes. When asked about OpenAI pausing frontier reinforcement learning training, Tibo said the goal was to give safety teams room to understand and harden the system, then resume once clearer principles were in place. That's an important example: a real "rapid release" culture also knows when to stop. The organization has to be able to accelerate, but also hit the brakes when risk climbs.

What to watch isn't which model is temporarily ahead

When the host pressed Tibo on competition with Anthropic, he kept steering the conversation back to the same answer: he doesn't spend much time watching rivals. Instead, he focuses on what OpenAI can uniquely do well, its values, and how to move forward as quickly as possible.

That's corporate narrative from a product lead, not neutral market analysis.

Another answer deserves scrutiny too. When asked about job automation, environmental impact, and public anxiety about AI, Tibo focused on efficiency gains, falling costs, broader access, and existing use cases. He answered the question "how does AI become cheaper and more useful?" but didn't directly address the costs to employment or the environment. That doesn't mean he lacks answers, but it does suggest the conversation offered a product lead's upbeat framing rather than a full social impact assessment.

A more complete checklist, then, might include five items:

1. Will the organization disrupt itself? Will a profitable product block the next generation from getting resources?
2. Is complexity being hidden from users? Are people actually accomplishing their own goals, or still managing Skills, memory, sub-agents, and permissions?
3. Does speed protect attention? Are faster models producing more notifications, or keeping ideas and feedback in a single flow?
4. Are efficiency gains being passed back to users? After cost reductions, does capability actually become cheaper and more widely available, or does it stay on the company's balance sheet?
5. Is the vendor addressing costs beyond price? When it comes to job automation and environmental impact, is it offering evidence, or just redirecting back to product utility?

On whether capabilities will spread widely, Tibo offered a concrete timeframe: six months ago, Luna might still have been at the frontier; today, it's priced very low, and the host noted it has even entered the free tier. That suggests frontier capability can indeed become commoditized quickly. Specific pricing, free access limits, and usage terms should still be checked against official product documentation at the time.

Tibo used writing, personal advice, and pre-doctor-visit preparation as examples of how AI's value doesn't have to wait for a full personal agent. These examples show that people are already using AI, but they don't prove every answer is reliable, and they certainly don't replace professional judgment.

What's worth tracking is whether the human role starts to shift. In the past, people operated software. Today, people give agents tasks and then manage the work system on their behalf. OpenAI's vision is that people eventually only provide goals, boundaries, and key decisions, while the agents handle the coordination themselves.

If that path holds, the clearest sign of a next-generation personal AI might simply be this: you no longer have to manage your agents yourself.