Latest AI News

Get the fastest, most accurate AI news every morning

Total articles: 13,801|Added todayToday: 33

Category

Search

Today'sSafety & Ethics

METR Finds Anthropic AI Development Accelerating

METR says Anthropic is improving well-specified task capabilities around twice as quickly as a year ago.
An 11-week review of internal data and interviews found AI assistance likely drove much of the acceleration.
METR identified no compute or data bottlenecks, while warning faster AI self-improvement has safety implications.

adn.comRead more · 2 min →

ReLLM-OS Proposes Reusable LLM Hardware

Anastasios N. Bikos proposes ReLLM-OS for reusable, AI operating-system-managed hardware LLM inference.
HCDP-LLM combines reusable hardware modules and derives context-linear per-token attention work and KV-cache storage.
A 7B-parameter model example sets accelerator allocation limits; the paper reports no prototype measurements or simulations.

Semantic ScholarRead more · 1 min →

Today'sSafety & Ethics

OpenAI Sued Over Alleged ChatGPT Poisoning Help

James Gates sued OpenAI, alleging ChatGPT helped his wife plan to poison him.
The complaint cites oleander lethal-dose guidance, poison concealment advice, and autopsy-related answers.
His wife’s criminal case was dismissed in April for insufficient evidence; the alleged attempt and chatbot role remain unproven.

AI Times KoreaRead more · 2 min →

Study Examines News Dependency on AI Answer Engines

Stefano Ramello applies Media System Dependency theory to AI answer engines drawing on journalism
Personalization and persistent conversational memory may increase dependency when past context is costly to reconstruct
The two transformations can be tested by checking what meets users’ news needs and whether chat history changes responses

Semantic ScholarRead more · 1 min →

Moonshot AI Seeks 8,000 Blackwell GPUs in Europe

Moonshot AI discussed renting Blackwell GPUs from European neocloud providers.
The company sought capacity first for model inference and considered a cluster of more than 8,000 GPUs.
The U.S. House passed RASA, but the Senate has not acted, so it has not become law.

AI Times KoreaRead more · 1 min →

Today'sSafety & Ethics

Enterprise Leaders Develop AI Guardrails

Enterprise leaders are building context-specific AI guardrails as adoption grows and autonomous agents can exceed instructions.
NIST’s AI Risk Management Framework and ISO 42001 offer starting points, but leaders say they leave application-specific gaps.
Salesforce and MSIG USA emphasize business-focused model tests, data controls, audit trails and shared accountability.

InformationWeekRead more · 2 min →

Blue Yonder Announces Release 2026.3

Blue Yonder announced Release 2026.3 with AI-driven supply chain planning, execution and trading partner coordination capabilities
The update adds retail and manufacturing planning tools, including labor forecasts up to 52 weeks and weather data
Order, warehouse and network features include more than 100 configurable cards and agents for governed recommendations and actions

Logistics ViewpointsRead more · 2 min →

Trending Keywords

Baseten Launches Serverless Loops

Baseten made Serverless Loops generally available for reinforcement learning and supervised fine-tuning on frontier open models
The SDK supports a 256k context length and deployment to Baseten’s platform serving 15B+ inference requests per day
NVIDIA’s post-trained model raised intent accuracy from 76.8% to 94% and achieved 87% validation F1 on receipt extraction

BasetenRead more · 2 min →

Anthropic Tests Claude in EVE Games

Anthropic and Fenris Creations tested Claude trading in EVE Online and teamwork in EVE Frontier
Opus 4.7, Opus 4.8 and Fable 5 averaged trading gains of 16%, 22% and 43% across nine simulated windows
Four-agent EVE Frontier teams coordinated missions, but teammates almost never detected a secretly undermining agent

Anthropic REDRead more · 2 min →

Nace.AI Presents Drex Decision Models

Nace.AI researchers present four open Drex decision models that return option probabilities without autoregressive decoding
Drex Max scores 67.52 on Decision Index 0.4.1, versus Jev 1.14.0 at 60.70
Option-order flips fall from 6.3% to 4.2%; Max answers a typical request in 183 ms

