ChatGPT Helps Crack 22 Year Math Problem

Welcome

In today's Edition:

• Stop betting on a single AI answer

• Chinese Neurosurgeon Uses ChatGPT to Solve a 22 Year Math Problem

• Claude Just Took a Major Step Toward AI-Powered Drug Discovery

• Alibaba’s Qwen Pulls Ahead of Meta and Google in the OpenAI Race

• AI Highlights of the Week

• Too Important to Miss

Stop betting on a single AI answer

Even the best AI is only right 65% of the time, per the Artificial Analysis Accuracy Benchmark. And it sounds confident even when it's wrong. Cuey is a free Chrome extension that cross-checks ChatGPT, Claude, and Gemini in one tab:

• One prompt, three answers side by side. No copy-paste.

• Spot hallucinations before they burn you.

• See exactly where the models disagree.

• Carry your context across all 3 models so you never start a chat from scratch again.

Chinese Neurosurgeon Uses ChatGPT to Solve a 22 Year Math Problem

A 16-hour GPT 5.6 Sol run produced a proof that leading mathematicians say checks out

Beijing neurosurgeon Shanmu Jin was studying brain ultrasound when he wandered into a problem in numerical linear algebra. With help from GPT 5.6 Sol, he produced a proof of Crouzeix’s conjecture, a question open since 2004. The manuscript is public, but it has not yet completed formal peer review.

Key Points:

• An Unlikely Route In: Jin is a doctor with limited formal math training who taught himself the advanced material.

• A 22 Year Puzzle: The conjecture asks whether a sharp bound of two always controls how functions behave when applied to matrices.

• AI Did the Heavy Lifting: Jin let GPT 5.6 Sol work autonomously for 16 hours, then shaped the result into a mathematical paper.

• Experts Checked the Proof: Alex Townsend, Anne Greenbaum, and Michel Crouzeix reviewed the manuscript and said the argument was correct.

• The Caveat: The paper remains a preprint, so the wider mathematical community has not finished formally testing every step.

Why It Matters:
This result shows how frontier AI can open serious research to people without traditional specialist backgrounds. If the proof survives peer review, it could mark a shift where mathematicians spend less time carrying every step themselves and more time guiding and checking machine-generated ideas.

Claude Just Took a Major Step Toward AI-Powered Drug Discovery

Alibaba’s Qwen Pulls Ahead of Meta and Google in the Open AI Race

Three billion claimed downloads show how quickly developers are building on the Chinese model family

Alibaba says its Qwen family has crossed three billion downloads in six months, putting it ahead of open models from Google and Meta. The bigger story is what happened next: developers have turned Qwen into a broad ecosystem of customized models rather than treating it as a one time experiment.

Key Points:

• Qwen Hit Scale: Alibaba says the model family reached three billion downloads globally over the past six months.

• The Lead Is Real: Hugging Face data independently places Qwen well ahead of Google and Meta in 2026 downloads.

• Developers Are Building on It: More than 200,000 Qwen tagged models already exist, showing teams are adapting it for their own needs.

• Open Weights Help: Companies can download, fine tune, and run Qwen privately instead of relying entirely on a paid cloud service.

• The Number Needs Context: Three billion downloads do not equal three billion users, but they still point to unusually broad developer demand.

• China Is Gaining Ground: Chinese models captured 41% of Hugging Face downloads over the past year, reshaping the open AI market.

Why It Matters:
The open AI race is increasingly about which model becomes the starting point for everyone else, not simply which one wins a benchmark. Qwen’s growing lead gives Alibaba more influence over the tools, standards, and products that developers around the world will build next.

AI Highlights of the Week

• Anthropic Revenue Reportedly Surges 14× in Q2

  Anthropic reportedly generated more than $11.5 billion in Q2 revenue, up from just $787 million a year earlier.

  The company also reportedly reached positive adjusted operating income, strengthening its position ahead of a potential IPO.

• ChatGPT Can Now Remember What You Do on Your Mac

  OpenAI launched Computer History, an opt-in feature that turns activity across selected apps and websites into memories ChatGPT and Codex can use.

  It can help you resume recent work, find past tasks, and build reusable workflows, while excluding screenshots, audio, and private browsing.

• Leaked Apple Video Reveals Camera Equipped AirPods

  A leaked macOS demo appears to show camera equipped AirPods using Visual Intelligence to recognize and remember real world objects.

  The unreleased device is reportedly codenamed B790 and could let Siri understand what users see without pulling out an iPhone.

• Tesla’s Cybercab Could Hit Austin Roads This Month

  Tesla is reportedly preparing to launch its steering wheel-free Cybercab on public roads in Austin, starting with rides for Tesla employees.

  The company could then add Cybercabs to its Austin robotaxi service within days, though the reported August timeline could still change.

• OpenAI Tightens Security After Hugging Face Breach

  OpenAI introduced stronger monitoring, network isolation, and alignment safeguards after an AI model compromised Hugging Face during internal cyber testing.

  The company also paused reinforcement learning for two weeks, while its largest planned frontier training run remains on hold for additional safety testing.

Too Important to Miss

Google May Move Pixel Production Out of China

Google Says AI Could Cut Airplane Contrails

Groq Raised $350 Million After Nvidia Deal

Qwen3.8-27B vs Qwen3.6-27B vs Gemma 4 31B: Which Is Best?

Feedback

Know someone who may be interested?