Nvidia halves its OpenAI guarantee, Claude watermarks land, SpaceX closes Cursor

📰 News

Nvidia cuts its OpenAI guarantee from $250B to $120B — investors forced the haircut. Same week, Anthropic’s Q2 jumped $4.7B → $11.5B, with a $190–200B 2028 target and a ~$1T IPO in Sept/Oct. Bubble talk and real demand, side by side. → The Decoder
SpaceX officially closes the Cursor deal — $60B option in April, closed today. Cursor says it now sits on “the largest fleet of GPUs in the world.” xAI is already under the SpaceX roof. → TechCrunch
Claude watermarks via SynthID-Text — EU AI Act. Readers can’t see it; anyone with the key can. Light edits usually leave a trace; code gets less of one. → TechCrunch
Grok used to turn a childhood photo explicit — a woman says her stepfather generated 7,000+ images from a picture taken when she was 11. She joins a Tennessee teens’ suit against xAI. → TechCrunch
World Labs turns one robot task into thousands of sims — Fei-Fei Li’s lab. Train in a reconstructed world, then run hours on real hardware. The bottleneck is experience, not the model. → The Decoder
Invisible AI instructions in a court filing — a Connecticut plaintiff hid white-on-white text to steer automated review. The judge likened it to slipping a note to a juror. → The Decoder
PerceptionBench: no frontier model clears 60% — GPT-5.6 Sol 59.7%, Kimi K3 58.5%, Fable 5 57.2%. A lot of “reasoning errors” start earlier: the model just didn’t see the image. → The Decoder

citrolabs/ego-lite — a browser for agents. Share logins with Claude Code or Codex without giving up your tabs.
holaboss-ai/holaOS — local-first agent workspace. Claude Code and Codex on one shared memory.
github/spec-kit — make the spec executable. Spec-driven development for any coding agent.
cathrynlavery/diagram-design — a Claude Code skill for editorial diagrams. HTML + SVG, not Mermaid slop.

AI News: ChatGPT Ultrafast, Grok 4.6, 3 New Open-Source Models — Matthew Berman | Ultrafast, watermarks, Cursor, and GLM-5.3 in one sweep.

Latent reasoning models are more readable than we thought — Coconut/CODI barely use hidden steps on logic; on math, the intermediate calc shows up 93% of the time when they’re right.
Don’t classify. Hallucinate! — Simon. Let the model invent tags, then embed-match them to your real vocabulary.