OpenAI Product Head Tibo: Google Had ChatGPT But Did Not Release It
OpenAI's core product head Thibault Sottiaux stated in the Pragmatic Engineer podcast that he participated in a conversational product called LMChat at Google, which was developed about a year before ChatGPT. However, the company was too cautious, and DeepMind was blocked from launching a product that could impact its core search business, so it was never released.
In 2023, he learned that ChatGPT was maintained by only about 20 people but had become widely popular. He believed that such a small team indicated a high degree of autonomy, and he subsequently joined OpenAI, arriving just in time for the reasoning model and o1-preview preparations. Previously, he worked at Google for about six years, holding positions in Maps, Ads, and DeepMind infrastructure, and contributed to support work for projects including AlphaGo.
Codex's command line was written in Rust. At that time, the company's models were better at Python and TypeScript, and the team focused on performance, security, and future scalability with millions of cloud instances, separating the Agent core from the product interface to avoid future rewrites. After open-sourcing, its features were quickly imitated by competitors, which he admitted was "a bit painful," but accepted it as a cost of diffusion. Internally, newcomers often ask, "Have you consulted Codex?" He stated that code review has reached an extraordinary level, with security and logical flaws being caught in merge requests, allowing updates to reach hundreds of millions of users the same day.
He mentioned that the latest models have reduced dependency on scaffolding and can be requested to work continuously for about a week. The challenge is integrating the fully local Codex with the fully managed ChatGPT into a $20 Plus package that serves hundreds of millions. Codex users were previously reported to be around 20 million, while ChatGPT's active users are measured in billions. When hiring, they look for three qualities: the ability to quickly understand systems, a constant questioning attitude, and taste. His exact words were: if you can't clearly explain your intentions, you can't create truly good things.
In terms of market mechanisms, this is organizational storytelling, not a price list. Buyers are looking for engineers who can manage Agents; sellers are using "small teams with high autonomy, and models that can work weekly" to attract talent. Funding is still focused on the cloud costs of computing power and merging the two stacks, not on opening a new independent subscription. Beneficiaries are those who have handed over the review and release loop to the model's internal rhythm; those under pressure are competitors who still rely on thick scaffolding to run their proxies, as well as the large company product processes that previously blocked LMChat. The event was driven by interview disclosures and does not change the API pricing on that day.
On a supplementary level, Rust and open-source are established engineering choices, and the week-long unattended operation is the description of the latest model by the parties involved, not an external SLA. Merging into the $20 tier is a product goal, not all capabilities being equally represented in the same package.
Source: Public Information
ABAB AI Insight
Google developed a conversational model but did not launch it, while OpenAI turned it into a default entry point with about 20 people. Sottiaux's reason for switching jobs was not salary but release rights: similar products were seen as risky to search at DeepMind, while at OpenAI they were viewed as products that could be updated weekly. He joined just in time for o1-preview and then led a small team to make the programming agent one of the fastest-growing revenue lines, later managing the merger of ChatGPT and Codex. The path was "launch first, then merge, and then integrate the local Agent into a low-cost package."
The capital path involves using open-source to build an ecosystem and Rust for scalability. The core was chosen in a language that the models were not proficient in at the time, betting that Agents would need to run on millions of cloud instances, with security boundaries needing to be established before syntactical comfort. Open-sourcing was acknowledged as being copied but resulted in a default position in terminals. Handing over review to models and pushing merge requests to hundreds of millions of users on the same day shifts human effort from writing code to writing intentions. Money is invested in moving local runtimes to the cloud, allowing the $20 tier to accommodate cycles that previously only developer computers could run.
In contrast to Google's internal struggles to launch products due to policy barriers, and Anthropic writing programming agents in TypeScript for model distribution: OpenAI chose release speed and system language, sacrificing short-term accuracy in Rust for the model. The industry phase has shifted from creating chat boxes to developing agents that can operate unattended for a week, with control resting on who dares to dismantle scaffolding and still launch. Hiring standards have shifted from language lists to "being able to articulate intentions," which aligns with stronger models and shorter prompts.
Structural changes belong to technological substitution. The mechanism is: when review and security interceptions can be completed faster than human hands, maintenance costs shift from engineer hours to model calls; when two stacks are forcibly merged into the same low-cost package, programming agents cease to be independent businesses and become backend capabilities for chat entry points. Companies that are afraid to release keep technology in the lab, while those willing to release define default interactions with small teams.
ABAB News · Cognitive Laws
- Being first to produce does not guarantee winning; being first to launch defines the default.
- The stronger the model, the clearer the intentions of the people the company needs.
- Open-source being copied is tuition, resulting in becoming someone else's default runtime.