Saturday, 17 October 2026
Weather Models Were Guessing Two to Six Weeks Out. A Correction Doubled Their Skill
Weather Models Were Guessing Two to Six Weeks Out. A Correction Doubled Their Skill
A Microsoft Research team, working with MIT climatologist Judah Cohen, published a method in Nature on Wednesday for the hardest stretch of weather forecasting: two to six weeks ahead, after the ten-day forecast runs out and before the seasonal outlook begins. It is called probabilistic bias correction, and it replaces nothing — it learns the habitual errors in a model’s past forecasts and corrects them. On the European Centre’s systems it doubled the subseasonal skill of ECMWF’s AI model and improved its physics model on 91% of pressure, 92% of temperature and 98% of precipitation targets. In ECMWF’s 2025 real-time competition it came first on every variable and every lead time, ahead of six national weather centres and 34 teams. On real floods of the past decade, prediction skill rose by 76 to 92 percent — by 86 to 150 percent for the deadliest ones. It is open source, runs without a supercomputer, and Kenya’s weather office is testing it on the monsoon. “Just by correcting them, we can take an unskillful model and make it skillful,” Cohen said. The most useful AI result of the week does not talk at all — it just stops being wrong in the same direction.
Source: Gizmodo, October 14, 2026
Four Gemini Agents Finished Minecraft in 27 Hours, Dying 605 Times on the Way
A paper posted to Hugging Face on Friday, which Google DeepMind’s own account promoted, describes four Gemini 4 Argon agents, each driving a bot through the open Mineflayer library, starting in a random world and killing the Ender Dragon 27 hours and 14 minutes later. Nobody assigned roles or wrote game tools for them. They coordinated through in-game chat — 229 messages — and a shared memory, wrote 2,783 programs, and died 605 times. The authors call it the first end-to-end autonomous completion of Minecraft by AI agents. Then they list the small print themselves: one run, the first they tried; Easy difficulty, where hunger cannot kill you; bots that read the world as data rather than pixels; 22% of programs ending in errors, six server restarts, and nearly 14 hours on the dragon alone, at an estimated $810 to $1,620 for the session. A team that dies 605 times and still finishes is either very human or a strong argument for respawning.
The White House Superintelligence Task Force Gave Itself 120 Days
The superintelligence accord the AI chief executives signed with the White House earlier this month now has a body behind it. The task force adopted its charter after its first official meeting on Thursday, Semafor reported, and started a 120-day clock. It will ask the President’s Council of Advisors on Science and Technology to assess AI risk in three areas: cybersecurity and critical infrastructure, biosecurity, and “model control and accountability.” Director of National Intelligence Jay Clayton chairs it. Defense Undersecretary Emil Michael will draft rules for labs to report breaches and hacks; FTC chair Andrew Ferguson gets a law-enforcement committee with the Justice Department; a working group will put the accord into practice. Members include JD Vance, Scott Bessent and Pete Hegseth. What nobody has said is how any of it is enforced. A White House official offered export controls as “one tool” — the lever Commerce already pulled on Anthropic. A voluntary accord now has a clock, a chair and, if needed, the export office.
Source: Semafor, October 15, 2026
Perplexity Gave Away the Box It Keeps Its Agents In
Perplexity open-sourced SPACE on Thursday, the sandbox that every session of its Computer agent has run in since July, when it replaced a commercial provider. Each sandbox is its own small virtual machine with its own operating system; network access and the credentials the platform holds stay outside it. “SPACE treats the sandbox as untrusted,” the post says. In Perplexity’s published security test, agents made 108 attempts to break out: none got from the virtual machine to the host, though some found ways around the network restrictions, which the company says it has since closed. Akamai has launched it on its cloud, Crusoe is adding it to its managed services and CoreWeave has published setup guidance. All the numbers are Perplexity’s own. The line worth keeping is its premise: “Agent governance is an engineering problem.” Don’t ask the agent to stay in the room. Lock the door.
Source: Perplexity, October 15, 2026