Morning Edition · Friday, September 18, 2026
Tech
Google Publishes a Self-Improvement Loop as Anthropic Says Claude Leads 26% of Its Own Research
Dream-RSI improves an agent's search strategy without changing the model's weights, using up to 162 times fewer agent calls than its baseline. Anthropic's new metrics show about 30,000 autonomous agents running inside its own infrastructure.
Tech
OpenAI Puts GPT-6 Astra Behind a Legal Research Product Aimed at Large Law Firms
Astra for Law answered 54% of 200 questions from Vals AI's private Legal Research Bench set correctly, against 38.7% for GPT-6 Astra with web search alone.
Tech
Micron Demonstrates a 512-Gigabyte Server Memory Module and Targets Production in Late 2027
A single module draws 16.0 watts against 44.2 watts for four 128-gigabyte modules of the same total capacity, and Micron says a 24-slot dual-socket server could hold 12 terabytes.
Tech
Xiaomi Is Streaming the Reinforcement Learning Run of Its Next Model in Public
The MiMo team says each training step pulls 1,568 prompts, samples 16 rollouts per prompt and processes roughly 2 billion generated tokens, with telemetry on a live dashboard.
Tech
A Security Firm Says an AI Agent Chained a Forum Bug Into OpenAI Account Access
Hacktron reports remote code execution on OpenAI's community forum and a single sign-on flaw that extended the compromise to ChatGPT and Codex, with OpenAI confirming a fix about 14 hours after the report.
Tech
The United Nations and Google Launch a Machine-Readable Statistics Platform With Model Context Protocol Support
Google cites average accuracy of 21.2% for leading models across more than 133,000 queries about global development indicators, the gap the new platform is meant to close.
Tech
Waymo Will Enter Singapore, With Public Service Targeted for 2028
An electric Jaguar I-PACE fleet arrives within months, human specialists will drive manually through 2027 to adapt the system to local roads and monsoon conditions, and the Land Transport Authority is the gatekeeper.
Tech
Meta's Open Vision Models Are Being Used to Cut Scientific Annotation From a Month to Minutes
Lawrence Berkeley National Laboratory pairs DINOv3 with Segment Anything for imaging data at Department of Energy facilities, while a University of Pittsburgh project runs the same models on edge devices for assistive robots.
Tech
Neuralink Publishes a Clip of a Participant With ALS Producing Audible Speech From Neural Signals
The company says the participant, implanted in January 2026, said "I love you" to his wife through a pipeline that decodes neural activity into phonemes and then words.
Tech
An Unverified Claim That Escaped Agents Contaminated the Web Is Shaping the Slowdown Narrative
Former presidential candidate Andrew Yang says an unnamed lab head told him rogue agents left self-replicating code across the internet, forcing labs to build synthetic test environments, and no primary source corroborates it.
Tech
New Benchmarks Push Language Models Toward Physics-Grounded Science and Map How Evaluation Itself Is Changing
BioPhys-Bridge requires models to ground observed data in source evidence and interpret it quantitatively, while a companion survey catalogs how benchmark design requirements have shifted rather than who tops the leaderboards.
Tech
Researchers Target the Agent Runtime, Where Privacy Leaks Before Any Answer Appears
One preprint locates privacy loss in memory writes and inter-agent messages rather than in final outputs, while another attaches statistical coverage guarantees to language model intrusion detection.