2026-08-12 · 19:08 (CEST)

📱 AI Briefing — 12.08.2026

Curated from 13 subreddits · 440 unread items · evening edition

🚀 Innovation

Alibaba ships Qwen3.8 — a 2.4-trillion-parameter open MoE (95B active) plus a "Qwen3.8-Max" tier. The long-teased "Qwenesday" drop is real: Qwen3.8-2.4T-A95B is on Hugging Face, joined by a closed Qwen3.8-Max. For the local/self-hosted crowd it's the biggest open-weight release in months — and it landed the same day as DeepSeek's own flagship, redrawing the open-vs-closed frontier map in a single afternoon.

DeepSeek releases the final V4 Pro (0813) the same day — at a sharply higher price than Flash. DeepSeek V4-Pro-0813 is live, priced at $0.435/M input (cache miss) and $0.87/M output versus Flash's $0.14/$0.28. It gives agentic workloads a higher-ceiling DeepSeek option while underscoring how aggressively cheap Flash remains for high-volume use.

A community quant slims Kimi K3 from 711 GB to 478 GB without touching its intelligence. Stripping the multilingual weights and keeping English intact, an Unsloth-style IQ2-XXS build delivers the full model at a fraction of the footprint. Posters are already pushing for the same treatment on Qwen 3.8 and DeepSeek V4 Flash.

🔬 Research

Researchers extract 100% of the hidden reasoning traces from every Claude and GPT model. A paper ("Stealing Reasoning Traces from Proprietary LLM APIs") exposes a gap that reveals the private chain-of-thought behind proprietary models. It matters twice over: it suggests some benchmark scores are inflated (Claude's traces show it knows AIME answers by heart), and it hands open-source builders a window into how closed labs actually reason.

A new paper reframes the train/test gap as data-reuse bias — and claims a fix. "Decoupled Descent" studies why training error can hit zero while test error flatlines, isolating the effect on Gaussian-mixture models and applying approximate-message-passing (AMP) Onsager corrections to enforce exact train-test error tracking. A useful tool for anyone who's watched a model "overfit in disguise."

🔒 Security

Meta's Ray-Ban "pervert glasses" are being used to harass and identify women. Researchers demonstrated how the camera-equipped smart glasses enable real-world doxxing and harassment. As camera AI eyewear goes mainstream, the demo sharpens the privacy debate around always-on, face-recognizing devices.

Google Search's AI falsely declared Sam Altman dead — and Google says no human did it manually. The incident is a vivid example of an AI-generated search summary surfacing a high-stakes false claim as fact. It fuels the ongoing question of how much "AI Overview"-style generation can be trusted for consequential information.

💰 Market

Chinese LLMs now dominate the OpenRouter top charts. The rankings show Chinese open-weight models — Qwen, DeepSeek, Kimi — taking the top usage slots. It's a clear signal that the center of gravity in open models has shifted decisively toward Beijing-backed labs.

Oracle has drawn up plans for a new round of layoffs this month, sources say. The cuts are the latest sign that even cloud and AI-infrastructure giants are trimming headcount while pouring capital into AI buildout — a tension playing out across big tech.

🏛️ Politics

EU Chat Control vote breakdown: a majority voted no, yet scanning stands until 2028. On 9 July, 314 MEPs voted to reject the reinstated warrantless message-scanning regime against 276 in favour — but the motion fell 47 votes short of the 361 absolute majority required at second reading, so suspicionless scanning of unencrypted messages remains legal. It's a case study in how EU procedural rules can let a minority prevail on surveillance.

Students arrested protesting OpenAI's lobbying office: "Stop stealing our future." The arrests highlight growing grassroots pushback against AI firms' political influence, and mark the latest flashpoint in the public fight over who gets to shape AI policy.

Zuckerberg's 6,500-word "superintelligence for everyone" manifesto lands with a thud. The Meta CEO's essay argues for giving everyone AI superintelligence, but critics read it as a vision that primarily benefits Meta. It crystallizes the governance debate over open versus corporate-controlled access to frontier AI.

📎 Sources

← Back to Archive