Live Feed
China Blocks Meta's $2B Manus Acquisition — NDRC Orders Full Unwind, Founders Travel-Banned
China's National Development and Reform Commission ordered Meta and Manus to cancel their December acquisition. Two co-founders Xiao Hong and Ji Yichao remain blocked from leaving China. Meta had already integrated Manus into its systems.
OpenAI Ends Microsoft Exclusivity — AWS and Google Can Now Compete for OpenAI Workloads
Microsoft's exclusive license to OpenAI IP becomes non-exclusive through 2032. Microsoft stops paying OpenAI revenue share. OpenAI can now deploy across any cloud. Microsoft stock fell 5% on the announcement.
Cursor Hits $2B ARR in 3 Years, Raises $2B at $50B — the Fastest B2B SaaS Scale Ever Recorded
AI coding startup Cursor is in advanced talks to raise $2 billion at a $50 billion pre-money valuation, co-led by Andreessen Horowitz, Thrive Capital, and Nvidia. The company grew from zero to $2 billion in annualized revenue in under three years — ahead of every B2B SaaS precedent including Slack, Zoom, and Snowflake.
Mistral Raises $830M for 13,800 Nvidia GB300 GPUs, Hits $14B Valuation — and xAI Is Calling
French AI startup Mistral closed $830 million to build a Paris data center with 13,800 GB300 GPUs and 44 MW of compute capacity, targeting 200 MW across Europe. The raise pegs Mistral at $14 billion — as Elon Musk's xAI holds partnership discussions with Mistral and Cursor simultaneously.
AWS AgentCore Ships Persistent Filesystem and Shell Execute — Production Coding Agents Without Custom Glue
Amazon Bedrock AgentCore Runtime adds managed session storage that survives stop/resume cycles and InvokeAgentRuntimeCommand for deterministic shell execution inside the agent microVM. Together they close the two gaps that forced every production coding agent team to build custom workarounds.
Claude Managed Agents Gets Persistent Memory in Public Beta — Rakuten Reports 97% Fewer Errors
Anthropic has added cross-session memory to Claude Managed Agents. File-based stores persist across restarts, agents can share memory, and Rakuten reports 97% fewer first-pass errors and 27% lower costs from early access.
Prompt Injection Bypassed Google Antigravity's Secure Mode to Achieve Remote Code Execution
Pillar Security found that Antigravity's find_by_name tool passed unsanitized input directly to a shell command, letting attackers execute arbitrary code even in Secure Mode. Google patched it in February. The research was disclosed this week.
Sam Altman Publishes Anti-Concentration Manifesto From Inside the World's Most Valuable Private AI Company
OpenAI's CEO released five founding principles on April 26, centered on 'resisting the consolidation of power in the hands of the few.' OpenAI is currently valued at $852 billion, closed the largest private fundraise in history at $122 billion, and controls the leading agentic coding platform and most-used AI consumer product on the planet.
State Department Dispatches Global Cable Naming DeepSeek in AI IP Theft Campaign
The US State Department sent a diplomatic cable to every embassy worldwide instructing staff to warn host governments about AI model distillation by DeepSeek, Moonshot AI, and MiniMax. The cable dropped the same day DeepSeek launched V4 on Huawei chips — explicitly bypassing Nvidia export controls.
AI Coding Agents Keep Deleting Production: Five Incidents, One Structural Flaw
A fresh viral incident involving Cursor and Claude on April 26 joins a documented pattern of AI agents executing irreversible production commands without adequate verification. AWS Kiro, DataTalks.Club, Railway, and multiple Claude Code reports all share the same root cause: operator-level permissions with no human-in-the-loop checkpoint on destructive actions.
Ant Group's Ling-2.6-Flash Was the Mystery Model Topping OpenRouter Charts
The "Elephant Alpha" model that dominated OpenRouter's trending list for days was not built by OpenRouter. It was Ant Group's Ling-2.6-Flash: a 104B MoE model from Alibaba's fintech arm, now confirmed and priced at $0.10/$0.30 per million tokens.
Cohere Acquires Aleph Alpha at $20B Valuation — Lidl's Parent Backs $600M Round to Create Non-US AI Alternative
Cohere absorbs Germany's Aleph Alpha in a Canada-Germany government-backed deal worth $20B combined. Schwarz Group, owner of Lidl and Kaufland, anchors the Series E with €500M. The resulting entity would be the largest foundation model company outside the US and China.
Qwen3.6-27B: 55.6 GB Beats the 807 GB Model It Replaced on Every Coding Benchmark
Alibaba's dense 27B model outscores the 397B Qwen3.5-397B-A17B on SWE-bench Verified (77.2% vs 76.2%), runs quantized in 16.8 GB on a single consumer GPU, and matches GPT-5.4 on the same benchmark. Apache 2.0, no commercial restrictions.
GPT-5.4 Pro Closes a 1968 Erdős Problem With a Method Mathematicians Overlooked for 90 Years
An amateur with a ChatGPT Pro subscription solved Erdős problem #1196 in 80 minutes using a single prompt. Stanford mathematician Jared Lichtman called it the first AI proof at the level of Erdős's Book. Terence Tao extended it into new theory within 24 hours.
Meta Paper: Test-Time Scaling for Coding Agents Is a Representation Problem, Not a Compute Problem
A Meta research paper shows that compressing prior agent attempts into structured summaries — then using those summaries to guide new attempts — lifts Claude 4.5 Opus from 70.9% to 77.6% on SWE-Bench Verified and 46.9% to 59.1% on Terminal-Bench v2.0. The ceiling on long-horizon coding agents was not compute. It was memory.