2026-08-13 · 07:07 (CEST)

📱 AI Briefing — 13.08.2026

🚀 Innovation

Qwen3.8-2.4T-A95B released — Alibaba's Qwen team shipped Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter open-weight MoE model with 95B active parameters, immediately sending the LocalLLaMA community into planning mode for running it locally. It's flagship-scale open weights closing the gap with closed frontier systems — and a reminder that open models have reached sizes unthinkable on consumer hardware a year ago.

Meta's Muse Glimmer 30B targets "always-on" agents — Meta released Muse Glimmer 30B, an open-weight model purpose-built for 24/7 autonomous agents, and the community already has it running up to ~3.3× faster on Apple Silicon via speculative decoding (mlx-dspark). Output stays byte-identical to normal decoding, so it's a practical speedup with no quality trade-off for Mac-based agent workloads.

DeepSeek V4 Flash 0731 runs locally on CPU — The new DeepSeek V4 Flash 0731 can be served on a single machine with 128GB RAM plus a 4090/P40 combo, bringing a frontier-class model within reach of budget local setups. The Unsloth 4-bit build fits in ~144GB, letting hobbyists approximate a "GLM 5.2-class" experience without a datacenter.

🔬 Research

Ablating one attention head kills Morphy's queen sacrifice — A chessformer-lens interpretability demo shows that removing a single one of 128 attention heads makes a chess transformer stop discovering a famous queen sacrifice. It's a striking, concrete illustration of how localized and brittle a specific model capability can be.

"The Loss Does Not See the Basis, But Adam Does" — New research argues Adam's per-coordinate second moment breaks the rotational invariance of the loss in factored models (W=UVᵀ), destroying gradient descent's implicit low-rank bias. Across nine update rules, GD, shared-scalar Adam, Muon and Shampoo kept the bias while Adam, RMSProp, Lion and Adafactor lost it — a practical lever for choosing optimizers in low-rank training.

🔒 Security

Nvidia Jetson Orin found inside a new Russian cruise missile — Ukraine's defence intelligence identified an Nvidia Jetson Orin AI module in the S-71 "Monochrome" cruise missile, a sign Russia is folding smuggled Western AI hardware into weapons for machine-vision terminal guidance. It's the latest evidence that sanctions evasion keeps feeding dual-use chips into Russian munitions, sharpening pressure to close export-control loopholes.

Twitch feeds creator content to Amazon AI and hides the opt-out — Twitch is now training Amazon AI models on streamer content and buried the opt-out option, with its chief product officer admitting "if it was opt-in, nobody would opt in." The quiet default-in consent shift shows how platform ToS changes are becoming the front line of AI training-data disputes.

Farmer kills 25 acres of crops on AI's advice — A farmer destroyed 25 acres of his own crop after following an AI tool's recommendation, a stark real-world example of the harm unverified AI guidance can cause when acted on without human checks. It's the kind of incident fuelling the debate over AI accountability in high-stakes decisions.

💰 Market

Nvidia doubles RTX PRO 6000 Blackwell MSRP to $16,000 — The 96GB RTX PRO 6000 Blackwell — which opened pre-orders below $8,000 last year — now carries a $16,000 list price, nearly double. For local-AI builders, the surge on the flagship 96GB workstation card signals that AI demand is inflating GPU pricing across the board, and raises fears DGX Sparks could follow.

74% of enterprises run AI in production — half can't prove it pays off — A large majority of enterprises now run AI in production, but roughly half can't demonstrate a return on the investment. The widening gap between deployment and provable ROI is becoming the central tension in enterprise AI as CFOs push back on spending without measurable payoff.

🏛️ Politics

EU AI Act forces Anthropic to watermark Claude globally — Under the EU AI Act's Article 50 transparency rules (effective 2 August), Anthropic is adding imperceptible watermarks and signed provenance metadata to all Claude output — and applying it worldwide, a textbook "Brussels effect" where EU compliance becomes the global default. Non-compliance carries fines up to €15M or 3% of global turnover, and users are already complaining the marks will expose AI use at work and school.

AI agents aren't legally responsible for harm — so who is? — Legal experts say autonomous AI agents can't themselves be held liable for damage they cause, leaving a murky gap over whether the developer, deployer, or user answers. As agents gain the ability to take real-world actions, the unresolved liability question is becoming a barrier to trust and a rising priority for regulators.

📎 Sources

← Back to Archive