The Intake
The Intake — Tuesday, July 28, 2026
On the substrate
Anthropic publishes open-weights position naming Chinese government AI development as primary concern
If you've been watching the open-weights debate and trying to parse where Anthropic stands, Dario Amodei published an essay on July 27 that names the position directly.
Amodei says Anthropic has "never advocated for a ban on open-weights models." He describes non-dangerous open-weight models as "a public good." He names two primary concerns. The first is the Chinese Communist Party achieving superior AI capabilities. He cites permanent military dominance and domestic repression as the specific risks. The second is AI-enabled biological attacks. His proposed mitigations include restricting China's access to advanced chips and cracking down on model distillation practices. He also proposes creating global model safety testing organizations that include China as a participant.
If you've been treating Anthropic's safety posture as equivalent to opposition to open-weights releases, the essay marks a different boundary — chip access and model distillation practices, not the model files themselves.
---
Moonshot AI releases Kimi-K3 weights — 2.8 trillion parameters with a revenue-triggered commercial license
If you're evaluating open-weights models for commercial deployment, Kimi-K3 ships with a license term that becomes a planning variable once revenue scales.
Moonshot AI is a Chinese AI company. It released Kimi-K3 weights on HuggingFace on July 27. The model has 2.8 trillion total parameters with 104 billion activated per inference. It uses a MoE architecture with 896 experts. The model also includes 24 Gated MLA layers. Total file size is 1.56TB. The custom "Kimi K3 License" applies to operators running Model-as-a-Service businesses. A separate commercial agreement is required if revenue exceeds $20 million over any consecutive 12-month period. OpenRouter is offering K3 through seven providers. Pricing is $3 per million input tokens and $15 per million output tokens.
If you're below the $20 million revenue threshold, the model is available now via HuggingFace or OpenRouter. If you're building a commercial product where that threshold is plausibly in scope, the license is the thing to read before the model is in production.
---
Ilya Sutskever's Safe Superintelligence announces Nvidia compute deal after two years building in stealth
Safe Superintelligence is a research company founded by Ilya Sutskever. Sutskever co-founded OpenAI and served as its Chief Scientist until 2024. SSI announced a long-term compute partnership with Nvidia on July 27. The partnership centers on the Vera Rubin GPU platform. Bloomberg reported the deal at $5 billion. SSI describes the arrangement as providing an order-of-magnitude increase in its compute access. SSI has raised $7 billion total at a $32 billion post-money valuation. Backers include Nvidia, Andreessen Horowitz, Alphabet, Lightspeed, GV, and Sequoia. SSI has not released any commercial products. Its stated research goal is developing safe, aligned superintelligence without interim product deployments.
No near-term practitioner implication.
---
For operators
Shadow AI agent activity leaves no incident-response trail unless logging is in place before the incident
Help Net Security BleepingComputer
If you've onboarded AI agents into an organizational environment without an explicit logging policy, the forensic gap Brandy Wityak of LevelBlue describes is likely already present.
Wityak notes that AI agents typically record their activity outside an organization's central logging systems. The records needed to reconstruct what an agent did are often rotated or deleted before an incident is discovered. Wityak draws a distinction between shadow apps and shadow AI agents: shadow apps are passive. Shadow AI agents hold persistent permissions and connect to corporate systems. They take autonomous actions on behalf of users. When those actions are later found harmful, operators without pre-incident logging cannot scope the incident. They also lack the records needed to satisfy regulatory reporting obligations.
If your agent deployments write activity logs outside your central incident response stack, those records may be gone before an incident creates the need to read them.
---