Welcome back! OpenAI's coding agents pushed so much code that the company rebuilt the machinery behind Git, three weeks after Anthropic ditched it. Sam Altman says the model that cracked a $1 million math problem is coming soon. Claude can now take 4x more data in one go. GPT-6 Astra also made Alibaba's Qwen 6x better at math by editing its brain.
In today's Generative AI Newsletter:
OpenAI: Why did OpenAI keep the Git that Anthropic killed?
ChatGPT: What model is the math genius?
Claude: How much more can you send Claude at once?
Qwen: What did ChatGPT's model change inside Qwen's brain?
OpenAI rebuilt Git to keep up with Codex
Codex agents were pushing so much code that OpenAI's Git servers started failing, with reliability dropping below 90% at times.
So OpenAI built Pando, a new backend for Git, and it's been running since April.
Git is the tool almost every developer uses to save and share code. Three weeks ago, Anthropic said Claude had outgrown it and swapped it for a system of its own.
OpenAI's fix sits behind Git, so its engineers and agents didn't have to change a thing.
Commits to OpenAI's main codebase are up 1,378% since early 2024, and it expects 2,442% by year's end.
Pushing code is now about 5x faster, and Pando serves almost a petabyte of code a day.
One engineer designed it and teams of Codex agents built it in 10 weeks.
Anthropic's own post suggested companies rebuild Git's servers instead of leaving Git, which OpenAI had already done.
OpenAI says it's exploring ways to share the tech with its users.
Don’t miss out
Alex Wang, our head of education strategy, goes live with Ho Bae, co-founder and CEO of CUBIG, on what actually changes with AI in the near future.
CUBIG helps you get data ready for AI and keep it that way as their rules change.
What they'll cover:
→ AI-ready data vs. agent-ready data
→ Why context matters as much as the data itself
→ How to tell when a wrong AI answer is really a data problem
→ Build vs. buy as AI use cases start to scale
Thursday, Nov 12, at 8 AM ET, live on LinkedIn.
Altman reveals the GPT that cracked a $1 million math problem
Sam Altman just said the next GPT is the model OpenAI used to crack one of math's $1 million Millennium Prize problems, and that it's coming soon.
He said it on stage at ADIPEC, a giant energy conference in Abu Dhabi.
The problem is Navier-Stokes, the equations that describe how water and air move. In September, OpenAI said 10,000 copies of an unreleased model worked for about four days and proved the equations can break down.
An Anthropic researcher and an NYU mathematician were working on a similar problem, and Altman called their plagiarism accusations unfounded.
Altman said AI went from trailing the best scientists in January to mostly beating them now.
OpenAI also just launched LifeVal, a test of everyday chores like finding childcare and budgeting, where GPT-6 Astra already scores 80.6%.
He didn't say when soon is.
Claude quadrupled how much you can send it
Apps built on Claude can now send it up to 128 MB in a single request, up from 32 MB.
The old cap kept rejecting image-heavy jobs like big PDFs, even when Claude had plenty of room left to read them. Long computer-use sessions, where Claude runs a computer by looking at screenshots, hit the same wall.
It's live now with no price change.
Developers don't need to update anything.
The new limit only works on Anthropic's own API, while Amazon Bedrock stays at 20 MB and Google Cloud at 30 MB.
On Monday, Anthropic also let long chats keep their history on its servers, so apps only send the newest messages.
ChatGPT made Alibaba's Qwen 6x better at math
MIT researchers handed GPT-6 Astra, the model behind ChatGPT, the raw insides of a small Alibaba AI called Qwen and asked it to make Qwen better at math.
Astra edited 0.016% of the numbers that make up Qwen's brain. Qwen's score on a standard math test jumped from 7.9% to 48.1%.
It found the problem on its own. Qwen kept repeating itself until it ran out of room, so Astra traced the loop to two of its layers and cut it out.
That let Qwen finish answers it already knew how to solve.
Before the fix, 93% of Qwen's math answers hit the length limit. After, it was 33%.
Its instruction following tripled too, from 21% to 62.3%.
The job took about 20 hours, almost 5 of them spent waiting out a service outage.
It couldn't help Google's Gemma 3 270M, which didn't have a looping problem to fix.
Starting from scratch, Astra also wrote a Breakout-playing AI that scored 20x higher than one trained the usual way.
The researchers say the same trick could one day let an AI improve its own weights.
Tool of the Day: GitDiagram
GitDiagram turns any GitHub project into an interactive map of how its code fits together. Swap "hub" for "diagram" in any GitHub link and you get the diagram in seconds, then click any part to jump to the code behind it. It also has a library of narrated one-minute videos that explain how popular projects work. It's free and open source.
Try this yourself:
Go to gitdiagram.com and paste a GitHub link, or try an example like Monkeytype, the typing test site.
Shortcut: on any GitHub page, change "hub" to "diagram" in the address bar.
Click any part of the diagram to jump to its code.
Open Videos for a one-minute narrated tour of a project.
Who it's for: anyone who wants to see how an app works before reading a single line of its code.
Everything else you shouldn't miss
OpenAI teamed up with the ex-DOGE startup that got into Anthropic: Cathedral's researchers will use OpenAI's models to test its defenses, and OpenAI more than quadrupled its bug bounty rewards for critical findings.
Abu Dhabi will turn its government rules into something AI can read: it's one of four deals OpenAI signed there Monday, as it raises $20 billion with UAE investors co-leading.
DeepSeek is building a boss AI to manage your other AIs: it says people can only watch 3 to 5 agents at once, and its early prototype finished one task in under 16 minutes against more than 29 the usual way.
Meta reportedly built a humanoid robot that could give its Muse AI a body: the 105-pound prototype, called Max, was built with China's Fourier in three months, though Meta says it has no plans to sell it.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Check out courses made by AI experts on GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
