Microsoft and OpenAI end their exclusive and revenue-sharing deal

Microsoft and OpenAI have ended their exclusive distribution agreement, allowing OpenAI to pursue partnerships with competing cloud providers like Amazon. In exchange for relinquishing exclusivity, Microsoft will no longer pay OpenAI a revenue share on OpenAI products resold via its cloud platform. This restructuring marks a significant shift in the strategic partnership as both companies seek broader market integration.

Read more

4TB of voice samples just stolen from 40k AI contractors at Mercor

The Lapsus$ leak of 4TB of Mercor data has compromised voice biometrics and government IDs for 40,000 AI contractors, providing ideal inputs for high-fidelity synthetic voice cloning. This breach facilitates sophisticated vishing and MFA bypass by pairing studio-quality reference audio with verified identity credentials. Technical mitigation involves transitioning away from voice-based authentication and employing forensic analysis to detect anomalies in codecs, prosody, and formant trajectories.

Read more

China blocks Meta's acquisition of AI startup Manus

China’s state planner has blocked Meta’s $2 billion acquisition of Manus, a Singapore-based AI startup with Chinese origins that develops general-purpose AI agents for coding, data analysis, and market research. The intervention by the NDRC follows a probe into export controls and reflects Beijing's crackdown on "Singapore-washing" by domestic tech firms seeking to bypass geopolitical investment restrictions. Meta intended to leverage Manus’s technology—which reached $100 million ARR in record time—to integrate advanced automation into its Meta AI assistant and enterprise offerings.

Read more

Talkie: a 13B vintage language model from 1930

talkie is a 13B LLM trained on 260B tokens of pre-1931 text to study model generalization, forecasting, and contamination-free evaluation. It utilizes a "modern twin" benchmark to analyze the performance gap caused by OCR noise and historical data distributions. The researchers employed a novel post-training pipeline using historical manuals and RLAIF to develop conversational capabilities without modern data leakage, with plans to scale to a 1T token GPT-3.5 class model.

Read more

The Prompt API

The Prompt API enables developers to interface with Gemini Nano locally in Chrome, supporting multimodal inputs including text, audio, and images. It features robust session management with context window tracking, session cloning, and structured output capabilities via JSON Schema. Currently available through origin trials on desktop, the API provides both request-based and streaming interfaces for low-latency LLM interactions directly in the browser.

Read more

ASI-Evolve: AI Accelerates AI

ASI-Evolve is an agentic framework for automated AI research that employs a closed-loop cycle to optimize data curation, neural architectures, and learning algorithms. By integrating a cognition base for human priors and a dedicated analyzer for insight distillation, the system discovered SOTA linear attention architectures and RL algorithms that significantly outperform GRPO. These results demonstrate the feasibility of using AI to accelerate foundational AI development through autonomous, long-horizon experimentation.

ArXiv

Microsoft Paper: LLMs Corrupt Your Documents When You Delegate (Arxiv.org)

DELEGATE-52 is a benchmark designed to evaluate LLM reliability in delegated document editing across 52 professional domains. Testing 19 LLMs reveals that even frontier models corrupt an average of 25% of document content over long workflows, with errors compounding silently over time. Performance degradation is exacerbated by document size and interaction length, and agentic tool use fails to mitigate these reliability issues.

ArXiv

Guess-Verify-Refine: Data-Aware Top-K for Sparse-Attention Decoding on Blackwell

Guess-Verify-Refine (GVR) is an exact Top-K algorithm designed to accelerate sparse-attention decoding on NVIDIA Blackwell by exploiting temporal correlation across decode steps. Integrated into TensorRT-LLM for DeepSeek-V3.2, GVR replaces standard radix-select with a predictive thresholding and ballot-free verification mechanism. It achieves up to a 1.88x operator speedup and a 7.52% TPOT improvement at 100K context while maintaining bit-exact outputs.

ArXiv

The Platonic Representation Hypothesis

Deep networks are converging toward a shared "platonic representation" of reality across diverse domains and modalities. As vision and language models scale, their internal data representations become increasingly aligned, suggesting a universal statistical model driven by common selective pressures.

ArXiv

Gatekeeping: A Partial History of Cold Fusion

This case study explores scientific gatekeeping through the history of cold fusion, which was ostracized in 1989 following reproducibility failures. Despite its long-standing status as a pariah field, the discipline has persisted and is currently experiencing a renaissance with renewed US and EU government funding.

ArXiv

49Agents – Infinite canvas IDE for AI agents

49 Agents is an open-source, 2D agentic IDE that replaces traditional terminal tabs with an infinite zoomable canvas for managing agents, terminals, and projects across multiple machines. It features zero-SSH connectivity, integrated Monaco editors, and real-time monitoring of Claude API usage and system resources. The architecture uses a WebSocket relay to enable multi-device access and self-hosting while ensuring terminal data remains ephemeral.

Repo

Waiting for LLMs Suck – Give your user a game

react-waiting-game is a zero-dependency React library providing 1-bit, single-button mini-games to improve UX during high-latency tasks like LLM inference or builds. It features five modular games with SSR-safe rendering, localStorage persistence for achievements, and a unified framework for power-ups and high scores. The library is highly customizable via props and designed for easy integration into async loading states.

Repo

Memory Guardian – open-source memory governance for AI agents

Memory Guardian is an active governance layer for AI agent memory that provides intelligent ingestion, task-aware retrieval, and automated lifecycle management. It addresses common agent failures by implementing multi-signal scoring—combining semantic similarity, importance, recency, and frequency—alongside NLI-based conflict detection and explainable retrieval traces. Designed for regulated environments, the system features strict multi-tenant isolation, memory consolidation, and deterministic audit trails for every storage and retrieval decision.

Repo

Tera – A Compiler‑Native UI Framework with Shared Runtime/AI Context

Terajs is a compiler-native UI framework for route-first, local-first web applications that replaces VDOM diffing with fine-grained reactivity and direct DOM bindings. It features a specialized local-first runtime with built-in mutation queues and conflict resolution, alongside an @terajs/adapter-ai and dedicated <ai> SFC blocks for providing structured state snapshots and metadata to LLMs. The framework includes integrated DevTools and a VS Code bridge that surface rich diagnostic context, facilitating AI-assisted debugging and development workflows.

Repo

Defeating AI by making knowledge accessible to Humans

PeakSlab is an offline-first PWA dictionary framework designed as a high-performance, hallucination-free alternative to using LLMs for knowledge retrieval. It utilizes a custom .peak binary format and a 37kb C-based WASM runtime to achieve significantly faster load times and smaller footprints than SQLite. By leveraging Zstandard compression and direct struct casting for binary search indexes, the project provides a lightweight, client-side solution for obscure language data that remains inaccessible to AI scrapers.

Repo