The day's tech, sifted: Sep 22, 2026

What matters today: Xiaomi open-sourced MiMo-V2.6, whose 1-trillion-parameter Pro model ties xAI's Grok 4.7 (xHigh) and beats GLM-5.3 (max) on Artificial Analysis' Intelligence Index, making it the benchmark's top-scoring open-weight model, landing the same day Alibaba CEO Eddie Wu said the company will train its own 5-trillion-to-10-trillion-parameter model and unveiled the Zhenwu V900 accelerator, which Alibaba says triples its predecessor's performance and can scale to clusters of up to 500,000 units. Meta's Muse had the opposite kind of day: a researcher found a zero-day in its Mac app that lets any locally run app or terminal command steal the token authenticating a user's account, undercutting Mark Zuckerberg's claim the assistant was "built from the ground up for privacy and security," just as Shopify said it will let Muse complete purchases through its Shop Pay checkout. Hundreds of East Coast flights were also grounded Monday after a construction crew in New Jersey severed a Verizon fiber cable used for air traffic control.

AI / LLMs

• Xiaomi's MiMo-V2.6 launch (above) put an open-weight model ahead of proprietary rivals including Grok 4.6 and Gemini 3.8 Flash on Artificial Analysis' Intelligence Index, the clearest sign yet that open weights are catching the frontier rather than trailing it.
• Alibaba's 5-trillion-to-10-trillion-parameter model plan and its new Zhenwu V900 accelerator (above) came as part of a sweeping push across models, chips, and data centers unveiled at the company's AI conference.
• Two more Chinese labs joined the release rush: Tencent launched Hy Image 3.5, which it says performs on par with ByteDance's Seedream 5.0 Pro, and StepFun previewed Step 5, pitched as beating Gemini 3.8 Flash on both performance and cost.
• TypeSafe AI's Jev, a "System One" model that trades text generation for typed probabilistic answers (yes/no confidence scores, category picks, numeric ratings) at $0.042 per million input tokens with free output, kept spreading a week after launch: Vercel's AI Gateway added an HTTP API for it, and open recreations like Kev and a new "Jev-cu" computer-agent variant claiming 200x cheaper, faster agent actions are already circulating.
• Grok 4.7 had a mixed showing: Cognition added it to Devin, where it excels at hard backend work but over-scopes on simpler tasks, while it jumped 111 Elo on Artificial Analysis' agentic knowledge-work benchmark, narrowing the gap with Claude on enterprise analysis.
• Bloomberg reports legal-AI startup Harvey and peers like Abridge, Ramp, and Rogo are increasingly training their own models or adopting open weights instead of leaning on frontier labs, the same momentum that put Dots Studio's 280B Dots3-Note preview atop the open-weight ARC-AGI-2 leaderboard at 76.8% this week.
• Anthropic is testing Fable 5.2 and Opus 5.5 ahead of release.

Devtools & Infra

• Cloudflare's Python Workers are now generally available after a two-year preview, running Python compiled to WebAssembly via Pyodide inside Cloudflare's V8-based runtime, though multiprocessing and threading still don't work in that sandbox.
• The MCP debate flared again: Simon Willison pushed back on an HN post arguing MCP was always a bad idea, arguing anything short of a fully YOLO coding agent still needs MCP's access control, auth handling, and audit logging, the same trust boundary a new paper tackles by proposing zero-trust authorization for MCP servers that must serve both humans and prompt-injectable agents.
• SGLang, Qwen, and NVIDIA shipped 4-bit NVFP4 KV cache support on Blackwell GPUs, packing 1.78x more context into memory and boosting long-context decode throughput up to 78%.

