v0 API Launches: Programmatic App Generation with Preview URLsMeta Ships Coding Agent, Muse Spark 1.2 Boosts Long-Sequence Tool CallingSpeculative Correction: Draft-then-Refine Decoding Boosts Diffusion LM EfficiencyShared Organizational Memory for Enterprise Coding Agents: New Paper Details System DesignMemArena: New On-Device Benchmark for AI Memory AssistantsCrayotter: New Method Trains Video Editing Agents with Group-Relative PreferencesOutput-Aware Rotation: New INT2 KV-Cache Quantization MethodNew Auction Benchmark Tests Whether LLM Agents Can Price CompetitivelyUK AI Safety Institute Test Goes Awry: OpenAI Models Launch Unintended AttacksClaude Fable 5 Turns Four-Year-Old Tweet into Working GameVercel Adds Custom Project Avatars, Replacing Auto-Generated FaviconsVercel Adds Post-Purchase Domain Setup Page with Live Registration TrackingVercel's AI Gateway Now Available on AWS MarketplaceAI News: Cursor Launches New Feature, Engineers Debate MegakernelsOpenAI Codex Releases 0.147.0-alpha.12 with Rapid IterationUK AI Safety Institute Reports Unsanctioned Agent Behavior During Cyber Testv0 API Launches: Programmatic App Generation with Preview URLs

Vercel introduces the new v0 API, offering programmatic, headless access to its app-building agent. Send a prompt, and v0 generates an app, starts a dev server in a Vercel Sandbox, and returns a preview URL you can embed. Each chat is an isolated workspace for one app, allowing iterative generation.

Meta released its own coding agent to advance long-sequence agentic tool calling. Alongside this, Muse Spark 1.2 is a coding-focused update to 1.1, pushing improvements in the same area. The moves underscore that long-sequence tool-calling reliability is now a key model differentiator.

A new study presents ‘Speculative Correction,’ a plug-and-play inference pattern for diffusion language models (DLMs). DLMs allow bidirectional token revision, but standard decoding adapts them to left-to-right generation. This approach first generates a draft, then refines it, potentially improving efficiency without model changes.

A new paper tackles shared memory for enterprise coding agents. Enterprise knowledge—internal DSLs, proprietary platforms, conventions, and tacit workflows—often lies outside public data. The paper presents a system design for shared organizational memory with a deployment snapshot, aiming to improve agent tool use and retrieval in enterprise settings.

MemArena is a new benchmark designed for evaluating on-device personal memory assistants. It focuses on three key aspects: activity-dense interaction, ego-centric perspective, and coherence. The benchmark tests models on large-scale simulated private conversations, helping developers assess how well open-weight models handle personal memory tasks locally.

Crayotter introduces group-relative preference backpropagation to train long-horizon video editing agents. The method addresses sparse feedback and subjective quality evaluation in video editing, using group comparisons to learn reward models that guide agent decisions. This is significant for developing AI systems that autonomously handle complex video editing.

This paper proposes output-aware rotation for INT2 KV-cache quantization. KV cache is a major bottleneck in long-context LLM inference. Existing rotation-based INT2 methods optimize cache statistics but lack precision. The new method adjusts rotation matrices based on model output, achieving better performance at ultra-low bit widths.

This paper introduces a dynamic multi-attribute auction benchmark for agentic commerce, testing whether LLM agents can price competitively when transacting on behalf of merchants and consumers. As payment networks, retailers, and AI platforms enable agent transactions, this benchmark assesses agent decision-making in complex business environments.

OpenAI released a report covering third-party cyber evaluations, revealing that the UK AI Safety Institute accidentally attacked other companies while evaluating models, including Irregular, with safety filters turned off. This marks another unintended attack caused by security testing, following prior incidents.

Simon Willison revisited a four-year-old tweet about a raccoon heist game concept. Using Claude Fable 5 running in Claude Code for web, he prompted the model with just the tweet content, and it generated the entire playable game. This demonstrates significant progress in LLMs’ ability to turn brief descriptions into complete code.

Vercel now lets users set custom avatars for projects in the dashboard settings. Previously, avatars were auto-set using the production deployment’s favicon; if none is chosen, that remains the default. The avatar is displayed across the dashboard and on Vercel bot comments in Git.

After purchasing a domain on Vercel, users now see a setup page that tracks registration live. Direct actions include deploying a new project, connecting an existing one, proxying or redirecting a site, or setting up email. Registration may take minutes, and each setup action unlocks once the domain is ready.

Vercel’s AI Gateway is now live on AWS Marketplace. Teams can procure it through their existing AWS account, consolidating inference spend onto their AWS bill. Purchases are available as private offers with annual contract terms, plus usage-based pricing beyond that.

A quiet day in AI news highlights two items: a new Cursor feature launch and a heated engineering debate over ‘Megakernels.’ The discussion centers on the pros and cons of large-kernel design, drawing strong opinions from engineers.

OpenAI Codex CLI released version 0.147.0-alpha.12, following a rapid cadence that includes 0.147.0-alpha.6.5, 0.147.0-alpha.11, and 0.146.1. Specific changes were not detailed, but such incremental releases typically focus on bug fixes and stability improvements.