Welcome to Jumble, your go-to source for AI news updates. This week, three Chinese labs shipped open models that closed the gap with the best closed systems in the world. Meanwhile, an unreleased OpenAI model solved ten impossible math problems. Let’s dive in ⬇️
In today’s newsletter:
🌏 Chinese open models catch the frontier
🧮 Ten open math problems fall at once
📜 Washington quietly finishes its AI blueprint
👻 Snapchat stops rewarding purely AI clips
📓 Weekly Challenge: Make your appliances talk
🔓 Three Labs, Three Weeks, One Vanished Gap
DeepSeek, Moonshot and Alibaba all shipped inside a single stretch, and the distance between open and closed models effectively disappeared. Open models spent three years trailing the frontier by about six months, and that gap has now functionally closed.
Would you swap a paid model for a much cheaper open one?
🤗 Hugging Face Says China Is Ahead
CEO Clement Delangue told CNBC on Monday that China is clearly dominating on open models and could take the frontier within a year. He says American labs are building in silos.
🐳 Kimi and DeepSeek Take the Edges
Kimi K3 is the biggest open model ever built at 2.8 trillion parameters. DeepSeek went the other way, running a full benchmark suite for three cents against $3.15 on Claude Fable 5.
🛒 Alibaba Closes the Last Gap
Qwen3.8-Max landed earlier this week at 2.4 trillion parameters, scoring 37 points behind Claude Opus 5 on Frontend Code Arena, making it the fourth best model overall. Alibaba says it opens the weights next week.
👇 This is probably one of the best non-AI-doom model release videos we’ve seen yet.
🧠 An Unreleased Model Just Cracked Ten Open Problems
On August 1, OpenAI published ten results produced by an internal version of Astra, its next major model. Each one resolves or substantially advances a problem that had sat open for years.
📐 What Actually Fell
The list spans sphere packing, group theory, quantum complexity and lattice cryptography, and it resolves three Erdos problems outright. Every argument was formalized in Lean so anyone can machine check it.
💵 Two Thousand Dollars of Tokens
OpenAI says the tokens needed to find all ten solutions would have cost about $2,000 at Sol API rates. Humans prepared the manuscripts, but the mathematics came from the model.
✍️ Who Gets the Credit
Claiming human authorship for an AI-generated proof would misrepresent both sides, the company argues, so it is naming the system as the source. It is also handing 100,000 researchers free access.
Weekly Scoop 🍦
🔐 Weekly Challenge: Make Your Own Hype Video
Challenge: Analysis that cost hundreds of dollars a year ago now costs cents, so stop saving it for work. Google renamed NotebookLM to Gemini Notebook last week, and it is very good at the paperwork nobody in your house has ever read.
Here's what to do:
📸 Step 1: Photograph your stickers Snap the model number plate on the fridge, water heater, furnace, washer and car.
🧾 Step 2: Build one notebook Search each model number plus the word manual, then drop every PDF in beside your lease and insurance policy.
🔧 Step 3: Ask what you normally search badly Try: "The dishwasher is flashing three times. What broke, and what part number do I need?"
🗓️ Step 4: Get your maintenance year Ask for a month by month schedule of filter changes and warranty expiry dates. Live code running is on Ultra and Workspace today, Pro soon.
🔔 Step 5: Find the coverage gap Ask which appliances your insurance would not cover if they died tomorrow.
Is nearly free open intelligence a gift to everyone, or a national security bill nobody has priced yet? And what would you build first inside Gemini Notebook now that it can run its own code? See you next time! 🚀
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
