AI News
Daily AI industry news — funding, products, policy, and major moves from global AI companies and startups.
Why GLM-5.2's 10x Cost Reduction Signals an AI Pricing Collapse
Uber exhausted its annual AI budget in just four months. GLM-5.2 delivers superior coding performance at 10% of GPT 5.5's cost. AI infrastru
Why SongRyeon Core Splits LLM Judgment and Absolute Facts
SongRyeon Core is a local agent runtime that separates LLM judgment from facts. The system categorizes data into absolute, relative, and mix
Unconventional AI Aims to Slash Inference Power by 1,000x
Unconventional AI aims to reduce inference power consumption by 1,000x. The Un-0 simulation matches the performance of Stable Diffusion. Nav
Anthropic Opus Defends Fiu AI Against 6,000 Prompt Injection Attacks
Anthropic Opus powered an AI assistant that blocked 6,000 attacks. The experiment resulted in zero leaks of the secrets.env file. High instr
Claude's 75% Paid User Surge Signals a Crack in ChatGPT's Monopoly
Anthropic's Claude saw a 75% increase in paid users since January 2026. Course demand for Claude on DataCamp now exceeds that of ChatGPT. Ch
NVIDIA Rubin Cuts Water Use to Zero With 100% Liquid Cooling
NVIDIA Rubin implements 100% liquid cooling to remove all server fans. The closed-loop system saves 2.6 million gallons of water per MW annu
General Intuition Raises $320M to Solve the Robot Data Bottleneck
General Intuition raised $320 million to train robots using game data. The company uses action labels to teach AI spatial and temporal reaso
How isgen.ai Identifies the Mechanical Patterns of AI Content
The isgen.ai tool helps creators identify predictable patterns in AI text. It distinguishes between AI-generated style and traditional plagi
WorkOS Launches auth.md to Standardize AI Agent Registration
WorkOS introduced auth.md as an open protocol for AI agent registration. The standard uses OAuth and ID-JAG to automate account creation. Ma
The Cognitive Debt Trap in the Age of LLM-Generated Code
LLM dependency creates cognitive debt by replacing understanding with replication. True productivity shifts from measurable output to long-t
LFM2.5-230M Outperforms Models 4x Its Size in Data Extraction
Liquid AI released LFM2.5-230M for high-efficiency data extraction. The model beats larger competitors on BFCLv3 and CaseReportBench. A hybr
Anthropic Exposes Alibaba's 28.8 Million Query Attack on Claude
Anthropic reported that Alibaba used 25,000 fake accounts to extract Claude's data. Chinese labs used model distillation to bypass massive c
Why the Trump Administration is Mandating Customer Approval for GPT 5.6
The Trump administration has ordered a restricted release of OpenAI's GPT 5.6. Access to the model now requires individual government approv
AI Political Bias Map Reveals 4 of 6 Models Lean Left Economically
A new bias map shows four of six major AI models lean left economically. Researchers used a neutral classifier and Condition A to isolate in
OpenKnowledge Launches Local-First LLM Wiki for Markdown
OpenKnowledge released a local-first Markdown editor and LLM wiki. The tool supports integrations with Claude and Codex LLM harnesses. It is
Patronus AI Secures $50 Million to Build Digital Worlds for Agent Testing
Patronus AI raised 50 million dollars to verify AI agent reliability. The company uses digital world models to simulate real enterprise envi
The 52.5% Hallucination Drop Powering OpenAI's GPT-5.5 Instant Update
OpenAI released GPT-5.5 Instant to improve complex constraint handling. The model reduces hallucinations by 52.5 percent in high-risk prompt
Netris Raises $15M from a16z to Automate 1 Million GPU Clusters
Netris secured $15 million in Series A funding led by Andreessen Horowitz. The platform currently optimizes networking for one million GPUs
Adobe's Topaz Labs Acquisition Shifts AI Video Upscaling to the Local GPU
Adobe is acquiring Topaz Labs to integrate high-end AI enhancement into Creative Cloud. The deal focuses on moving heavy video upscaling fro
Micron Hits $1.2 Trillion Valuation as AI Memory Shortages Peak
Micron reached a market valuation of 1.2 trillion dollars. The company's third quarter profit surged to 28.2 billion dollars. AI memory shor
SignalFire Data Shows AI-Era Engineer Hiring Now Hits 55%
SignalFire reports engineer hiring share rose to 55 percent. AI efficiency is increasing demand via the Jevons Paradox. Software roles are s
The 19% Revenue Risk Driving ASML's Fight Against the MATCH Act
The Netherlands is lobbying the US to block the MATCH Act's equipment bans. ASML faces a potential 19% revenue loss if China exports are hal
Hang Ten Systems Launches With $32M to Automate Enterprise Software Delivery
Vishal Sikka founded Hang Ten Systems to replace linear IT labor costs with AI agents. The startup secured $32 million in seed funding led b
Harness-Level Loops: The AI Pattern Trading Code Quality for Speed
Harness-level loops use external systems to force AI agents into repetition. This pattern accelerates porting but degrades long-term code ar
RubyLLM Unifies GPT, Claude, and Ollama Into a Single Ruby Interface
RubyLLM provides a unified interface for over 12 major AI providers. The framework integrates deeply with Rails via ActiveRecord and UI gene
Code as Agent Harness: Redefining AI Agents as Operational Substrates
UIUC, Meta, and Stanford researchers introduced the Code as Agent Harness framework. The paper argues that agent failure stems from infrastr
Why Yongho Ha Redefines AI Expertise as an Operations Lead
Yongho Ha identifies a Verification Tax that slows AI productivity. AI-native firms must shift from production to verification layers. Profe
Why Facebook is Shipping a Standalone AI Companion for Creators
Facebook launched a standalone AI Companion app for creators. The tool integrates content planning and analytics into one interface. Meta ai
The 0% Exposure Rate That Redefines AI Search Optimization
A developer found that new services have a 0% exposure rate in AI search. The study tested ChatGPT, Claude, Gemini, and Perplexity with 100
VibeThinker-3B Compressed Flagship Reasoning Into 3B Parameters
VibeThinker-3B achieves flagship-level reasoning with only 3 billion parameters. The model uses CLR and Spectrum-to-Signal training to optim