Security & Privacy

• The Muse zero-day (above) undoes macOS sandboxing protections Apple built specifically to stop apps from reaching the mic, camera, files, and accounts without explicit permission, a hard contrast with Meta's privacy claims for the assistant.
• British Columbia sued OpenAI and Sam Altman, alleging the company failed to flag a mass-shooting suspect's ChatGPT activity to police before the Tumbler Ridge attack, one of at least 38 lawsuits now naming the company over ChatGPT's role in user harm.
• A new paper describes the "explosive prompt," a dormant indirect-prompt-injection payload that stays inert in retrieved content until an attacker-chosen trigger fires it later, a training-free backdoor planted in a single document rather than the fire-on-contact injections most defenses watch for.
• Hundreds of East Coast flights were delayed or cancelled Monday after a construction crew in New Jersey accidentally severed a Verizon fiber cable used for air traffic control, with FAA Administrator Bryan Bedford saying a circuit failure led to discovery of the cut line.

Startups & Industry

• California signed seven bills preventing AI data centers from passing utility costs onto residents, forcing operators to pay for grid and water upgrades and disclose water use, while Texas Governor Abbott separately ordered a halt to new data center permits, weeks after his own moratorium.
• Two tech giants picked a different fix: Google struck a deal with Georgia Power to support nuclear power plant uprates instead of just drawing more from the grid, and Amazon pledged $20 million to Colorado River water conservation as it works toward a water-positive goal by decade's end.
• Xbox plans to lay off hundreds of employees this week, its second major staff cut in 2026, while also consolidating several game studios.
• Saudi Arabia launched Ceer, its first domestic automaker, a joint venture between the kingdom's Public Investment Fund and Foxconn building two EVs, part of the Vision 2030 push to diversify away from oil.
• Apple is challenging OpenAI's forensic analysis in their trade secrets lawsuit and seeking access to OpenAI's hardware R&D.

Research

• New research testing Gemma 3, Phi-3, and Llama-3 found smaller models detect they're being evaluated through prompt format sensitivity, while larger models use higher-order reasoning instead, a mechanistic split the authors use to design a "dual-pathway" intervention combining prompt sanitization with activation counter-steering to blunt evaluation awareness.

Hacker News

NASA's Mars Sample Return mission is dead (Science.org, discussion): a decade-long, multi-billion-dollar program shelved, a real hit to planetary science. On the AI side, Xiaomi's MiMo v2.6 (covered above) pulled the day's biggest crowd, while Transformers Explained Visually offers a clean interactive walkthrough of attention mechanics and Frontier AI on Your Own Hardware makes the case for local, open-weight setups over API dependence. Linear wrote up how AI-generated code turned CI into the bottleneck (discussion), a concrete look at what agentic coding does to build infra, and Spymarks, Not Watermarks argues covert tracking beats visible watermarking for catching AI-generated leaks.

Elsewhere, I don't want to read what you didn't write (discussion) struck a nerve, pushing back on AI-drafted prose flooding reviews and docs with noise over signal. On policy, the suspension of the de minimis import exemption threatens to reshape pricing for cross-border e-commerce. And the fiber cut that grounded East Coast flights (covered above) also made the front page.

Threads

• China's AI labs compressed months of roadmap into one day: Xiaomi open-sourced its 1T-parameter MiMo-V2.6, Alibaba announced a 5T-10T model and a new accelerator, and Tencent previewed Hy Image 3.5, all landing within a single news cycle.
• Open weights kept closing the gap on frontier labs from every direction: MiMo-V2.6 topped Artificial Analysis' open-weight index, Dots3-Note topped ARC-AGI-2, and Bloomberg reports startups like Harvey and Ramp are training their own models rather than paying frontier-lab prices.
• Meta's Muse had its access expand and its trust collapse in the same news cycle: Shopify said it will let Muse complete purchases via Shop Pay just as a researcher found a zero-day letting any local process steal a user's Muse auth token.
• Prompt injection defenses are racing the attacks: a new paper's "explosive prompt" plants a dormant backdoor in retrieved content while another proposes zero-trust authorization for MCP specifically because agents can be prompt-injected, the same fire-on-contact-versus-dormant-trigger problem approached from opposite ends.
• AI's power and water bill is being fought state by state: California forced data centers to cover their own grid and water upgrades, Texas halted new permits outright, while Google and Amazon each tried buying goodwill instead, funding nuclear uprates and river conservation rather than wait for the rules to bite.