📰 AI News Daily — 11 Feb 2026

TL;DR (Top 5 Highlights)

OpenAI pilots ads inside ChatGPT, signaling a power shift in digital marketing and new revenue levers for conversational AI.
Isomorphic Labs unveils IsoDDE, claiming >2× AlphaFold 3 accuracy and faster pocket discovery—potentially compressing drug development timelines.
Nebius buys Tavily for $275M, fusing agentic search with its AI cloud to accelerate autonomous agent development.
Cadence launches an AI “super” agent for chip design, automating code, test, and debug with reported 10× productivity gains.
India mandates labels for AI‑generated content on social media, raising the bar for transparency and deepfake mitigation.

Google Gemini Skills: A library to extend Gemini API/SDK with reusable capabilities. Simplifies tool introduction and discovery, helping teams ship richer agent workflows without bespoke glue code.
devagents‑cli (WebAssembly): Now runs fully in‑browser, removing local installs and easing enterprise adoption. Ideal for locked‑down environments and quick prototyping of agent automations.
NextPlaid: Production‑ready multi‑vector database bundling an ONNX late‑interaction engine. Improves retrieval precision and latency, enabling higher‑quality RAG in real‑world apps.
OpenAI Responses API + Agent Skills: Lets developers package workflows, scripts, and assets the model can discover and execute. Moves agents beyond chat toward reliable, file‑aware automation.
Samaya AI Agent Control Plane: Real‑time, expert‑grade financial analysis across complex workflows. Already serving >10,000 users at a major bank, pointing to faster, auditable decisions.
Medable Agentic AI: An autonomous assistant for clinical trial review and approvals. Reduces investigator workload and backlog, accelerating evidence generation and oversight.

GLM 5: Reported parameter gains and sparse attention for long context. Targets lower latency and stronger reasoning, improving large‑scale enterprise use and complex document understanding.
Qwen 3.5 & MoE hybrids: Hybrid SSM–Transformer MoE designs fuel strong open‑model adoption, with GLM‑4.7‑Flash‑GGUF leading downloads—evidence open models are catching up on cost/perf.
Kimi K‑2.5: Sets striking inference records (very low TTFT, high TPS) and launches a multimodal API. Signals practical, high‑throughput deployment for content and research assistants.
OpenAI Codex 5.3: Scores 90% on Next.js tasks and improves usability. Real‑world coding productivity rises as assistants better handle frameworks, scaffolding, and end‑to‑end integration.
Claude Opus 4.6: Mixed reports versus earlier variants on some benchmarks, but many users see narrower gaps to OpenAI. Highlights benchmark volatility versus lived developer experience.
Qwen‑Image‑2.0: Adds text‑to‑slides and 2K image generation. Speeds creative and marketing workflows from storyboard to production, reducing handoffs and tool fragmentation.

Stanford–Harvard math benchmark: Unpublished, proof‑requiring problems test genuine research reasoning. A harder yardstick that pressures models beyond pattern matching and exam‑style tasks.
Scaling law rethinks: Studies derive neural scaling exponents from language statistics and challenge Chinchilla‑style token‑to‑parameter ratios, suggesting task‑dependent data‑compute trade‑offs.
EPFL Stable Video Infinity: Open‑source model for seamless, extended video generation (ICLR 2026). Eases long‑form, consistent storytelling—key for ads, entertainment, and simulation.
Yale cell‑interaction AI: Real‑time insights into cellular “conversations” could unlock targets for complex diseases like cancer. Bridges single‑cell data and actionable biology.
Masked hypertension detection: ML flags hidden high blood pressure missed in clinics, enabling earlier intervention and reduced cardiovascular risk at population scale.
Lancet Digital Health study: AI systems can be misled by realistic medical misinformation. Underscores the need for guardrails before broad clinical deployment.

OpenAI pilots ads in ChatGPT (Free and Go) with partners like Fever. Clear labeling and privacy promises aim to woo brand budgets and challenge incumbent ad platforms.
Nebius acquires Tavily for $275M, integrating agentic, real‑time search with its AI cloud. Lowers friction for autonomous agents, strengthening enterprise retrieval and orchestration.
Google AI Max blends search with in‑app ads; meanwhile, Google warns AI threatens its core ad model—driving record AI infra spend and new security considerations.
India mandates labels for AI‑generated content on social platforms. A strong transparency move to curb deepfakes, likely influencing regional policy harmonization.
Enterprise agent security: Over 80% of Fortune 500 now use AI agents. Shadow AI risks grow; Cisco expands oversight to manage dependencies and AI supply chains.
Nvidia stock surges as AI infrastructure demand climbs and investors pivot toward hardware. Reinforces compute as the primary value capture layer in the AI stack.

Claude + NSA Ghidra: Agents analyze binaries without source to find backdoors—demonstrating practical, high‑stakes security workflows beyond code‑assist demos.
Kimi K‑2.5 on Mac Studio: MLX Distributed runs massive inference on Apple Silicon clusters, showcasing accessible, on‑prem alternatives to cloud for heavy workloads.
Seedance/SeeDance 2.0 & Kling 3.0: Hyper‑realistic video impresses creators, pressuring leaders like Sora and Veo. Community challenges catalyze rapid, open benchmarking.
MOVA multimodal: Tightly synchronized video‑audio generation hints at richer assistants and media tools with higher temporal coherence and narrative control.
RTX 4090 vs DGX Spark: Consumer GPUs outpace enterprise boxes on key inference/fine‑tuning tasks, underscoring that “good enough” hardware can deliver outsized, practical throughput.

Data centers in space? Critics argue training—not inference—drives co‑location needs; data gravity and bandwidth make terrestrial proximity more valuable than orbital novelty.
Labor shift, not collapse: AI favors skilled tool users, yet targeted layoffs (e.g., legal ops) show uneven disruption. Upskilling and redeployment remain crucial strategies.
Diffusion world models often hallucinate on complex tasks; progress likely needs multi‑step reasoning, self‑verification, and grounded planning beyond raw scale.
Scaling laws are not one‑size‑fits‑all: Deviations from Chinchilla suggest dynamic tokens‑per‑parameter targets, with task‑specific data curation increasingly decisive.
Single‑agent coding plateaus; dynamic multi‑agent teams with automated role generation show promise—if inter‑agent communication and coordination become first‑class design goals.
Safety as “industrial accident”: Anthropic frames failures as complex system incoherence, advocating layered oversight and continual learning over purely adversarial misalignment fixes.

Curated from 250+ RSS feeds, Twitter expert lists, Reddit, and Hacker News.