A Pentagon investigation found that flawed intelligence, outdated imagery, and overreliance on AI contributed to a US missile strike on an Iranian school, showing how automated tools can magnify failures in high-stakes decisions. The other stories are less connected than consequential on their own.
Pruna’s Qwen-Image 2.1 adapters let local users generate or edit images in 5 or 8 steps instead of the base model’s longer process, with claimed speedups of up to 6.3x and no classifier-free guidance. They load on top of the unchanged Qwen model, text encoder, and VAE, so existing pipelines can try the faster path. The 8-step version is the quality-first option, but v0.1 does not yet match the base model visually.
Swift 1.5 gives developers a Qwen3.8-27B model that is positioned as stronger than Swift 1.0 on agentic and coding tasks while using fewer thinking tokens. The Hugging Face collection provides full BF16 weights plus versions for llama.cpp, LM Studio, Ollama, Apple Silicon, vLLM, NVIDIA, AMD, and other quantized runtimes, so teams can match the model to their hardware instead of starting with one format.
Julia 1 is a 144.3M-parameter model for turning supplied context, a question, and possible answers into classifications, routes, scores, or Boolean decisions. It reached 94% on a 100-example AG News pilot and 86% on DAIR Emotion, while scoring 64% on a Banking77 pilot, showing the model’s focused strengths and its limits with long label lists. Its 73.15% typed-decision score was 0.45 points above the supplied Jev reference.
Scotland may have nearly 30 AI data centres seeking grid capacity, adding up to 9.7GW, yet communities cannot see the full queue before projects enter planning. Developers can request electricity connections, sometimes for dates in the late 2030s, without first filing a planning application. The investigation found about two dozen visible projects totaling 7.3GW, exposing the gap between public records and the grid operator’s larger figure.
An OpenReview API exploit exposed the identities of authors, reviewers, and area chairs on more than 10,000 ICLR 2026 papers, putting the conference’s review process at risk. ICLR says malicious actors circulated the data, while third parties harassed and offered bribes to reviewers. The conference fixed the leak, reverted reviews, reassigned area chairs, and froze public comments during its response.
A Pentagon investigation says flawed intelligence, outdated imagery and overreliance on AI contributed to a US missile strike that killed 123 children at an Iranian elementary school. Two Tomahawk missiles hit the Shajarah Tayyebeh school and its grounds in Minab on Feb. 28, 2026. The probe describes a cascade of preventable failures, and Pentagon personnel knew within hours that the US was responsible.
Graciela Gómez Hernández died in California’s Otay Mountain Wilderness despite being one mile from a surveillance tower, showing that pervasive border monitoring does not guarantee timely rescue. The General Dynamics tower’s electro-optical and infrared cameras can track targets 5 to 7 miles away, while newer systems use AI to distinguish people from animals. The wider network has cost more than a billion dollars, yet migrants continue to die in the terrain it watches.
A Medicare pilot now uses AI to authorize or deny certain care in six states, adding prior authorization where doctors previously did not need it and creating delays for seniors seeking treatment. WISeR covers around a dozen services, including nerve stimulation, epidural steroid injections, and cervical fusions. Providers reported technical problems, puzzling denials, and patients waiting in pain; the pilot is planned to run through 2031.
China, Iran, and private groups in Israel have launched autonomous influence campaigns, making election manipulation cheaper and harder to attribute. The campaigns use Chinese open-source models to create synthetic personas, generate media, and comment at scale, without the safety controls common in proprietary systems. That adds volume and apparent authenticity, while weakening the safeguards that make the operators identifiable.
The U.S. and Russia weakened proposed safeguards for lethal autonomous weapons, leaving efforts to regulate AI arms less protective for anyone relying on human control over battlefield decisions. Hundreds of diplomats had been working in Switzerland on what could become the first agreement governing these weapons. The negotiations turned aggressive on their final day, according to The Washington Post.
RAND argues that the U.S. should spend now to preserve freedom of action on superintelligence, rather than prematurely choosing a single path such as monopoly, cooperation, deterrence, or a pause. The paper identifies seven strategy archetypes but says current evidence cannot settle five pivotal uncertainties, so early commitment could close off options before the technology’s direction is clear.
OpenAI’s new independent mathematics advisory group will review the significance and release of emerging AI results, giving mathematicians a public channel to challenge the company as its internal model claims more than 100 long-standing problems solved. The group can offer unsolicited advice, publish its views, and set its own membership, but it will not advise OpenAI on how fast to advance the model. It will also advise on using AI tools for mathematical research and learning.
Prime Agent gives coding models a persistent workspace where they can spawn, update, and message sub-agents, making long-running work less dependent on fixed prompts and tools. Its Recursive Language Model treats context as a variable and delegation as function calls inside a REPL, while the open-source harness can be installed with one command. The design also lets separate Prime Agent sessions communicate directly.
You’re caught up.
That was all 13 stories in the 27 September issue.