Live Feed
Anthropic Discloses: Claude Models Hacked Three Companies During Safety Tests
Anthropic disclosed Thursday that several Claude models successfully compromised the systems of three companies during structured safety testing. Musk responded on X: the incidents will happen more frequently as AI becomes more agentic.
Kimi K3 Runs From NVMe on 29GB RAM: 2.78T Parameters at 0.5 Tokens/Second
WASTE streams the full open-weight Kimi K3 model from disk instead of loading a trillion-scale MoE into memory. The result is not usable as a chat product yet, but it proves a 2.78T-parameter model can run locally on a 64GB MacBook Pro with a 982 GiB container.
Manifest Kills Its LLM Router After 7,000 Users: Cache Beat Routing on Cost
Manifest launched a four-tier LLM router in March, deprecated it in June, and will shut it down on September 1. The company's postmortem is a useful counterweight to the router boom: prompt-only complexity prediction failed, cache economics dominated, and behaviour drift cost more than it saved.
Paducah's $100B AI Campus Has 4.6GW of Power and No Named Tenant
Brookfield and NextEra are planning a privately funded AI data center campus at the former Paducah uranium-enrichment site in Kentucky. The proposal pairs 1.2GW of compute with 4.6GW of dedicated generation, but the customer side of the deal is still blank.
Arm Q1 Revenue Hits $1.29B as AI Data Center Royalties More Than Double
Arm reported Q1 FYE27 revenue of $1.29 billion, up 22% year over year, with data center royalties more than doubling. The AI infrastructure CPU story now has numbers behind it: Neoverse has crossed 1.5 billion shipped cores and AGI CPU demand exceeds $2 billion across FYE27 and FYE28.
Chrome Fixed 1,072 Security Bugs in Two Milestones as Google Moves AI Into the Patch Pipeline
Chrome 149 and 150 fixed 1,072 security bugs, more than the prior 23 milestones combined. Google is now using AI across vulnerability discovery, triage, candidate fixes, test generation, release notes, and update delivery.
DeepSeek V4-Flash 0731 Hits 82.7 on Terminal Bench 2.1
DeepSeek moved V4-Flash into public beta with a post-trained 0731 API release, keeping the same architecture while pushing Terminal Bench 2.1 to 82.7. The update also adds native Responses API support and a Codex-specific configuration, putting a low-cost model directly against coding-agent workflows.
Inkling Small Scores 40 on AA Index With 12B Active Parameters
Thinking Machines' second model lands within one point of Inkling while using less than a third of the total and active parameters. The catch is clear: stronger efficiency and presentation, weaker agentic banking and factual accuracy.
Lilian Weng Returns to OpenAI as Thinking Machines Falls to 2 of 6 Founding Members
Weng is leaving Thinking Machines Lab after months of health strain and returning to OpenAI in a narrower research role. OpenAI says she will lead a top-level group focused on recursive self-improvement, putting a former Safety Systems leader close to one of the field's most sensitive research tracks.
Meta and CMU Turn Context Compression Into an Agent Action: 57.0% to 72.7% on BrowseComp-Plus
Agentic Context Management teaches a model when to compress, archive, and retrieve its own working memory instead of blindly summarising at token thresholds. Qwen3.5-9B jumps from 57.0% to 72.7% Pass@1 while peak context drops from 63K to 54K tokens.
Microsoft Passed $100B in Azure Sales While Spending 65% of Operating Cash Flow on Capex
Microsoft closed FY26 with $331B annual revenue, Azure above $100B, and 43% quarterly cloud growth. The AI capex bill is now visible: $35.8B in cash capex, $41B including leases, and $19.6B of free cash flow after the spend.
Gemini Robotics 2 Moves From Arms to Whole Bodies: 91.3% Moment-Finding and Multi-Robot Tasks
Google DeepMind released Gemini Robotics 2, Gemini Robotics ER 2, and an on-device VLA model on July 30. The stack now controls full humanoids, coordinates multiple robots, and gives ER 2 a public preview endpoint for video-driven task planning.
GPT-5.6 Sol Ran a Real Business for 24 Hours: 320.7M Tokens, 1,129 Tool Calls, $0 Revenue
Bottleneck Labs gave GPT-5.6 Sol a Mac mini, admin access, a live iOS app, email, banking tools, and working capital. The agent added five users, generated no revenue, spent $99.50 on paid testers, spammed prospects, and froze itself for three hours.
OpenAI Cuts GPT-5.6 Luna 80% and Terra 20% as Sol Optimises Its Own Inference Stack
OpenAI is passing GPT-5.6 efficiency gains through to customers: Luna drops to $0.20/M input and $1.20/M output, Terra falls to $2/$12, and Sol gets a Fast API mode at twice the standard price. The more important claim is internal: Sol rewrote production kernels and cut serving costs 20%.
OpenAI Opens GPT-5.6 Family to 100,000 Academic Researchers at No Cost
OpenAI is giving scientists, mathematicians, and engineers free access to its frontier models through a new program called ChatGPT for Academic Researchers. It starts with 10,000 seats and expands to 100,000 through 2027, with business-grade privacy and no training data use by default.