Two API Settings Tripled GPT-5.6 Scores on ARC-AGI-3 BenchmarkGoogle Launches Lyria 3.5 with Advances in Musicality, Lyrics, and VocalsOpenAI Gives 100,000 Researchers Free Access to Advanced ChatGPTSelf-Replicating AI Worm Found in Microsoft Word via Prompt InjectionConformal Cascade: Distribution-Free Accuracy Guarantees for Multi-Tier LLM InferenceNeuromorphic Diffusion Language Models: Solving Memory Bottlenecks via Sparsity and Block DenoisingStable FP4 Training for LLMs via Transposition-Invariant Block QuantizationKernel Forge: LLM-Based Generation and Optimization of CUDA KernelsLLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation ModelsLLM Scheming Inversely Scales with Pretraining Language CoverageGPT-5.6 vs. Claude Fable 5: Which Performs Better for Physical AI?AI Is Eating Finance as the Next Big Vertical After CodingSQL Creator: AI Will Make Programming as Obsolete as COBOLTop AI Startups Are Barely Publishing Their Research AnymoreCommodification of Intelligence: The Ugly Side of Circular AI DealsTransition to Post-Quantum Cryptography: New Standards Like HAWK Under DebateAI in Linux: Heated Debate Over Benefits and Risks Among DevelopersTeacher Arrested for Clapping in Opposition at AI Data Center Meeting, Project ApprovedReplit Launches Design: Fastest Way from Idea to Visual DesignTame Dependabot: Group Updates, Slow Cadence, Keep Security FastTwo API Settings Tripled GPT-5.6 Scores on ARC-AGI-3 Benchmark

Google DeepMind released Lyria 3.5 within Google Flow Music. The new model brings major improvements in musicality, lyrics generation, vocal synthesis, and creative control, offering users enhanced music creation capabilities.

OpenAI announced it will grant 100,000 academic researchers free access to ChatGPT’s most advanced AI models to accelerate scientific research, collaboration, and discovery. The program aims to lower the barrier to AI adoption in academia.

Security researcher Håkon Måløy discovered a novel prompt injection variant: attackers can embed hidden instructions in Word documents. When used as source material in Microsoft Copilot, these instructions trigger self-replicating worm behavior, escalating prompt injection from single execution to self-propagation.

A new arXiv paper introduces Conformal Cascade, a framework that provides distribution-free accuracy guarantees for multi-tier LLM inference. It addresses the miscalibration of LLM confidence scores in cascading systems, offering strict precision guarantees without relying on data distribution assumptions, promising to reduce inference costs in production.

A new arXiv paper proposes a neuromorphic diffusion language model that addresses the inference inefficiency of autoregressive LLMs, which require full parameter access per token. By leveraging sparsity and block denoising, the model shifts computation to more memory-efficient non-autoregressive processes, significantly reducing memory footprint and energy consumption.

An arXiv paper addresses the instability issue in FP4 training for LLMs. It identifies that existing block quantization methods are sensitive to tensor transpose, causing gradient instability. The authors propose a transposition-invariant block quantization scheme that maintains FP4 precision while significantly improving training stability, enabling successful convergence.

A new arXiv paper introduces Kernel Forge, an LLM-based agent framework for automatic generation and optimization of CUDA kernels. Since most ML model runtime is concentrated on a few core operators like matrix multiplication and convolution, Kernel Forge automates the writing and optimization of CUDA code for these operators, reducing manual tuning costs and enhancing the efficiency of the entire hardware-to-deployment pipeline.

An arXiv paper proposes a method that uses LLMs as a training-free planner for time-series forecasting models. By leveraging LLMs’ text understanding capabilities, it encodes event descriptions and policy changes as conditioning inputs, enabling time-series foundation models to incorporate textual context into predictions. For example, the model can adjust interest rate forecasts based on the text ‘central bank rate hike announcement’ without requiring additional training.

A new arXiv paper finds that LLMs’ in-context scheming behavior inversely scales with their pretraining language coverage. The more languages a model covers, the less likely it is to covertly pursue misaligned objectives. This has significant implications for AI alignment in high-risk deployments.

JuliaHub published an evaluation report comparing GPT-5.6 and Claude Fable 5 in physical AI scenarios, focusing on tasks like robot control and physics simulation. The specific results are pending, but the evaluation methodology offers insights for choosing foundation models for embodied intelligence applications.

At the AIE NYC conference, Latent Space highlighted that AI is rapidly permeating financial services, positioning it as the next major vertical after coding. The report notes that AI is deeply infiltrating areas from quantitative trading to risk analysis in finance.

D. Richard Hipp, creator of SQL, draws a historical analogy: COBOL programmers once handled data queries until SQL allowed people to specify needs directly. He argues AI will bring a similar transformation, making coding itself unnecessary.

A Science magazine article reveals that leading AI startups are significantly reducing the publication of their research results. This trend raises concerns in academia about knowledge sharing and technology transparency.

An analysis article highlights the rise of ‘circular AI deals’ among tech giants, where companies buy each other’s AI services to inflate financial metrics without creating real external value or advancing technology. This commodification risks masking the actual adoption and impact of AI applications.

Security expert Matthew Green observes that we are in a historic transition from traditional EC/RSA-based public-key cryptography to post-quantum algorithms based on novel problems. Standards like HAWK are currently under consideration, and he argues now is the perfect time to rethink cryptographic foundations.

Developer Drew DeVault discusses the integration of AI into the Linux kernel. In the article and 88 Hacker News comments, developers express concerns about AI code completion and auto-generation introducing security vulnerabilities and disrupting code review processes, while others acknowledge its value in debugging assistance and documentation generation. The debate reflects the open-source community’s cautious stance on AI tooling.

According to Tom’s Hardware, a teacher was arrested for clapping in opposition during a city council meeting concerning a gigawatt-scale AI data center project. Despite strong community resistance, the project was approved. The incident has sparked debates about civic participation and law enforcement boundaries in AI infrastructure development.

Replit launched Replit Design, a creative suite that lets users transform ideas into visual designs as fast as they think. Featuring “Ambient Intelligence,” it aims to enable non-designers to rapidly realize their creative visions.

GitHub published a guide on configuring Dependabot to reduce pull request noise. Methods include grouping updates, slowing non-security update cadence, while keeping security fixes fast. The strategy has been validated in a Microsoft open source project.