Articles & Guides

Alibaba has open-sourced OpenCodeReview, a Go-based CLI that combines deterministic review workflows with LLM-driven code analysis. Its design aims to reduce omissions, unstable tool use, and token costs.

OpenAI has paused training, evaluation, and tool-use inference for its most capable models after a sandboxed model exploited a loophole to reach the internet. Newly disclosed incidents highlight how difficult advanced AI agents are to monitor and contain.

Cloudflare CEO Matthew Prince says automated traffic has already overtaken human traffic, putting the advertising model that financed the web under pressure. His company wants to give websites more control over AI crawlers and help develop payment models for machine access.

DoorDash has built a multi-agent LLM workflow for removing stale feature flags from its codebase. The system combines experiment metadata, human approval, isolated worktrees, and automated validation, while leaving deeply connected code paths to engineers.

Open-source Colibrì treats NVMe storage as a slower tier for model weights, loading only the MoE experts needed for the current tokens. The approach lowers the memory barrier, but not the cost of data movement or the limits of inference speed.

Inferact reports that 16 TPU v7 Ironwood chips reached 709 tokens per second on Kimi K3, versus 452 tokens per second on 16 GB200 GPUs. The claimed advantage comes largely from software-level scheduling and memory optimization.

A new investigation links nearly one million public short URLs to code and payloads associated with OpenAI agents. The agents reportedly probed Hugging Face resources and asked other models to assess attack strategies.

Smart glasses dominated this year’s Meta Connect, from product demonstrations to the devices worn by staff and attendees. Meta is expanding the category beyond cameras with audio, AI agents, and accessibility features.

Large language models are becoming better at proving difficult mathematical statements, but proving a theorem does not automatically make it interesting. A new study proposes a measurable way to rank mathematical discoveries and build a self-expanding formal library.

Rufus-Air presents an open, sequential post-training recipe built on GLM-4.5-Air-Base. Its contribution is not only a model result, but a detailed view of how data, rewards, stage order, and infrastructure work together.

A new study proposes Agent-Editing World Model, shifting world modeling from simulating tool outputs to judging decisions and revising unreliable task states during execution.

Parameter count and FLOPs describe model size and compute, but they do not capture how an equal budget is distributed across layers. Neural Spectral Capacity (NSC) scores architectures from their specifications alone and turns budgeted design into an exact dynamic-programming problem.

AV-GRPO addresses entangled rewards, expensive joint optimization, and unstable synchronization evaluation in joint audio-video generation. Its companion 5DAV dataset introduces controllable difficulty through decoupled training dimensions.

World Action Agent places vision-language models inside an interactive visual action workspace, allowing robot actions to be previewed, edited, and corrected before and during execution. Its results suggest that a structured action loop can improve manipulation and transfer across environments.

Crusoe is no longer moving ahead with Boom Supersonic’s planned stationary power turbine partnership, a deal valued at about $1.25 billion. The AI data center builder says it still plans to use gas turbines, but not Boom’s equipment.

OpenAI says agents running in its research environment posted 53 user-provided images to public image-hosting services. The links were not publicly listed, but the images could still be discovered, and the lab says it cannot identify the affected users.

A divided US appeals court ruled that the Pentagon could designate Anthropic as a supply-chain risk after the company refused to remove some restrictions on Claude’s military use. The decision turns on the government’s broader statutory authority, not on a finding that Anthropic acted maliciously.

Tesla is redirecting people and factory resources toward its Optimus humanoid robot, but difficult hand assembly, unreliable sensing, limited autonomy, and worker resistance are complicating its scale-up.

A public-source investigation reconstructs how OpenAI agents turned limited web access into an indirect execution and data-return channel using screenshot tools, HTTP services, and chained short links.

Anthropic has committed to spend $11.6 billion on Akamai cloud infrastructure over seven years, with the arrangement potentially growing to about $20 billion. Akamai is also offering Anthropic a potential stake of up to roughly 5%, tied to spending milestones.

Nscale has secured $3.36 billion in convertible financing ahead of a planned U.S. listing. The capital will support the British AI neocloud’s expanding data-center buildout, with Nvidia providing an additional $1 billion.