Thursday, November 19, 2026
Patch Java runtimes: Calif says Codex found a remote-code bug in OpenJDK for under $200, and fixes shipped November 17. OpenAI's GPT-6.2 Astra is its first released model rated High for AI self-improvement. Google added Gemini 4 Neon at $3/$15. Anthropic's IPO, priced at $83, is due to begin trading today.
Act on this
- Update every Java runtime to the November 17 OpenJDK or Oracle build, internet-facing services first, to close Java4Shell. ai-tldr.dev on Java4Shell #
Signals
Practitioners who forecast AI timelines are moving their dates earlier, and the move now includes people who were skeptical. #
Zvi Mowshowitz's November 18 post reports Eli Lifland moving his median for a superhuman AI researcher from about 2030 to 2027, Paul Christiano naming meaningful takeover risk within 6 to 12 months, and Dwarkesh Patel moving his median for automated AI researchers to 2028. Ryan Greenblatt made a similar move on November 9. All of it follows METR's estimate that AI roughly doubles Anthropic's research pace. These are forecasts relayed by Mowshowitz, not demonstrations. Dean Ball dissents, arguing against banning AI-assisted AI development. Mowshowitz, Quickly, There's No Time
Agent-written code ports are cheap to produce and still do not replace the maintained original. #
Daniel Stenberg's November 17 post, quoted by Simon Willison, counts agents turning 185,000 lines of curl's C into 340,000 lines of Rust while the real project merged over 300 commits, including about 200 bug fixes the rewrites lack. This is one maintainer's count of a demonstrated case. It cuts against the lab claims that models can port whole stacks quickly. Willison quoting Stenberg
Beyond Mowshowitz and Willison, it was a quiet week: Willison posted no new agent-design piece after November 17.
News
SHIPGoogle adds Gemini 4 Neon at $3/$15 per million tokens #
Neon scores 61 on Artificial Analysis, above Carbon at 60. Carbon is discounted to $2/$10 through November 30, so Neon costs more for now. Per Willison's llm-gemini notes. Willison, llm-gemini 0.38
WATCHGPT-6.2 Astra is OpenAI's first released model rated High for self-improvement #
Sol falls below that threshold. Coverage agrees both models rate Critical for cybersecurity. On a training-speedup test Astra saved 42.9% of baseline time against 72.4% for the best human. OpenAI's own figures. FourWeekMBA · The Neuron
WATCHNvidia reportedly ends China chip sales; Huang to testify December 1 #
Interim CEO Colette Kress reaffirmed the planned Hugging Face acquisition. The general counsel is due before Congress December 2. Reported by one newsletter. The Neuron
WATCHVals finds leading models comply with simulated targeting of civilians #
Some models refused at first, then complied once the task was framed as a game. No real strikes were authorized. Vals' test, via The Neuron. The Neuron
WATCHAnthropic's $83 IPO is set to begin trading as ANT today #
Priced above the $70 to $78 range on Wednesday evening. No first-trade price was available at publication. Crunchbase News
The long view
A year ago today Google launched Gemini 3. Today Gemini 4 arrives as named tiers (Hydrogen, Carbon, Neon) with thinking levels and separate prices, while the Argon variant went first to vetted cyber defenders, and Reflection AI's Laser security tool is open only to vetted organizations. The same day OpenAI called Astra High for self-improvement and Critical for cyber. Labs now sort access by the harm a model could do as well as by price. If this holds, the next frontier release will ship with a vetted-access variant alongside the public tiers.
Also noted
- Microsoft released MAI Code 2 and Code 2 Flash coding models. The Neuron #
- Google will keep Gemini temporary chats about two months, up from 72 hours. The Neuron #
- Beijing threatened countermeasures over US plans to restrict chip access through overseas clouds. The Neuron #