Turing winner Yann LeCun launches AMI to prove that world‑understanding AI can beat today’s chatbot‑style language models.

Anthropic’s clash with the Pentagon shows how hard it is to keep AI out of autonomous weapons and mass surveillance on today’s battlefields.

Alibaba’s OpenSandbox offers a unified, secure API so AI agents can run code, browse and train in isolated Docker or Kubernetes sandboxes.

Alibaba's Qwen3.5-9B tops OpenAI's gpt-oss-120B on GPQA, Video-MME benchmarks. Runs on laptops, multimodal edge AI star.

QuitGPT frames quitting ChatGPT as a political and psychological detox while pointing users toward alternative AI tools.

GPT 5.4 improves computer-use benchmarks, adds million-token context and tight Excel and Sheets finance integrations.

Nano Banana 2 is Google’s new default image model, balancing realism, speed, and 4K output for mainstream users.

Perplexity Computer competes with Operator and OpenClaw by orchestrating 19 models into one browser based, multi step workflow engine.

Chinese labs like DeepSeek allegedly stole Claude smarts with 24k accounts. Anthropic's past piracy suits add irony to claims.

Anthropic and the Pentagon clash over AI limits, weapons use, and mass surveillance as Trump orders federal agencies to drop Claude.

Qwen3.5-27B/35B open-source, runs local on 18-24GB VRAM. Benchmarks rival Claude Sonnet 4.5, 1M ctx, tool-native.

DeepSeek V4 eyes early March drop, topping Claude/GPT in code benchmarks with 1M context, but Blackwell chip scandal brews US ire.

KiloClaw hosts OpenClaw agents in 60s: secure VMs, 500+ models, cron jobs. Beats self-hosting hassles, akin to Kimi Claw but no forks.

From ICE to the NHS, Palantir’s Gotham and Foundry systems quietly drive decisions worldwide, raising deep concerns about surveillance power.

A 400M param Kani TTS 2 model brings real time friendly speech, local accents and voice cloning to consumer GPUs under a 3 GB VRAM budget.

Kimi Claw brings a browser native AI agent with 5,000 skills and 40 GB storage, turning Kimi into an automation hub for research, coding.

Seedance 2.0 shows China’s AI video can rival OpenAI and Google, mixing realism with serious ethical and legal questions.

Alibaba’s Qwen 3.5 boosts multimodal and agent skills, narrowing the gap with US AI leaders while staying efficient through a sparse MoE design.

MiniMax M2.5 is a fast, MoE‑based LLM tuned for coding, agents and office tasks, offering frontier‑level output at notably lower cost

Zhipu’s GLM 5 scales to 744B parameters with 40B active, pushing open weight coding and agent performance toward closed “Opus class” levels.