Tsinghua and CUHK open-sourced AKV: agents drop selected KV cache entries and reposition the rest, so retained context is reused with no re-prefill. vs summarization: 27.7% lower cost, +1.7 points on 3 models, 4 agent benchmarks. Apache 2.0 on SGLang. github.com/thu-nics/AKV

Google open-sourced its Elephant-Goldfish Model plugin: a context-rich Elephant session drafts a design doc, then stateless Goldfish agents attack it before any code is written. Apache 2.0; Google says thousands of engineers use it daily. github.com/google/elephant-g…

Nanbeige4.3-33B is out: Apache 2.0, 33B agentic coding model, 512K context, teacher for smaller distilled models. Team-reported: ahead of Qwen3.8-27B on Terminal-Bench 2.1 (80.9 vs 73.0) and DeepSWE (65.5 vs 42.5), behind on SWE-Bench Pro (60.3 vs 61.7). huggingface.co/Nanbeige/Nanb…

Google DeepMind trained a model to argue well about moral dilemmas, judged by the model itself. It became more decisive and stopped flipping when options were reordered, but sycophancy and reframing moved it. The verdict: consistency is not validity. arxiv.org/abs/2610.30349

A dermatologist pitted GPT-6 Astra and another frontier model against a 2015 Inception v3 on 200 biopsied lesions. The CNN, a 20-minute fine-tune, got the right answer 62% of the time, the LLMs 42%. At its default threshold it cleared 18 of 30 melanomas. lynchlab.net/llm-skin-cancer

TokenGate learns to skim: a small gate plus LoRA on a pretrained LLM keeps only top-K tokens past one layer; the rest skip upper layers and their KV cache. Authors report 1.4-11.3x faster first token, 1.5-3.4x lower peak prefill memory, no accuracy loss. arxiv.org/abs/2610.26754

Anthropic's Applied AI team says enterprise agents can often drop their router and specialist subagents for one loop that loads skills and tools on demand: one system prompt, four or five thin data tools, firm guardrails. One customer's single loop ran roughly 2x faster.

NVIDIA open-sourced Flora, the data engine behind GR00T: a captioning, segmentation and reward model that label robot video together. It labels 200,000 hours a day on 512 GB300 GPUs, at $0.72 per source-video hour versus $88.49 for GPT-6 Astra at similar quality.

Cognition open-sourced the remediation workflow it used for the PHP and Node.js RCE advisories as an agent plugin: find affected repos, ask before changing anything, patch to a verified fixed version, test before and after, one PR per repo. MIT, works in Devin and other agents.

The Chancellor's AI package puts £165m into a National AI Biosecurity Initiative: £100m for an AI Biosecurity Lab, backed by up to £200m in model credits from Anthropic and another frontier lab, plus £65m for bio surveillance. Another £120m goes to the AI Security Institute.

EU member states adopted their Digital Omnibus mandate. Article 88bis lets AI development and operation rely on legitimate interest under GDPR Art 6(1)(f), but the Council text says AI alone is not a legitimate interest, to be assessed case by case, with data minimisation.

Cerebras runs a 3-trillion-parameter model across wafer-scale silicon: Kimi K3's 69 linear and 24 dense layers take 59 WSE-3T wafers, one dense layer plus its KV cache each, two linear layers each. Weights stay put; only activations hop between wafers, 2 microseconds.

Reservoir Media sued Anthropic over song lyrics: 853 works, Joni Mitchell and Johnny Cash among them. It asks up to $150,000 per willfully infringed work, or $127,950,000, plus $23.5M in DMCA claims for scraping past paywalls and stripping credit metadata.

Apollo's chief economist looked for AI in the US macro data and found it only in tech: revenue per worker, profit margins and productivity show no lift outside the sector. Spending is lopsided too: the top 10% of firms account for 99.5% of model serving and inference spend.

All nine frontier and near-frontier models in a new study use more inanimate subjects with animate verbs and more low-frequency n-grams than human writing does, and most also use longer dependencies. The authors blame training pressure for math, code and reasoning.

Dream Labs' Dream-1 is a robot policy sold as an API: point it at a bimanual arm with parallel grippers and it works, with no fine-tuning and no data from that robot. It zero-shots to 27 unseen robots in sim and 4 real ones by predicting end-effector poses, not joints. $3/hour.

Pennsylvania's AI Deceptive Advertising Act is law: ads using AI-generated images, video or audio that misrepresent a product need a label saying so, in the same medium, on first display. Text-only ads, parody and evidently fictional content are exempt. Fines top out at $3,000.

LiteLLM's Oct 21 release was poisoned: a master key left in a public repo since May gave an attacker owner-level GitHub access, CI secrets and a swapped deployer. Seven litellm==1.107.0 wheels sat on PyPI for about two hours, carrying credential theft and remote code execution.

A Convovo study asked GPT-6.1 Sol and Grok 4.7 who to vote for in 479 Senate, House and governor races. GPT named the Democrat in 1,082 of 1,111 partisan answers, Grok the Republican in 952 of 1,038. Every GPT answer with no web search declined: 99 of 99.

iFLYTEK's developer platform went agent-native at its 1024 festival: Agent OS schedules agents, staff and business systems, APA turns a screen recording into a reusable skill wired to ERP/CRM, and one CLI command reaches 700+ atomic AI capabilities. It claims 11.87M developers.