Abu Dhabi's Institute of Foundation Models has released K2 Horizon, a family of six artificial-intelligence models ranging from 0.9 billion to 375 billion parameters. The smallest is designed for constrained devices such as watches and phones, while larger versions target local servers and enterprise systems. IFM says it has published the models' weights, training code, data, methods, intermediate checkpoints, logs, and evaluations so researchers can inspect, reproduce, and adapt the work. Performance claims still come from the developer, yet the release exposes far more of the manufacturing process than the downloadable weights commonly marketed as openness.

A researcher receiving only model weights gets the finished object without a useful record of how its behavior was formed. Access to data mixtures, filtering methods, code, and checkpoints makes different questions possible. Investigators can trace a bias toward a source category, examine whether an evaluation rewarded an unintended shortcut, or compare the point at which a capability appeared during training. Reproducing the largest model still demands money, hardware, and electricity beyond most universities or independent laboratories. Transparency therefore does not erase the inequality of scale. It gives outsiders evidence with which to challenge a laboratory's account of its own system. K2 Horizon places six machines on the table with their casings removed; the practical test begins when researchers inspect the files, rerun a checkpoint, and publish what the original announcement missed.

An Environmental Protection Agency proposal would remove the federal requirement that states notify the public and invite comments before granting air-pollution permits to facilities classified as minor sources. The Associated Press reported that the change could affect data centers and supporting equipment as companies build infrastructure for artificial intelligence. States could keep their own participation rules, although the federal minimum would disappear. The EPA says the proposal leaves emissions standards intact, reduces administrative burdens, and gives state and local agencies control. A coalition of nearly 200 environmental, health, and community groups warns that residents in states with weak rules could learn about projects only after approval.

The dispute makes public participation part of AI infrastructure. A data center is assembled through permits, utility agreements, cooling systems, backup generators, and land decisions long before a model answers its first prompt. Notice and comment give neighbors a narrow moment to examine that bargain while changes are possible. Removing the common floor would make a household's influence depend on local rules and its ability to discover an application buried in an administrative process. The result could be a geography of unequal visibility. Some communities debate the machines arriving beside them; others meet the building after the decision. Artificial intelligence may be sold as a national race, yet its air, noise, water use, and power demand settle at specific addresses. A permit issued without witnesses turns technological ambition into a local fact before the public can speak.

OpenAI has begun rolling out GPT-6 Astra, a model it says reaches 99.9 percent on ARC-AGI-3, a benchmark for artificial general intelligence (AGI), and human-level action efficiency on 96 percent of that benchmark's levels. Astra is the first OpenAI model classified at the Critical cybersecurity threshold, meaning that with tools and access it can discover unknown flaws and develop exploits across hardened systems without step-by-step human direction. OpenAI delayed parts of the release to strengthen containment and is initially limiting access. It does not call Astra AGI.

OpenAI's charter defines AGI as a highly autonomous system that outperforms humans at most economically valuable work. One saturated benchmark cannot prove that claim, and Astra still falls below the company's High threshold for AI self-improvement. Yet a turning point can arrive before its official name. Astra combines reasoning with direct computer use across long professional workflows. That changes the unit of automation from an answer to a finished assignment. Scientists may compress weeks of technical work. Offices may remove layers of junior labor. Cyber defenders and attackers may search for the same unseen vulnerability at machine speed. The human consequence depends on who can delegate, who owns the infrastructure, who is displaced and who retains authority to stop the task. If Astra is remembered as an AGI milestone, the evidence will come from workplaces, laboratories, public agencies and security incidents rather than the launch page.