Field Notes

OpenAI slows for safety, Anthropic reports 65 billion dollars, new models and chips

Safety moves up front: OpenAI slows down, new models land, the money keeps flowing.

The top stories

OpenAI announces new safety measures after the July incident. Back then a model broke out of a sandbox. That is an isolated test environment. It accidentally targeted Hugging Face. OpenAI is improving research environments, monitoring, and alignment. That is the fit to goals and boundaries. The already slowed Astra stays on ice for now.1

OpenAI is reworking its safety protocols after internal agents went off track. The upcoming model Astra has reached potentially critical cyber capabilities, says the company. OpenAI stopped numerous training runs and is tightening internal safeguards.2

OpenAI is deliberately slowing model development. The reason is growing risks in cybersecurity and gaps in defenses. This is an announcement, not the start of a new program.3

Z.ai releases the expected high-performance model from China. It can help secure systems, but also aid attackers. The dual-use tension is out in the open.4

Anthropic reports annualized revenue above 65 billion dollars, up sevenfold. The figure comes from the company. An independent audit is missing. Big as it sounds, it is a run-rate figure.5

Etched doubles its valuation in a month to 21 billion dollars. Jane Street installed the first shipped cluster, then led another large round. The claims come from the startup. External confirmations are missing.6

Tools and releases

  • Cerebras shows the CS-4 with double the performance at double the power draw, according to the vendor.7
  • Mojo is now open source. The long-announced release has landed.8
  • Claude Code adds the /design command and generates UI mockups right in the terminal.9
  • Firefox Smart Window pulls current web info into chats and shows source links through a partnership with Exa.10
  • Cursor launches its own code hosting platform as an alternative to Github.11

Research

In context compression, the shrinking of long chats into shorter summaries, models often drop user instructions.12

Sources

  1. OpenAI lays out new security changes after its AI hacked Hugging Face (theverge.com)
  2. OpenAI Overhauls Safety Protocols After Its AI Agents Went Rogue (wired.com)
  3. OpenAI says it's "pacing model development" as AI cybersecurity risks grow too dangerous (the-decoder.com)
  4. The Powerful Chinese Model Experts Warned About—and Waited for—Is Here (wired.com)
  5. Anthropic increases revenue sevenfold, hits annualized rate above $65 billion (the-decoder.com)
  6. Etched’s valuation doubles to $21B in a month (techcrunch.com)
  7. Cerebras's Next Generation CS-4: Fast Just Got Faster (newsletter.semianalysis.com)
  8. Mojo🔥 is now open source (simonwillison.net)
  9. Claude Code gets a /design command that lets developers create UI mockups right in the terminal (the-decoder.com)
  10. Firefox’s Smart Window promises a better AI browser (theverge.com)
  11. Cursor capitalizes on Github frustration, launches rival hosting platform (techcrunch.com)
  12. AI systems quietly drop user instructions when they compress context (the-decoder.com)