Weekly Intel - 2026-09-13

The theme this week is where AI actually runs, and who gets to say no. Models are moving onto local devices and cheaper endpoints, while the guardrails around them, environmental review, corporate governance, frontier pacing, get tested or stripped away.

AI & Software

DeepSeek v4.1 Flash DeepSeek released DeepSeek-V4.1-Flash, the smallest model in a new architecture family and its first with native visual understanding. The company positions it for faster inference and higher throughput, and describes the architecture as designed to scale to larger models.

ChatGPT Images 2.5 OpenAI released ChatGPT Images 2.5, with sharper detail, more precise editing, and up to 50% lower generation latency than Images 2.0. The update covers natural lighting, texture, subject preservation from reference photos, and multi-turn editing reliability. Users create more than 3 billion images a week across ChatGPT Images and the GPT-Image API models. New features include Sketch, which lets users draw directly in ChatGPT as a reference, and Templates for formats like flyers and product pages.

Why are AI agents lying, cheating and coordinating? Yoshua Bengio writes that over the past few months AI agents have taken actions that would count as crimes if a human performed them, escaped containment to cheat on assigned tasks while attempting to evade detection, and coordinated toward goals nobody specified, including launching cyber attacks. The post situates these behaviors in the longer history of AI systems acting in unintended ways, which researchers call misalignment, offers hypotheses about the causes, and concludes that the behavior could grow more severe as capabilities grow.

AI Industry Moves

We must pace the frontier Anthropic CEO Dario Amodei argues AI could cure most major diseases within five to ten years, accelerate economic growth, and expand abundance, while carrying risks that include loss of control over AI systems, misuse for cyberattacks and bioterrorism, and economic disruption. Drawing on twelve years in the field and his own experience with cancer, he calls for pacing development at the frontier, and lays out a three-part plan for doing it. Anthropic is unilaterally committing to the first step: permanent, employee-level access for third-party evaluators.

Desert Ant Labs: local, fast models that run on device Desert Ant Labs, a European lab focused on on-device intelligence, launched with 18 small specialized models for audio, vision and text, 12 stable and six in beta, reachable through one SDK for Swift, Kotlin and JavaScript. The models run locally on hardware as old as a five-year-old phone, answer in milliseconds, and carry no inference cost. Voz transcribes 10 minutes of audio in two seconds on an iPhone, 4.7x faster than Whisper. Clear is a 9MB model that upgrades laptop recordings toward studio quality.

Legal & Regulation

The EPA is planning to scrap public review rules for data center pollution The Environmental Protection Agency plans to eliminate a federal requirement that states notify the public and allow public comment before approving certain air-pollution permits for industrial facilities, a category that includes data centers and the power plants built to supply them. Last month the agency also proposed letting developers begin building data centers before their permits are approved. Seven in 10 Americans oppose building AI data centers in or near their communities, and the rural South, home to a disproportionate share of Black communities, is among the areas affected.

Paramount caught using an astroturf group to manufacture support for its merger Paramount deployed an astroturf group, a fake grassroots organization, to create the appearance of public support for its $111 billion merger with Warner Bros. Discovery. The broader campaign includes labeling deal critics as antisemitic, threatening to relocate the company out of California if states enforce antitrust laws, and coordinating PR and press leaks with political allies to promote media consolidation. The merger already faces criticism over its effects on labor, creatives, consumers and market competition.

Tech Industry

Automattic’s board forces out Matt Mullenweg, who says days later he is back Automattic’s board voted on 9 September to place founder and CEO Matt Mullenweg on a paid leave of absence against his wishes, naming CFO Mark Davies interim CEO. Mullenweg said the members who voted for it included Davies, Ann Dunwoody, Toni Schneider and Sue Decker, that he received the resolution 50 minutes before the meeting, and that he voted against it. Two days later he posted in Automattic’s Slack that he is back in control as CEO, saying board members had “conspired” behind his back. Automattic has not responded to requests for comment. Mary Hubbard, Executive Director of WordPress.org, wrote on X that she was “relieved and pleased” to see him return.

Economy & Labor

What will our economic future look like? Anthropic released Version 1.0 of its Economic Scenario Explorer, a tool built on the framework in its technical report, Economic Scenarios for Transformative AI. The explorer isolates a few key economic forces and omits others, including policy responses, business cycles, aggregate demand and financial market disruptions, and potential catastrophic risks. Anthropic describes it as a work in progress, and notes that economists including Daron Acemoglu, David Autor and Emi Nakamura reviewed an early draft of the underlying technical report.

That’s what I’m watching. What caught your attention this week?

-Eric