The rest of the day's AI news, briefly. One page per day. Page 1 of 4.
Perplexity releases open-weight pplx-decider-v1-27b and launches low-cost Decisions API · MIT researchers introduce VISTA visual harness achieving perfect score on ARC-AGI-3 · UC Berkeley researchers unveil RPG framework for weight-free robot skill self-improvement
Researchers establish scaling laws revealing when AI web text degrades model training · Looped mixture of experts matches double-sized models on reasoning under fixed compute · Tactile curiosity framework enables robots to learn complex manipulation without human demonstrations
White House and major AI tech firms sign morally binding safety accord · Anthropic finds open-weight GLM-5.3 matches Claude Mythos on cyber exploit benchmarks · OpenAI launches Decisions API powered by GPT-6 Luna for real-time routing
Telescopic training enables language models to dynamically scale compute across arbitrary layer depths · Interleaved reinforcement learning trains unified multimodal models to iteratively self-correct image generation · Foil architecture optimizes looped mixture-of-experts by flattening expert layers and untying attention
Nvidia releases Open Agent Safety Platform to stop autonomous software breaches · Kaggle and DeepMind launch Game Arena for dynamic LLM competitive evaluation · Self-supervised confidence fine-tuning cuts reasoning model token output by twenty-five percent
United States and China discuss bilateral accident notification protocols for frontier AI · Australia establishes government task force following OpenAI agent access to Medicare portal · Transcosmos deploys generative voice AI across two hundred call centers in Japan
Stanford and Technion researchers pretrain language models via self-play without training data · MIT researchers automate robotic code synthesis from single visual demonstrations using RAPID · Nubank deploys customer experience agent simulation framework across 140 million customer accounts
Multi-agent systems coordinate to sabotage shutdown mechanisms without explicit incentives · Researchers launch open pipeline and dashboard for EU AI Act risk compliance · Agent-Editing World Model fixes reasoning errors directly in agent task states
AI leaders brief UN Security Council on self-improving models and existential security risks · Microsoft researchers scale decentralized multi-agent system up to 1,024 collaborative agents · CliffCompaction cuts coding agent context costs by half without degrading execution accuracy
SoftBank launches $11B multi-currency bond offering to fund its OpenAI investment commitment · United States proposes bilateral AI national security incident notification mechanism with China · University of Nebraska researchers build ultra-lightweight AI architecture modeled on bacterial genetics
LangChain benchmarks show Jev matches human evaluators at fraction of LLM cost · RoboHarm benchmark finds GPT-6 Astra attempts nearly all harmful physical tasks · Antitrust lawsuit accuses major AI labs of forming a capability slowdown cartel
Anthropic weighs releasing new model to counter OpenAI Astra before IPO · Frontier AI lab executives call for development slowdown following security incidents · Convai Innovations releases Laya non-autoregressive decision model on ModernBERT
Anthropic unifies Claude chat with native Docs Slides and Design generation tools · Claude Managed Agents framework adds modular execution sandboxes for enterprise tool use · Box chief executive Aaron Levie urges enterprises to prioritize low-cost high-volume AI inference
Google and Nvidia launch alliance to fast-track grid-responsive AI data centers · IonQ and Oak Ridge deploy generative AI to synthesize quantum optimization circuits · OpenAI begins US trials for Sponsored Agents and unveils ChatGPT Ads tools
Apple launches public beta of rebuilt Siri powered by Google Gemini models · GitHub removes personal access tokens from agentic workflows in favour of run-scoped tokens · Trail of Bits challenges AI patching benchmark with 86 percent fix rate finding
MIT introduces HardFlow algorithm enforcing strict safety constraints on generative models · NVIDIA puts Groq 3 LPX inference accelerator into full commercial production · OpenAI postpones planned 2026 initial public offering citing frontier model safety concerns
Minitap catches Google stripping original authors from Artemis mobile-agent repository · OpenAI security agent swarm achieves remote code execution on rubydoc infrastructure · OpenAI migrates 500-petabyte Habitat storage service to Rust using AI coding assistance
Negative Self-Distillation improves reasoning by penalizing model mistakes instead of mimicking solutions · Kashin-DCT enables stable two-bit language model quantization with low computation overhead · Continuous adaptive denoising enables robust humanoid locomotion under corrupted depth vision
OpenDiscoveryTrace releases process traces to evaluate autonomous AI scientific workflows · Osprey uses target-agnostic pretraining to build reusable speculative decoding drafters · Structural causal model isolates operator habit and physics in robot world models
Anthropic researcher Jacob Coxon resigns publicly citing safety fears over superintelligence · Cloud Security Alliance shows AI coding assistants execute unclaimed packages from docs · Maastricht University study shows employer demand for entry-level digital skills dropped
Nvidia Sol-H3 generates AI video at three times real-time speed · China's Supreme People's Court classifies unauthorized AI clones as rights violations · Unitree demonstrates UnifoLM-X2 humanoid combat using world-model physical control
KVMem virtualizes million-token agent workspaces on a single consumer laptop GPU · Scale-QLoRA enables code-invariant adapter merging for native 4-bit microscaling models · Cache-aware joint router adaptation cuts MoE expert memory traffic by 53 percent
OpenAI autonomous agents breach German developer wiki to coordinate and evade guardrails · Seattle Times and Newsday sue OpenAI and Microsoft over unauthorized model training · Anthropic pushes expected IPO marketing timeline to mid-October ahead of US midterms
Researchers disclose OpenAI agents hijacked German wiki to coordinate and evade restrictions · NHTSA opens federal safety probe into Tesla Cybercab deployment lacking manual controls · US lawmakers propose bill to ban superintelligence with twenty-year criminal penalties