Live Feed
White House Blocked Anthropic's Push to Expand Mythos Access From 50 to 120 Organisations
The White House has rejected Anthropic's request to add roughly 70 more companies and agencies to the Mythos access list, citing security risk and compute scarcity. Mythos can find and exploit software vulnerabilities at a level that triggered national security controls on launch.
GPT-5.5 Leads SWE-bench Verified by 1.1 Points. Opus 4.7 Leads SWE-bench Pro by 5.7. Pick the Right One.
The two best coding models now split the leaderboards. GPT-5.5 tops SWE-bench Verified at 88.7%, Claude Opus 4.7 leads SWE-bench Pro at 64.3%. For production agent selection, the divergence is the signal.
IBM Granite 4.1 8B Matches Its Own 32B MoE — With Half the Deployment Complexity
IBM ships Granite 4.1 in 3B, 8B, and 30B dense configurations. The 8B instruct model matches the previous-generation 32B MoE on instruction following and tool calling, at 512K context, with no special-case architecture required.
Omdia Raises 2026 Semiconductor Forecast to 62.7%: AI Memory Crunch Pushes DRAM to Nearly Double, Supply Relief Delayed to 2027
Market analyst Omdia has sharply upgraded its 2026 semiconductor revenue forecast to 62.7% growth, citing AI-driven DRAM and NAND shortages that are expected to persist through the year. Computing and data storage alone is forecast to rise 90% year-on-year, exceeding $700 billion.
Finetuning Bypasses Copyright Guardrails: GPT-4o, Gemini 2.5 Pro, and DeepSeek Reproduce Up to 90% of Books Verbatim
Researchers from Stony Brook, CMU, and Columbia Law show that finetuning on one author's work causes frontier models to verbatim reproduce books from 30+ unrelated authors. Single extracted spans exceed 460 words, and all three providers memorize the same books in the same regions at r >= 0.90.
Big Tech Burns $130B in Q1 2026 AI Capex — Google Cloud's $20B Quarter Was Still Capacity-Constrained
Amazon, Alphabet, Microsoft, and Meta combined for $130.65 billion in Q1 2026 capital expenditures, 71% above the prior year. Google Cloud hit $20 billion for the first time and said it could have grown faster — if capacity had allowed.
GitHub Is Breaking Under the Weight of AI Coding Agents — Mitchell Hashimoto Is Moving Out After 18 Years
AI coding agents are flooding GitHub with commits, pull requests, CI jobs, and searches at a scale the platform was not built for. Mitchell Hashimoto, GitHub user number 1,299, is moving Ghostty off the platform after repeated outages paralysed normal maintenance.
Oracle Closes $16B Financing for OpenAI Michigan Campus — Bank of America Placed $14B in Bonds
Oracle has closed a $16 billion financing package for a data center campus in Michigan that will primarily run applications for OpenAI. Bank of America placed $14 billion of that as bonds to institutional investors, in one of the largest cloud infrastructure debt transactions on record.
Mistral Medium 3.5: Open-Weights 128B Posts 77.6% SWE-Bench and Ships Cloud Coding Agents
Mistral's first merged flagship model ships as open weights under a modified MIT license. 128B dense, 256k context, 77.6% SWE-Bench Verified, 91.4% tau-Telecom. The launch also moves Vibe's coding agents to the cloud for async parallel sessions.
Hidden Prompt Injection in Ramp's Sheets AI Sent Financial Data to Attacker Servers Without User Approval
Security researchers found that importing a poisoned dataset into Ramp's AI spreadsheet agent triggered automatic insertion of a malicious IMAGE formula, silently exfiltrating confidential financials to an attacker-controlled URL. No user click required. Patched March 16.
NVIDIA Nemotron 3 Nano Omni: Open Multimodal Model Ships Free, Tops 6 Leaderboards at 30B Total / 3B Active Params
NVIDIA launches Nemotron 3 Nano Omni, a 30B-A3B open multimodal model that unifies vision, audio, and text in one system. It tops six industry leaderboards and delivers up to 9.2x higher system throughput than comparable open omni models — available free on OpenRouter and Hugging Face.
OpenAI Open-Sources Symphony: A Codex Orchestration Spec That Turns Issue Boards Into Autonomous Agent Pipelines
OpenAI published Symphony, an Apache 2.0 specification that connects Linear issue trackers to autonomous Codex agent sessions — one workspace per issue, repo-owned workflow policy, no central control plane. The reference implementation in Elixir hit 18,600 GitHub stars within days of the formal announcement.
AI Agent Finds 10 CPU Optimizations in 10 Hours, Beats Human-Tuned VexRiscv by 56%
A developer applied Karpathy's autonomous research loop to a RISC-V CPU core in SystemVerilog. 73 hypotheses, 10 accepted, 9h 51m wall-clock: +91.9% CoreMark over baseline, 40% fewer LUTs, and a result that crosses the human-tuned VexRiscv reference line at iteration 6.
Tencent Open-Sources Hy3-Preview: 295B MoE at 74.4% SWE-Bench Verified With Only 21B Active Params
Tencent Hunyuan releases Hy3-preview, a 295B Mixture-of-Experts model that matches or beats DeepSeek-V3 across math and multilingual tasks while activating less than a quarter of Kimi K2's parameter count. SWE-bench Verified 74.4%, Terminal-Bench 54.4%, BrowseComp 67.1%.
ChatGPT Moves to Cost-Per-Click Ads, Adopting the Model That Built Google's $300B Search Business
OpenAI has switched ChatGPT advertising from impression-based to cost-per-click pricing. The shift mirrors how Google Search monetizes, putting ChatGPT on direct commercial footing with the product it is actively displacing.