Back to The Weekly Recall

SpaceX agrees to pay $60 billion for the company behind Cursor

The Cursor acquisition agreement, DeepMind’s agent monitoring, expert-reviewed AI mathematics and AI-assisted genetic diagnoses.

A CRT agent rides astride a rocket.

SpaceX has agreed to acquire Anysphere, the company behind the AI coding tool Cursor, for $60 billion in shares. Reuters reports that the transaction is expected to close in the third quarter.

SpaceX says it will soon release an AI model through Cursor. It also plans to release Grok Build, a coding agent it has been training with Cursor for several months.

The agreement includes a $4 billion termination payment if competition concerns prevent the acquisition, or $10 billion in other specified circumstances.

Top Stories#

DeepMind is treating powerful internal agents as potential insider threats. Google’s security framework adds monitoring and intervention to alignment, sandboxing and conventional security. DeepMind says it analyzed one million coding-agent tasks and built a monitoring prototype that runs asynchronously and escalates findings to humans.

The framework adapts MITRE ATT&CK threat modelling and measures monitoring coverage, detection recall and response time: what the monitors can see, what they catch and how quickly someone responds. Google acknowledges that visible reasoning becomes insufficient when a model conceals its reasoning or recognizes oversight.

Four AI systems produced passing solutions across seven of ten research mathematics problems. First Proof’s second batch reports a combined result across the systems, rather than a seven-out-of-ten score for any single model. The organizers ran the tests themselves, and around 30 expert mathematicians assessed 39 submissions, with at least two referees per solution.

Passing included essentially flawless solutions and solutions needing minor revisions. One solution to a stochastic partial differential equation used a novel approach.

No system made substantial progress on the metric-geometry problem. Failed submissions included unsupported difficult steps and inaccurate or missing citations.

AI-generated leads helped establish diagnoses in previously unresolved genetic cases. OpenAI describes a retrospective study in which o3 Deep Research generated hypotheses from clinical descriptions, genetic variants and literature. Expert review and clinical-laboratory confirmation established 18 diagnoses among 376 cases, a 4.8% yield after earlier specialist analysis.

Seven of the 18 diagnoses had already been established outside the local records under review.

Reviewers knew the model’s confidence, and the study did not measure time saved, false-positive workload or changes in care. One case had remained unexplained for nearly 20 years.

  • Demonstrate a workflow, save a skill (learn.chatgpt.com): Codex’s June 18 Record & Replay release turns demonstrated macOS workflows into reusable skills; Computer Use must be enabled, and initial availability excludes the EEA, UK and Switzerland.
  • Agents that survive interruptions (blog.cloudflare.com): Cloudflare describes durable execution and file storage for agents alongside Flue 1.0 Beta, an open-source framework built on the Pi harness.
  • 671 billion parameters, 37 billion at a time (globenewswire.com): MLPerf Training added DeepSeek V3, whose mixture-of-experts design activates 37 billion of its 671 billion parameters per token by routing tokens to specialized subnetworks.

The week in AI, once a week.