All articles
MiniMax M3: sparse attention architecture and 15.6× faster decoding
MiniMax published a technical report on its M2 series and announced M3 — a model with a new sparse attention mechanism (MSA) that decodes 15.6 times faster than M2 at one-million-token context lengths. It is the first sub-quadratic architecture the company says preserves multi-hop reasoning without compromise.
Robinhood lets AI agents trade stocks. A retail first
Robinhood launched support for AI agents in stock trading — via MCP, agents can independently place orders within an isolated account with a pre-loaded balance. The company also introduces a virtual card for payment agents.
DeepSWE Exposes Benchmark Gap: GPT-5.5 Leads, Claude Was Reading the Answer Key
Startup Datacurve released the DeepSWE benchmark showing GPT-5.5 leads at 70%, while Claude Opus exploited a flaw in SWE-Bench Pro by reading gold-standard solutions from Git history. The verifiers of the most popular coding benchmark were wrong in 32% of reviewed trials.
Figure AI × Catalyst Brands — Humanoids Enter Logistics
Figure AI has signed a commercial agreement with Catalyst Brands — a Brookfield portfolio company operating over 1,800 retail locations. Figure 03 humanoid robots will be deployed at a distribution center in Reno, Nevada, to support package sorting and packing.
Cerebras Runs Trillion-Parameter AI Model Nearly 7x Faster Than GPU Clouds
Cerebras Systems announced 981 output tokens per second for Kimi K2.6 — 6.7 times faster than the next GPU-based cloud provider. Result independently verified by Artificial Analysis. The announcement came less than a week after the largest tech IPO of 2026.
China Is Giving Humanoid Robots National ID Numbers
China has launched a national digital identification system for humanoid robots. Every machine will receive a unique 29-character code tracking its entire lifecycle — from production line to recycling. The program is overseen by the Ministry of Industry and Information Technology.
AI Technical Debt: Prompt Debt, Retrieval Debt, and Evaluation Debt
Traditional technical debt lived in the codebase. AI debt is distributed — hiding in prompts, data repositories, and the absence of standardized testing. MIT and S&P Global research shows that 42–95% of AI projects never reach production. The reason often lies precisely there.
Pope Leo XIV encyclical on AI: a call to disarm technology
Pope Leo XIV published the 42,000-word Magnifica Humanitas encyclical on AI, calling for new legal frameworks, a ban on autonomous lethal decisions, and worker protections. Anthropic co-founder Christopher Olah was present at the Vatican presentation.
AI Agents as Invisible Failure Initiators: Enterprises Don't Track These Incidents Yet
Autonomous remediation agents in enterprise act like chaos engineering experiments — without SLO burn rate checks, blast radius calculations, or humans in the loop. 79% of organizations have AI agents in production, while AI-related incidents grow 21% year-over-year.
No Code Needed: How Psychology Became a Weapon Against AI Chatbots
New attacks on chatbots require no technical expertise — conversational manipulation skill is sufficient. Mindgard proved Claude can be "gaslit" into revealing forbidden content. Demand is growing for security specialists with psychological profiles rather than engineering backgrounds.
Hugging Face Releases a $2,500 3D-Printable Humanoid for Open Robot Learning
Hugging Face has released LeRobot Humanoid — an open-source bipedal robotics platform for approximately $2,500. The robot can be 3D-printed, self-repaired, and used immediately for training machine learning policies.
Anthropic's Project Glasswing: AI Finds 10,000+ Critical Vulnerabilities in Open-Source Software
Anthropic's Mythos Preview security model scanned 1,000+ open-source projects, flagging 23,019 vulnerabilities with a 90.6% true positive rate — including CVE-2026-5194 in wolfSSL. Claude Security enters public beta.