NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-07-04ENGLISH EDITION
This issue
—
All time
—

AI Blog

1 story
01

State of the Art Models Exhibit Regression in Schema Enforcement for Tool Calls

Developer Armin Ronacher reported a performance regression in tool-calling reliability when testing advanced Anthropic models, specifically Claude Opus 4.8 and Sonnet 5, within the Pi application. While older Anthropic models successfully adhered to nested tool schemas, these newer models frequently generated non-existent arguments inside nested edit arrays, violating strict schema constraints. This behavior highlights ongoing challenges in how frontier models serialize structured inputs and adhere to developer-defined constraints, pointing to a potential need for sampler-level grammar-aware decoding rather than relying purely on model reasoning. (source: https://lucumr.pocoo.org/2026/7/4/better-models-worse-tools/)

Claude Opus 4.8Tool CallingSchema ValidationGrammar-Aware DecodingAnthropicLarge Language ModelAI AgentGenerative AI

Hacker News

5 stories
01

Agentic coding notes from Galapagos Island

Software engineer Dan Luu published an in-depth analysis detailing his hands-on experience using autonomous agentic AI loops for iterative code generation, testing, and debugging. The report documents how current agentic frameworks handle boilerplates and prototyping, while highlighting major obstacles such as state tracking across long contexts and error propagation where agents become trapped in infinite loops. This analysis underscores that while AI agents accelerate standard tasks, human guidance remains necessary to validate architectural logic and correct systemic deviations. (source: https://danluu.com/ai-coding/#appendix-agentic-loops-and-writing-this-post)

Agentic LoopsAI AgentsSoftware EngineeringCode GenerationAutonomous AgentsArtificial IntelligenceAI AgentGenerative AI
02

Potential session/cache leakage between workspace instances or consumer accounts

A potential security vulnerability has been identified and reported within the Anthropic Claude Code developer command-line assistant repository. The issue reports risks associated with stateful data, authentication sessions, or cached memory crossing security boundaries and leaking between distinct workspace instances or consumer accounts. This highlights a critical isolation and data governance vulnerability inside developer-focused LLM tooling, which requires runtime isolation and cryptographic session separation to prevent unauthorized context leakage. (source: https://github.com/anthropics/claude-code/issues/74066)

Claude CodeSession LeakageCache IsolationSecurity VulnerabilityWorkspace SecurityArtificial IntelligenceGenerative AIOthers
03

Google Books (or similar) all book scans – $200k bounty (2025)

Anna's Archive has announced a $200,000 bounty program for the year 2025 to acquire comprehensive book scans from major digital repositories like Google Books. This crowdsourcing initiative aims to compile high-quality, large-scale digital text corpora for cultural preservation and training foundation language models. The campaign underscores the rising industry demand for massive datasets required to train linguistic reasoning architectures. (source: https://software.annas-archive.gl/AnnaArchivist/annas-archive/-/work_items/234)

Google BooksDigital ArchivingDataset AcquisitionBounty ProgramData CrowdsourcingLarge Language ModelNatural Language ProcessingOthers
04

Fable created novel 4D splat format

Fable has announced a novel 4D splat format that expands on standard 3D Gaussian Splatting techniques to capture dynamic spatial data. By introducing a temporal fourth dimension, this format facilitates high-fidelity, real-time rendering of moving scenes and volumetric video with minimal computational overhead. The technology represents a notable advancement for virtual reality, gaming, and computer vision fields seeking to bridge static reconstructions with fully dynamic generative environments. (source: https://adamraudonis.github.io/splats4D/)

4D Splatting3D Gaussian SplattingComputer VisionReal-Time RenderingVolumetric VideoComputer VisionGenerative AIOthers
05

Plein Air

The Plein Air digital installation has showcased a new real-time generative art pipeline that translates organic environmental data into algorithmic visual renderings. Inspired by traditional French outdoor painting, the system processes continuous environmental data streams to generate dynamic visual landscapes that shift with simulated atmospheric and light changes. The project demonstrates a unique intersection of graphic frameworks, physical sensory inputs, and generative computer aesthetics. (source: https://art.joonas.wtf/)

Generative ArtCreative CodingDigital RenderingReal-time DataVisual GenerationGenerative AIComputer VisionOthers

Twitter

3 stories
01

Fable 5 System Card Documentation Released for Public Review

Thomas Wolf has highlighted the release of the Fable 5 System Card, offering detailed documentation for the advanced artificial intelligence model. This technical document details the model's architecture, safety guidelines, and operational capabilities to promote transparency within the AI community. By publishing this card, the project enables developers and researchers to evaluate the model's performance limits, technical foundations, and safety boundaries to facilitate safer integration and broader adoption. (source: https://x.com/Thom_Wolf/status/2073350943886758318)

Fable 5System CardAI TransparencyModel DocumentationArtificial IntelligenceArtificial IntelligenceLarge Language ModelResearch Progress
02

Luma Labs Releases Ray 3.2 For Advanced Vision Capabilities

Luma Labs has announced the release of Ray 3.2, a significant update to their vision-focused artificial intelligence suite. The new model update integrates advanced architecture to optimize multimodal interactions and high-fidelity video generation. Designed to bypass traditional trial-and-error workflows, Ray 3.2 delivers streamlined, precise control over generative outputs to improve efficiency in media production pipelines and visual storytelling applications. (source: https://x.com/LumaLabsAI/status/2073211416044069225)

Luma LabsRay 3.2Generative AIVision AIMedia ProductionArtificial IntelligenceMultimodalGenerative AI
03

Kling AI Hosts NextGen Campus Awards Ceremony For Creative Video Talent

Kling AI has announced its NextGen Awards Ceremony, celebrating emerging student and early-stage filmmakers who leverage generative video technology. The event highlights how creators use Kling AI tools to produce high-fidelity 4K video content from experimental conceptual designs. By showcasing these student productions, Kling AI demonstrates the growing integration of accessible generative video tools into academic and independent cinematic workflows. (source: https://x.com/Kling_ai/status/2073240591463223578)

Kling AIGenerative VideoNextGen AwardsCinematic Storytelling4K VideoGenerative AIMultimodalIndustry News