This website uses cookies

Read our Privacy policy and Terms of use for more information.

Welcome back! Anthropic's CEO described exactly what a dictator could do with AI, hackers took a Coca-Cola dairy offline this month, China published the biggest open model ever days after Washington accused it of theft and Dario finally said where Anthropic stands on open weights.

In today's Generative AI Newsletter:

  • Anthropic: What could a dictator do with AI?

  • Cyberattacks: Who got hit this month?

  • Moonshot: What can you actually do with Kimi K3?

  • Open weights: Is Anthropic for them or against them?

Dario Amodei spent this week explaining what an authoritarian government could do with powerful AI, and he was thorough about it.

Permanent military superiority over everyone else. Repression of your own people with no ceiling on it.

That ceiling is the part he keeps returning to in his own essays. A dictatorship has always needed people to carry out its orders, and people hesitate, leak and defect.

Machines don't. Run the surveillance and the enforcement through AI and the old limit disappears, which is how a regime becomes something you cannot vote out.

Sounds like he accidentally shared a feature list for dictators.

He named the Chinese Communist Party as the most capable government he has in mind, and said it isn't the only one.

He thinks the danger runs past cyberattacks into biological weapons, and he wants safety testing that even China signs up to.

Special highlight from the network

PR-AF ranks #2 of 42 on the Martian Code-Review-Bench, catching about 3x more real issues than the leading paid tools, at roughly 10x lower cost.

Built on AgentField, it runs on open or frontier models and drops straight into GitHub Actions with a single label — no new dashboards, no extra seats.

Fully open source under Apache 2.0

Coca-Cola's Fairlife shut down US production after intruders got into its systems. Abbott is investigating two separate break-ins and Clover Health caught strange logins on its network.

Ransomware is up about 20% this year and it isn't using the same old tricks anymore.

They steal the data first, encrypt what's left, then charge you once to unlock it and again to keep it off the internet.

The White House started a group this month to pass warnings between AI labs and the companies running hospitals, power and water.

Sounds like they raised the alarms. Do you think ransom attacks will hit unprecedented numbers this year?

Washington accused Moonshot of copying Claude, so Moonshot published its entire model for free.

Kimi K3's weights went up Monday, 2.8 trillion parameters, the biggest open model anyone has released, under a license that lets you use it commercially.

It currently ranks first on the Frontend Code Arena, above Fable 5, in blind tests where developers don't know whose model they're judging.

Here's what you can actually do with it.

If you have the hardware you run frontier-level coding on your own machines, no API bill, no rate limit, nothing leaving your building, which is the whole game for banks, hospitals and anyone with sensitive data.

Don't get too excited, hardware isn't cheap. You need roughly 64 H100s across eight servers, and the weights eat 1.4 terabytes even squeezed to four-bit.

For everyone else it shows up through hosting providers this week.

Dario wrote it out Monday after a week of the industry accusing Anthropic of lobbying to kill open models.

He says Anthropic has never called for a ban, and that open models without dangerous capabilities are a public good.

What he wants changed is three things.

  • Keep advanced chips out of China.

  • Crack down on industrial-scale distillation of US models.

  • Test every capable model before release, open or closed, including China's.

What set this off was Jensen Huang's first ever post on X, sharing an open letter from Nvidia, Meta, Microsoft, Mistral and Hugging Face telling policymakers not to rush into restrictions.

He also rejects the idea that open is automatically safer. Once the weights are out the guardrails come off and nobody can pull them back.

Anthropic's name is not on that letter. That's why half the industry read this as open source for us, not for them.

Models hallucinate. Fresh data doesn't. Get your access to fresh data.

That promise now has serious backing: Warburg Pincus invested $130M in Oxylabs, valuing the company at $3.6 billion and making it a unicorn.

A glimpse of where AI is headed, since models are only as good as the live web data feeding them.

Oxylabs delivers that real-time data to your AI agents at scale: 175M+ residential proxies across 195 countries, JavaScript rendering for dynamic sites and OxyCopilot, which builds scrapers from plain-English instructions.

New this round: Headless Browser for AI Agents and Fast Search API + an Agentic Skills repository teaching AI agents to use it all.

Named Best Proxy Service of 2026 by PCMag.

Claim up to 2,000 free scraped results with Oxylabs.

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.

Reply

Avatar

or to participate

Keep Reading