OpenAI acquires a mysterious team! Ultraman plans to rebuild GitHub in Codex.
OpenAI Acquires Mysterious Team! Ultraman Plans to Recreate a GitHub in Codex
Summary
Xinzhiyuan reports Today’s Silicon Valley is bustling! Just now, OpenAI announced the acquisition of a mysterious startup—Pierre Computer. Ultraman quickly approved the deal and immediately began reviewing its code,
Xinzhiyuan reports
Xinzhiyuan reports
</h3
Silicon Valley is so lively today!
Just now, OpenAI officially announced the acquisition of a mysterious startup—Pierre Computer.
Ultraman just closed the deal, and then turned around to fully incorporate its code browsing, inline review, and distributed Git capabilities into Codex’s ecosystem.
At roughly the same time, the parallel programming powerhouse with over 100,000 developersConductor was also taken over by Meta today.
Founder Charlie Holtz announced that the entire team is joiningMeta Superintelligence Lab (MSL).
They will join forces with former GitHub CEO Nat Friedmanand Chief AI Officer Alexandr Wangto create the “next‑generation interactive interface”.
Behind the sudden rush of AI giants buying tools, the calculators are clicking loudly—
Writing code is just the beginning; every step programmers take—from branching and code review to deployment—is now required.
Being singled out by OpenAI at a glance, the Pierre team has a very hardcore background.
A legendary front‑end open‑source frameworkBootstrap co‑founder Jacob Thornton, former Discord core engineer Amadeus, and esm.sh creator Je Xia are all part of this team.
Pierre’s team went out to a street basketball court to shoot a “no filming permit, only one take” dead‑pan short film announcing their joining OpenAI.
Don’t be fooled by the low‑budget look of the announcement video; OpenAI Codex lead Tibo likened Pierre to “RAM” in a computer when he shared it, promising to “reshape everything.”
The reason Tibo can give such a high rating is that although the current Codex can dispatch a dozen agents to write code simultaneously, delivering the final product becomes a big hassle.
Dozens of agents hammer away in the background, tens of thousands of lines of changes instantly fill the screen, and as a result not only can human programmers keep up, but such massive branches have nowhere to be stored.
And the three core skills Pierre has accumulated happen to fill Codex’s gaps in “reading code, reviewing code, and storing code” all at once.
-
View Code (Trees): Already built into Codex Desktop, dedicated to millisecond‑smooth expansion of massive project directories and cross‑branch file locating.
-
Review Code (Diffs): Previously on GitHub reviewing PRs, even if you just spot an AI‑generated variable name typo, you have to switch back to your local IDE, fix it, and push again;Pierre’s Diff now embeds an “Edit” mode directly in the comparison view, so you can fix any line you don’t like with a single click.
Code Storage: a distributed Git repository designed for high‑concurrency read/write of AI Agents, already the invisible infrastructure behind hot applications such as Lovable, Bolt, Poke. The unicorn Lovable alone creates over 8 million new Git repositories each week with it, and Poke uses it to perform 40,000 Git operations per minute to store Agent memory.
Going forward, this technology stack will be fully integrated into Codex; as for Pierre’s original open‑source projects and existing customer services, OpenAI promises to keep everything unchanged and will not cut off supply.
OpenAI is so eager to merge Pierre into Codex, but there’s a big embarrassment behind it—
its own Git system crashed after being written by AI not long ago.
Just two days before the official announcement of buying Pierre, OpenAI had just released its internally developed brand‑new Git backendPando.
It turns out that ever since OpenAI equipped everyone with Codex, engineers have been releasing a swarm of agents daily to branch, push code, and run CI tests in parallel, and the massive internal monorepo has been overwhelmed since the end of last year.
Pulls fail, pushes stall, merges frequently error out, and the reliability of OpenAI’s internal Git system once dropped below 90%.
Under pressure, OpenAI simply rewrote the Git core from scratch.
The result: push speed increased by nearly fivefold, with almost 1 PB of data transferred internally each day, and even CEO Greg Brockman personally posted the performance report.
Originally, people thought this was just OpenAI’s “internal rescue,” but surprisingly, just two days later, it turned around and bought Pierre.
With this series of moves, OpenAI essentially stormed straight into the backyard of GitHub, which is owned by its major shareholder Microsoft.
Of course, OpenAI is not the only one eyeing this juicy piece of meat.
LaunchedOrigin hosting platform’s Cursor, the self‑developedScion version‑control Anthropic, and last month’s newly founded former GitHub CEO Thomas Dohmke, are all encircling the traditional code‑hosting domain.
From the editor and review console to the Git repository foundation, AI giants no longer want to be merely a “plug‑in” that just completes code for people.
Although GitHub, with 150 million developers, still firmly holds the open‑source stronghold, even the platform’s former longtime leader has now turned to join this “stage‑crashing” army.
On the same night OpenAI acquired Pierre, Mark’s MSL also staged a lightning talent grab.
The most noteworthy name in this acquisition is Meta’s acquirer Nat Friedman.
He was the first CEO after Microsoft bought GitHub for $7.5 billion in 2018, and the previous generation of developers’ workflow was shaped by him.
Now, together with Meta’s chief AI officer Alexandr Wang, he has quickly slipped the hottest “next‑gen Agent command center” from Silicon Valley’s geek scene into his own pocket.
Conductor, launched only last July, has captivated 100,000 developers at top companies like Linear, Vercel, Notion, and Spotify by turning “working with agents” into a team‑play experience that feels like a gaming session:
-
Multi‑agent isolated concurrency: You can offload messy, heavy tasks to multiple Claude Code or Codex agents, letting them work simultaneously in their own isolated workspaces without stepping on each other’s toes.
-
Collaborate on the same screen like Figma: Press the shortcut “⌘⇧C” to hand the workspace to a colleague, and several people can watch the agent code together, relay commands, review the diff, and seamlessly click Merge.
Even Ilya Sutskever and YC partner Aaron Epstein were early supporters.
Take a closer look at the screenshot of the Conductor website’s dashboard above, and you can even spot Ilya’s avatar and name.
Nowadays, writing code has become the cheapest step. Type a couple of prompts, and a dozen or so agents can simultaneously modify code, add tests, and fix bugs.
The real bottleneck appears after the code is written: how to review tens of thousands of lines of changes, how to audit dozens of branches, and where to store massive repositories.
Big tech companies know better than anyone that models alone can’t lock in programmers: today you switch from Claude to Claude, tomorrow GPT updates to a new GPT, changing the model is just a click away.
But once the codebase, review process, and history are fully embedded in its tools, moving away becomes difficult.
As OpenAI said in its acquisition announcement:
Every workflow left over from the old software era must be redesigned for the era of humans and agents writing code together.
How the code is written is no longer important; which desk a programmer works at is what big companies will truly compete for next.
Reference:
Edited by: Moses
Get MirrorSpeed
Get MirrorSpeed