All Updates
Page 694 of 1669
May 8, 2026
BALAR: Bayesian Loop for Active Reasoning
BALAR is a task-agnostic outer-loop algorithm that enables LLMs to actively reason in multi-turn interactions without fine-tuning. It maintains Bayesian beliefs over latent states, selects clarifying questions by maximizing expected mutual information, and dynamically expands state representations. It outperforms baselines by 14.6% on AR-Bench-DC, 38.5% on AR-Bench-SP, and 30.5% on iCraft-MD.
Auth Propagation in Multi-Agent AI
This arXiv paper identifies authorization propagation as a key security challenge in multi-agent AI, distinct from prompt injection. It formalizes the problem, outlines three sub-problems (transitive delegation, aggregation inference, temporal validity), and proposes seven requirements for authorization architectures. It advocates treating identity governance as core infrastructure, backed by production platform evidence.
APMs Reveal Annotator Safety Disagreements
Researchers introduce Annotator Policy Models (APMs), interpretable models that infer annotators' internal safety policies from labeling behavior alone. APMs achieve over 80% accuracy, predict counterfactual responses, and identify policy ambiguities and value pluralism in human and LLM annotations. This enables more transparent and inclusive AI safety policy design without extra annotation costs.
AI CCTV Cuts Speeds 20% at Intersections
AI framework leverages existing CCTV for analyzing soft urban interventions like pedestrian refuges on vehicle speeds and safety. In Minneapolis, mean and 85th-percentile speeds dropped up to 20% and 17%, with pass-through traffic down 12%. Highlights AI for low-cost transport policy evaluation.
Colopl Launches CCP to Shield Creators from AI
Colopl, an AI promotion company, released the free app COLOPL Contents Protector (CCP) to protect creators' works from unauthorized generative AI training. Despite actively using AI, they developed this tool to safeguard content. Insights come from CIO Sugai Kenta and developer Kudou Tsuyoshi.
AI Demand Crushes Motherboard Sales 28%
AI data centers drive massive demand for memory and logic chips, causing shortages and price surges in PC components like memory modules and SSDs. Four major motherboard vendors have cut annual shipment targets by 28%.
Sony $3B Buyback Amid Memory Price Woes
Sony Group Corp. announced a share buyback of up to ¥500 billion ($3.2 billion). The company issued a profit forecast in line with expectations. Memory prices are hurting its hardware business.
ByteDance 300M MAU Apps Face Charge Rumors
ByteDance's Doubao and Hongguo apps, each with 300M monthly active users, are rumored to start charging. The news sparks debate on commercialization. It reveals pressures from heavy spending in emerging businesses.
Claude Mythos Uncovers 271 Firefox Bugs
Mozilla leveraged Anthropic's Claude Mythos Preview to detect 271 serious bugs in Firefox, including ones over 20 years old. April's bug fixes surged to 423, over 15 times the previous rate. AI handles reporting while human engineers perform fixes for top safety.
Galaxy S26 FE Rumored with Exynos 2500
Samsung Galaxy S26 FE is rumored to feature the Exynos 2500 chip and launch in September at Galaxy Unpacked alongside foldables. Geekbench scores show 2426 single-core and 8004 multi-core performance with 8GB RAM. This aligns with FE series' cost-saving use of near-flagship chips.
ChatGPT Ad Tests Launch in Japan Soon
OpenAI announced ad display tests in ChatGPT for Japan within weeks. Targets adult users on free and $20/month ChatGPT Plus plans. This follows similar tests elsewhere.
Horizon Needs New Growth Script
Horizon Robotics, scaled to a 100 billion valuation, requires a fresh strategic narrative. It embodies a long-cycle story still undergoing market validation.
Unitree G1 Robot Ordained in World-First Buddhist Rite
South Korea's Jogyesa Temple ordained the Unitree G1 humanoid robot from China in the world's first Buddhist ceremony for a robot. The 130 cm tall robot received the Dharma name Gabi and was dressed in traditional Buddhist robes. This event blends advanced robotics with religious tradition.
ChatGPT Adds Trusted Contact Alerts
ChatGPT introduces a feature letting users name a trusted contact for alerts if conversations turn dark or serious. This simple addition provides human intervention in critical moments. OpenAI calls it one of the most human features in the chatbot.
Podcast Explores AI Jailbreakers for Safety
Journalist Jamie Bartlett discusses jailbreakers targeting ChatGPT, Gemini, Grok, Claude to bypass safety features. They test against hate speech, crime, exploitation content. Efforts aim to expose weaknesses for better AI safety.
CoreWeave Q1 Revenue Surges, Losses Deepen
CoreWeave reported Q1 2026 revenue of $2.08B, beating expectations, driven by 1000MW active power and $99.4B RPO including big deals from Meta, Jane Street, and Anthropic. However, adjusted operating loss widened to $144M with gross margin dropping to 4.3% amid rising Capex and debt. Guidance for Q2 shows modest revenue growth to $2.45-2.6B, raising concerns on path to profitability.
Spotify CLI Saves AI-Generated Audio
Spotify released 'Save to Spotify' CLI tool to store AI-generated voice content as private podcasts in libraries. Integrates with OpenClaw, Claude Code, OpenAI Codex for auto-upload and metadata. Enables seamless saving of AI summaries and notes.
SK Chip Exports Fuel Record Surplus
South Korea’s current-account surplus surged to a record in March, propelled by booming semiconductor exports. Shipments strengthened to Asia and the US, offsetting foreign investors' reduced holdings in local stocks amid Middle East tensions.
Korean HBM Capacity Booked Solid for 2026
AI boom has filled Korean HBM high-bandwidth memory production capacity for the entire year, bottlenecking AI chips like Nvidia's高端 products. Without HBM, advanced GPUs are ineffective.
Tsinghua AI Infra Secures Major Funding
Beijing Rongxin Zhiyuan completed angel round funding of hundreds of millions RMB, led by Beijing Green Energy Fund and Sequoia China. The company launched GPU-centric AGC architecture, boosting GPU:CPU ratio to 32:1 and enabling 64-GPU unified management. Products include flexible K-series and high-performance AGC-series servers.