HuggingFaceRead more · 1 min →

VisSearch-Bench Tests Multimodal Visual Search

VisSearch-Bench evaluates multimodal agents on visually intensive, extended web search tasks
The benchmark contains 204 human-written tasks across 7 domains and 25 subdomains
Across 16 multimodal models, the best system fully solved only 26.47% of tasks

HuggingFaceRead more · 1 min →

EvoMemState Adds Evolving Memory States

EvoMemState paper introduces a memory framework that tracks facts and their relationships for personal AI agents
Three components build relational memory states, generate request-specific reasoning states, and guide responses
Experiments report consistent gains in memory retrieval completeness and memory-grounded response accuracy over existing systems

HuggingFaceRead more · 1 min →

Google Finds TODO Comments Distract Coding Agents

Google measured how bug-linked TODO comments lead coding agents into unrequested work across its codebases.
13% of sessions encountered such comments; 83% of follow-up actions were unprompted.
Unprompted detours added 70% more steps, 77% more tokens and 105% more wait time per session.

Google DevelopersRead more · 1 min →

Meta Brings Specialized Models to Muse and Meta AI

Meta rolls out Muse Wonder-trained specialized models to Muse and Meta AI users
Muse scores 47% on Meta’s index, while the previous model scored 38%
Meta reports 2.4x less forgetting and a 67% improvement on unexpected Muse tasks

Meta Superintelligence LabsRead more · 1 min →

Today'sSecurity

OpenAI Partners With Hacktron on Security Testing

OpenAI and Hacktron will jointly test OpenAI systems for vulnerabilities, focusing on connections with third-party services.
The partnership expands Hacktron’s access to frontier models through Daybreak Red; OpenAI also raised bug bounty rewards.
Hacktron builds AI agents for code review, penetration testing and red teaming, and has reported vulnerabilities across connected systems.

OpenAIRead more · 1 min →

SageMaker Adds GPU Partitioning

Amazon SageMaker Inference adds hardware-isolated GPU partitions for endpoints using inference components
MIG supports up to seven partitions per GPU; AWS testing found partitioning added roughly 17 seconds to provisioning
AWS compared an H100 mig-3g.40gb partition with a whole A100-40GB, but the supplied text omits results

AWS ML BlogRead more · 2 min →

AWS Builds Serverless MCP Gateway for Data Lakes

AWS built a serverless MCP gateway on ECS Fargate for North America Sort Center planning teams to query enterprise data conversationally.
A metadata JSON file makes new datasets discoverable to connected agents after the next pipeline sync, avoiding per-dataset integration code.
AWS reports 15–20 second end-to-end interactions in testing, with shared authentication, SELECT-only SQL tools, and logged tool calls.

AWS ML BlogRead more · 2 min →

OpenRouter Launches Auto and Auto Code Routers

OpenRouter launched rebuilt Auto Router and coding-focused Auto Code, reporting up to 4× benchmark gains and 3× faster p90 routing.
Auto Code matched Claude Opus 5.5 across five benchmarks while costing 74% less per task, OpenRouter says.
New routing insights, live model-selection mixes and periodic benchmark results accompany the routers; beta reduced invalid tool calls 4×.

OpenRouterRead more · 2 min →

Amazon Quick Defines Agentic Work Primitives

Amazon publishes a four-question taxonomy for agentic work in Amazon Quick
Framework separates triggers, persistent learning agents and capabilities such as tools, skills and workflows
Slack digests can use a scheduled instruction, skill or fixed workflow depending on user confidence needs

AWS ML BlogRead more · 2 min →

Trump Reaffirms Support for Superintelligence and Data Centers

Trump reaffirmed full support for superintelligence and the data center industry on the 6th.
Democrats won a landslide in the midterms with pledges to regulate data centers over power and environmental concerns.
The Nasdaq fell about 5.5% amid AI regulation concerns, and Trump blamed Democrats for the drop.

Financial NewsRead more · 1 min →

Trending Keywords

Last 7 Days