ABOVE THE NOISE
It's officially a geopolitical battlefield.
Anthropic just dropped a bombshell: three Chinese AI companies secretly siphoned Claude's capabilities through tens of thousands of fake accounts.
This isn't rumor. They have evidence.
THE BRIEF
What happened?
Anthropic revealed that three Chinese AI labs: DeepSeek, Moonshot AI, and MiniMax that they created over 24,000 fraudulent accounts to "distill" Claude's capabilities and train their own models. The total: more than 16 million exchanges with Claude. MiniMax led with 13 million interactions, Moonshot clocked 3.4 million, and DeepSeek logged over 150,000 but theirs were more targeted, including prompts asking Claude to generate "censorship-safe" alternatives to politically sensitive questions about dissidents and party leaders.
Distillation, in plain terms: you train a weaker model on outputs from a stronger one. Like copying someone's homework at a scale of millions of interactions, fully automated.
Why it matters?
This goes beyond a Terms of Service violation. Anthropic argues that models built through illicit distillation likely have their safety guardrails stripped out and can then be fed into military, intelligence, and surveillance systems. The most striking detail: Anthropic caught MiniMax actively distilling a model before it launched.
When Anthropic released a new Claude version during the campaign, MiniMax redirected nearly half its traffic to capture the new model within 24 hours. That level of coordination signals intent, not opportunism.
The timing matters too. This disclosure lands in the middle of a fierce U.S. debate over AI chip export controls to China. Anthropic isn't subtle about using this as a policy argument: restrict chips = restrict the compute needed for distillation at this scale.
LEFT CURVE TAKE:
The most interesting thing here isn't that DeepSeek distilled Claude. It's that Anthropic chose to go public with this level of detail. Named labs. Cited metadata. Staff profiles. Called for coordinated policy action. This is a calculated political move, not just a technical disclosure. When an AI company draws this much attention to a competitor by name, with evidence this specific, they're trying to shape a narrative as much as warn an industry.
The AI race has officially entered a phase where the line between commercial competition and national security no longer exists.
LEFT CURVE PICKS
Developer Tibor Blaho found a "Pro Lite" tier in ChatGPT's web app code, priced at $100/month. OpenAI currently has a glaring gap between Plus ($20) and Pro ($200) and users have complained for months. Pro Lite would reportedly offer top-model access and 3–5x the reasoning quota of Plus. No official confirmation yet.
→ The clearest signal yet that AI is shifting from "experiment tool" to daily work infrastructure, people will pay more when the value is undeniable.
Nvidia released benchmarks showing its Blackwell GPUs can run inference on open source models (Llama, Mistral, etc.) at up to 25x lower cost per token than previous generations. This has major implications for enterprises weighing self-hosted models against paying for OpenAI or Anthropic APIs.
→ When running your own model gets this cheap, the "build vs. buy" decision tips toward build faster than most people expect.
TechCrunch reports China's BCI sector is expanding rapidly, with a surge of startups and state investment flooding the space. Context: Neuralink leads in the U.S., but China is pursuing the same convergence of AI and biology through a parallel track and moving quickly.
→ Distillation attacks today, BCI race tomorrow - the U.S vs China tech competition is opening new fronts simultaneously.
TIP OF THE DAY
“Distillation” is learning from the outputs of someone smarter without understanding the underlying mechanism - doesn't just happen to AI. It happens to people constantly. Copying frameworks without knowing why they work. Applying best practices without understanding the original context.
Try this prompt in Claude or ChatGPT whenever you're learning a new skill or adopting a framework:
"I'm currently [describe what you're learning or doing]. Instead of giving me best practices, explain: (1) Why was this framework created, what original problem was it solving? (2) When does it NOT work? (3) If I had to derive this framework from scratch, what reasoning would get me to the same conclusion?"
Use it for: learning new skills, researching competitors, onboarding into an unfamiliar domain.
LEFT CURVE SIGNAL
Anthropic knew about DeepSeek's distillation campaign months before going public. They chose today to say something.
Information is never just information. It's always a decision about when and why to release it.
Found this useful?
👉 Forward it to someone trying to keep up with AI.
Read online: leftcurve.beehiiv.com
