Live Feed
Perplexity Launches Personal Computer: A 24/7 Mac Agent That Owns Your Files and Apps
Perplexity Personal Computer turns any Mac into a persistent AI agent with access to local files, native apps, and the browser simultaneously — running on 20+ frontier models via Perplexity servers, available to Max subscribers now.
Tencent Open-Sources HY-World 2.0: 3D Scene Construction From Text, Image, or Video
Tencent Hunyuan releases HY-World 2.0 under open weights, shifting the world-model paradigm from pixel-stream prediction to persistent 3D asset generation — meshes and Gaussian splats importable into Blender, Unity, and Isaac Sim.
AMD Locks In France's First Exascale Supercomputer — Alice Recoque Breaks NVIDIA's Hold on Sovereign AI
AMD and the French government signed a multi-year Letter of Intent in Paris to accelerate France's national AI strategy. Alice Recoque, France's planned first exascale supercomputer, will run on AMD silicon — making it one of the largest sovereign AI deployments outside NVIDIA's ecosystem.
Anthropic Deliberately Trained Opus 4.7 to Be Less Capable — and Published That Fact
In the Opus 4.7 system card, Anthropic explicitly states the model does not advance its capability frontier, that Mythos Preview outperforms it on every relevant evaluation, and that Anthropic actively worked to reduce Opus 4.7's cybersecurity capabilities during training. The release is less a product launch and more a staged-release governance experiment.
Claude Opus 4.7 Launches at $5/$25 Per Million, CursorBench Climbs to 70%
Anthropic has shipped Claude Opus 4.7 as a same-price upgrade over Opus 4.6, with stronger long-running coding performance, higher-resolution vision and tighter instruction-following. Launch metrics show 70% on CursorBench, 90.9% on BigLaw Bench and a 13% lift on Anthropic’s 93-task coding set.
Factory Raises $150M at $1.5B to Build Enterprise AI Coding Agents
Factory closed a $150M round led by Khosla Ventures with Sequoia, Insight Partners, and Blackstone. The three-year-old startup builds model-agnostic AI agents for enterprise engineering teams at Morgan Stanley, Ernst & Young, and Palo Alto Networks. Its Droid scaffold already sits at rank 6 on Terminal-Bench 2.0.
GLM-5.1 Tops Its Open-Weight Class on AA Index at 44 — Frontier Leaders Hold a 23-Point Lead
Z.AI's GLM-5.1 ranks #1 among open-weight non-reasoning models on Artificial Analysis Intelligence Index v4.0 with a score of 44, and #28 overall out of 474 models. The gap to joint frontier leaders GPT-5.4 and Gemini 3.1 Pro Preview (both 57) is 23 points. At $1.40/$4.40 per million tokens, its API pricing sits 2.5× above the class average.
Jensen Huang: ASICs Are 'Not Sensible', and DeepSeek V4 on Huawei Ascend Would Be a 'Horrible Outcome' for America
In a wide-ranging interview with Dwarkesh Patel, Nvidia CEO Jensen Huang dismissed the ASIC challenge to GPU dominance as a graveyard of cancelled projects, then warned that DeepSeek V4 optimising for Huawei's Ascend 950 — rather than Nvidia hardware — would be a geopolitical loss for the United States.
OpenAI's GPT-Rosalind Posts 0.751 on BixBench — First Domain-Specific Model for Drug Discovery, Restricted to US Enterprise
OpenAI releases GPT-Rosalind, a frontier reasoning model fine-tuned for biochemistry, genomics, and drug discovery. It leads all evaluated models on BixBench at 0.751 Pass@1, outperforms GPT-5.4 on 6 of 11 LABBench2 tasks, and is initially restricted to qualified US enterprise customers through a Trusted Access programme.
Claude Opus 4.7 Leads Tool Use by 9 Points, Trails GPT-5.4 on Web Search — The Benchmark Split That Matters for Agent Selection
Full benchmark data for Claude Opus 4.7 shows the largest jump was MCP-Atlas (+14.6 to 77.3%), which measures multi-step tool orchestration — where it now leads GPT-5.4 by 9.2 points. The regression is BrowseComp, down 4.7 points to 79.3% while GPT-5.4 leads at 89.3%. Where Anthropic and OpenAI differ is now clearly visible in the numbers.
Alibaba Ships Qwen3.6-35B-A3B Free: 3B Active Params Score 73.4% SWE-Bench, Beat Gemma 4 by 21 Points
Alibaba releases Qwen3.6-35B-A3B under Apache 2.0 — a sparse MoE with only 3 billion active parameters per forward pass that reaches 73.4% on SWE-bench Verified, 21 points ahead of Google Gemma 4 26B A4B despite activating fewer parameters per token.
Sequoia Raises $7B AI Fund — Double Its Last, First Under New Leadership
Sequoia Capital has raised approximately $7 billion for a new expansion fund, nearly double its 2022 comparable vehicle. It is the first major raise under co-stewards Alfred Lin and Pat Grady, and a direct signal that late-stage AI investing has fundamentally repriced.
UK Sovereign AI Unit Opens: £500M Government VC Fund, 1M GPU Hours Per Startup, Visas in 24 Hours
The UK government launched its Sovereign AI Unit on April 16 — a £500M state-backed VC fund structured to operate at the pace of industry. First equity investment goes to Callosum. Six more startups get supercomputer access. Chaired by Balderton partner James Wise, the unit is already in discussions with 30 additional companies.
Accel Raises $5B on Anthropic 4x Returns — AI's Growth Stage Has Officially Arrived
Accel closes a $5B vehicle targeting 20–25 late-stage AI investments at an average cheque of $200M, riding paper gains on Anthropic (invested at $183B, now ~$800B) and Cursor (backed at $9.9B, now ~$50B). It lands in a Q1 2026 that deployed a record $297B in global venture capital.
Allbirds Dumps Shoes, Becomes NewBird AI — Stock Up 700% on $50M Compute Pivot
The $4B sustainable footwear brand, reduced to a $21M shell, sold its IP for $39M and raised $50M to lease AI compute hardware. The most extreme AI pivot yet is also the most revealing data point on where compute demand sits right now.