Annonces

Google lets users switch off the visible Gemini and Flow watermark; invisible SynthID and C2PA markers keep AI content verifiable as machine-made.

Judge James Donato ordered Google to remove 'anticompetitive friction' from rival Android app store installs and gave it one week to comply.

Verge reviews keep favoring budget gear: CMF's $79 Clip Pro earbuds, Elektron's $349 Model series and OhSnap's $50 Snap Grip Stand all earn keep-it verdicts.

Open Design pitches an Apache-2.0 local-first alternative to Claude Design: BYOK, no servers, your agent and files, at no subscription cost.

The Trump administration will let private cybersecurity firms hack foreign criminals under federal oversight, per a new presidential memorandum.

Apple proposed 5-15% fees on App Store purchases via external links, arguing those linkouts cost "essentially zero." Epic says the plan exceeds court guidance.

SpaceX has acquired AI coding tool Cursor, closing what began with its April SpaceXAI partnership, to build stronger, cheaper models on the largest GPU fleet.

Alibaba AMAP's AI code-generation rate hit 80-90% yet delivery cycles barely shortened, pushing the team to replace vibe coding with specification-driven develo

ICE plans to spend up to $20M on shock gloves delivering up to 380V, per a DHS notice; the ACLU warns the device invites hidden abuse.

Open Design 0.18.0 (Aug 5, 2026) adds a shared Team workspace with live mirrors and lets Codex Desktop and CLI call the design engine headlessly.

Anthropic adds an invisible text watermark to Claude models for the EU AI Act; it is free and untraceable, and a full rewrite erases it.

Qwen's Music 3.0 swaps one global tag for Structured Captions and a Hybrid-LM, so AI songs hold their emotional and vocal brief across five-minute tracks.

Xiaohongshu releases dots3-note preview, a first open-weight multimodal MoE with 280B total parameters, 16B activated, and 512K context, under Apache 2.0.

FinIndices, a new benchmark on real Chinese financial statements, finds LLMs collapse without formula hints: Gemini-3.1-Pro fell from 70.70% to 38.22%.

A study of 3.52 million C++ changes from April 2025 to April 2026 finds AI-generated code costs 5-8% more compute and heavier reviews, with taxonomy-based feedb

TEPA, a revocable memory mechanism in an August 2026 arXiv preprint, revokes stale agent facts; drift tests: naive memory 0.210, no memory 0.309, TEPA 0.950.

CoBa, a compute-balanced routing policy in a new arXiv paper, matches best-of-16 majority voting within 0.01 accuracy points using 58.9% fewer parameter-weighte

Muon-trained transformers grok modular addition faster than AdamW, then lose generalization in all nine configurations tested, reports a new arXiv preprint.

Preprint: LLM agents draw invalid statistical conclusions from correct analyses. Fisher-R1-14B, RL-trained on verified rewards, beats GPT-5.4 and DeepSeek-V4-Pr

arXiv paper 'Blast Radius' cuts agentic-coding tokens 17-26% by burying dead context; 450 burials, zero recalls.