In partnership with
November 3, 2026 · READ ONLINE
Welcome back, AI enthusiasts.
An American AI just helped make a Chinese AI better, and the way it happened is pretty fascinating. MIT researchers used OpenAI's GPT-6 Astra to make changes inside Alibaba's Qwen, with surprising results. The experiment raises an interesting possibility for how AI models could be improved in the future. Let’s Dive In!
IN TODAY'S ISSUE
01 OpenAI's GPT-6 Just Rewired a Chinese AI's Brain
02 Anthropic Refiles IPO as AI Self-Improvement Risks Loom
03 Nvidia's New China Restrictions Reach Beyond Chip Exports
01 / MODEL IMPROVEMENT
Evolving AI: MIT researchers used OpenAI's GPT-6 Astra to fix a flaw in Alibaba's Qwen3.5 without retraining it.
KEY POINTS
Qwen's MATH-500 math score rose from 7.9% to 48.1%, roughly a 6× improvement.
Astra changed just 131,072 of Qwen's 800 million weights, only 0.016%.
Instruction-following scores nearly tripled, rising from 21.0% to 62.3%, extending gains beyond math.
DETAILS
MIT researchers gave GPT 6 Astra access to Qwen's weights to fix repetitive reasoning. Astra patched a pattern across two neural network layers, reducing math answers that hit the output limit from 93% to 33%. The team calls this "weight-space programming." Separately, Astra created a Breakout-playing AI without training in a simplified game environment, scoring 20× higher than a model trained for 500 million steps.
WHY IT MATTERS
AI assistants could improve faster if developers can patch problems instead of retraining models. If Astra's approach proves reliable, stronger models could repair recurring failures in smaller ones, giving users assistants that finish tasks more reliably and follow instructions better. Frontier AI providers could sell their models as repair tools to competitors, while specialist firms test and deploy patches. But Astra mainly corrected an existing reasoning problem. Researchers still need to demonstrate reliable edits across different models without introducing new failures.
One workspace where buyers, sellers, and AI close deals together
Instead of running your deals across scattered email threads and forgotten attachments, Aligned brings everything into one shared deal workspace.
Buyers get a single place to evaluate, loop in their team, and say yes. Sellers get visibility into what is actually happening when they're not in the room and AI Deal Insights that help you act accordingly, before it's too late.
Best part? Your first deal room takes only minutes to stand up.
02 / RSI WARNING
Source: WIRED
Evolving AI: Anthropic has revised its IPO filing with a new warning about self-improving AI, alongside major financial disclosures as it prepares to go public.
KEY POINTS
Anthropic warns future models may soon reach its threshold for dramatically accelerated AI research.
Q3 revenue jumped 47% to $16.96 billion, but Anthropic swung from a $974 million operating profit in Q2 to a $567 million loss.
Anthropic discloses $895.4 billion in future computing and infrastructure commitments, including $447.5 billion expected by 2030.
DETAILS
Anthropic’s initial S-1 mentioned recursive self-improvement (RSI), where AI helps build better models. The revision acknowledges that Anthropic noticed past research accelerations only afterward. No current model has been designated as crossing its Responsible Scaling Policy threshold. Anthropic may require human approval for some AI research decisions. It reports $5 billion in costly short-term computing.
WHY IT MATTERS
Anthropic's disclosures show why investors should look beyond revenue growth. Computing commitments stretch years ahead while future breakthroughs remain unpredictable. Breakthroughs could boost business, but safety restrictions could delay releases and raise costs. Founders retain 50.1% voting power, leaving public shareholders little influence. The IPO price will show what investors pay for the uncertain path to sustainable profits.
TOGETHER WITH DATADOG
Evolving AI: Moving applications to the cloud makes it easier to scale. But as your infrastructure grows, keeping track of what's happening behind the scenes becomes a whole new challenge.
Slow applications, unexpected downtime, and rising infrastructure costs can quickly become expensive problems. And when you can't see what's causing them, fixing them gets even harder. So how do you stay in control as your AWS environment becomes more complex?
Datadog's free eBook shows how to gain better visibility into your cloud applications, identify performance issues earlier, and keep your infrastructure running efficiently.
Inside, you'll discover how to:
Spot problems before they escalate: Identify performance bottlenecks and potential disruptions early.
See the bigger picture: Understand how your applications, services, and cloud resources work together.
Improve application performance: Find the causes of slow responses and reduce unnecessary delays.
Keep cloud costs under control: Identify inefficient resource usage as your applications scale.
03 / CLOUD RESTRICTION
Source: Forbes
Evolving AI: Nvidia is pushing cloud providers worldwide to reduce their dependence on Chinese customers, potentially limiting China's access to advanced AI computing.
KEY POINTS
Nvidia reportedly wants cloud providers to sell at least 51% of their computing services to non-Chinese companies.
The guidance applies to cloud partners worldwide for future contracts, leaving existing agreements unaffected.
Cloud providers are also being asked to identify who ultimately uses Nvidia chips, although the sales target is not mandatory or strictly enforced.
DETAILS
Nvidia's most powerful AI chips face US export restrictions in China, but Chinese developers can still rent access through overseas data centers. Nvidia began sharing the guidance verbally through sales teams this summer. Cloud providers with large Chinese customer bases, especially in Southeast Asia, could struggle to attract customers elsewhere.
WHY IT MATTERS
Nvidia's dominance in AI chips gives it influence over which customers cloud providers serve. Even voluntary guidance could force difficult decisions for firms dependent on Chinese demand. Chinese AI developers face uncertainty over resources needed to train and run advanced models, potentially increasing reliance on domestic alternatives. If Nvidia ties future chip sales to these expectations, cloud providers may have to turn away profitable Chinese customers to secure essential hardware.
04 / HUMAN OR MACHINE?
One is real. Trust your eyes?
Two images. One photograph. Click the one you think is real.
A
B
05 / QUICK HITS
🔐 Anthropic open-sourced FlashProve, an experimental way to prove which AI model generated an answer without exposing its internal weights.
📈 Palantir reported 96% quarterly revenue growth, reaching $2.31 billion as its U.S. commercial revenue jumped 148%.
🛢️ OpenAI signed a deal with ADNOC and G42 to use AI agents for oilfield planning, aiming to shrink complex planning work from months to hours.
🎥 Google Vids added image-based custom AI avatars, letting users upload a photo and refine its background, lighting, and appearance before using it in videos.
06 / TRENDING AI TOOLS
📝 Granola - AI notetaker that captures the real insights and turns every conversation into ready-to-share, action-driving notes*.
🔬 ClawsGO Science - Runs literature research, data analysis and scientific reports for you.
📬 Agent Inbox - Gives AI agents email, Slack and Telegram accounts for approvals.
🧠 Xmind 2027 - Turns mind maps into AI-assisted research and planning spaces.
*PARTNER LINKS
BEFORE YOU GO
Worth your four minutes?
Help us make the next edition better.
See you tomorrow,
The Evolving AI team
