22 stories
The Machine Intelligence Research Institute has published an informal summary of a November 2024 paper by its Technical Governance Team on how international agreements to halt or limit AI development could be verified, covering chip location tracking, training run detection, and certified model evaluations.
MIRI's 125th newsletter calls on its community to support the opening weekend of AI documentary 'The AI Doc: Or How I Became an Apocaloptimist', reports new language editions of Nate Soares and Eliezer Yudkowsky's book 'If Anyone Builds It, Everyone Dies', and lists recent advocacy and policy events.
The Machine Intelligence Research Institute has published a question-and-answer guide for viewers of the documentary The AI Doc, covering the pace of AI capability gains, why models are described as grown rather than programmed, and MIRI's view that international coordination to halt frontier AI development is necessary.
The Machine Intelligence Research Institute has published a memo collecting statements from Chinese officials, scientists, and government bodies since 2017 that express willingness to participate in international AI governance, including calls for monitoring mechanisms and warnings about existential risk from superintelligent AI.
The Machine Intelligence Research Institute has published an informal summary of its Technical Governance Team's paper on AI governance, describing four possible global responses to advanced AI development and concluding that only one avoids what the authors term unacceptably high levels of catastrophic harm.
The Alignment Research Center has released a summary of how its published research fits into a unified technical vision, mapping the relationships between heuristic explanations, mechanism distinction, and alignment robustness as subproblems of preventing AI misalignment.
The Alignment Research Center has published a formal conjecture stating that any reversible circuit exhibiting a highly improbable property must have a polynomial-length structural explanation, building on a principle from Fields medalist Timothy Gowers.
The Alignment Research Center has published a post describing its current technical research direction: building mechanistic explanations of neural networks that can estimate the probability of rare or catastrophic outputs more efficiently than random sampling alone.
The Alignment Research Center has published AlgZoo, a collection of RNNs and transformers ranging from 8 to 1,408 parameters trained on algorithmic tasks, arguing the field cannot yet fully interpret even the 432-parameter models in the set.
Mozilla AI has announced llamafile 0.10.0, a rebuilt release that adds multimodal model support, tool calling, and an Anthropic Messages API endpoint, while restoring the project's signature single-executable portability across operating systems and CPU architectures.
Mozilla AI has published cq, an open-source system that lets AI coding agents query and contribute lessons learned, aiming to reduce duplicated effort and token waste when agents repeatedly hit the same errors in isolation.
On March 24, 2026, malicious versions of LiteLLM were uploaded to PyPI by a threat actor who obtained the maintainer's publishing credentials — exposing SSH keys, cloud credentials, and API keys while the GitHub source remained clean throughout.
Mozilla AI has begun work on Clawbolt, a purpose-built AI assistant for contractors and small trade businesses, designed to handle quoting, invoicing, scheduling, and payment follow-up through messaging apps the users already have.
Mozilla AI's Encoderfile project has moved from a compile-time Cargo-in-Cargo approach to a pre-built executable with an appended payload, eliminating the Rust toolchain requirement and reducing build times to sub-second on Linux and macOS.
The AI Now Institute press page links to an op-ed published in The Nation. The full op-ed text is not accessible from the AI Now teaser.
A report from AI Now Institute documents how gig nursing platforms backed by venture capital and private equity have pursued legislative campaigns in at least 17 states to define themselves as technology platforms rather than healthcare staffing agencies, seeking to avoid existing labor and safety regulations.
A new AI Now Institute report identifies Clipboard Health, ShiftKey, CareRev, and IntelyCare as leading a push to carve gig nursing platforms out of state healthcare staffing regulations, drawing on roughly $1.4 billion in venture capital and private equity funding.
Stability AI has become a full member of the Tech Coalition, a global alliance of technology companies working to prevent online child sexual exploitation and abuse.
A Databricks blog post claims US regulators rescinded SR 11-7 and related model risk guidance and replaced it with a principles-based framework. The post outlines a Lakehouse-based implementation. 4o4 / AI has not independently verified the regulatory changes with the Federal Reserve, FDIC, or OCC.
A 40-year-old in South Korea faces up to five years in prison after he used AI to generate a false wolf sighting, which Daejeon city authorities say diverted emergency resources during an active rescue operation.
Ars Technica, citing a Financial Times report, says a White House memo warns of deliberate Chinese campaigns to extract and clone US frontier AI models; Congress has received a recommendation to treat model extraction as industrial espionage.
Meta's chief people officer told staff in a memo that roughly 8,000 jobs will go in May and around 6,000 open roles will be closed, with the stated rationale being cost offsets for rising AI investment as Meta's 2026 capex guidance reaches up to $135 billion. The Verge is the sole outlet to have reviewed the memo, citing Bloomberg's publication of it.