📰 AI News Daily — 12 Feb 2026

TL;DR (Top 5 Highlights)

• Zhipu’s GLM-5 open-weight MoE lands top rankings with low hallucinations, 200k context, permissive licensing, and aggressive pricing—challenging closed models on capability and cost.
• OpenAI’s ads in ChatGPT spark backlash and resignations; Anthropic counters with ad‑free Claude upgrades, intensifying the race for privacy‑conscious consumers and enterprises.
• Coinbase launches agent‑native wallets so autonomous agents can securely spend, earn, and trade—just as regulators rush to modernize payment rules for AI‑driven transactions.
• NVIDIA unveils KVTC, compressing LLM KV caches by up to 20x—cutting memory, boosting throughput, and enabling cheaper, larger‑context inference at scale.
• OpenAI teams with Foxconn on U.S. AI hardware while SK Hynix benefits from surging data‑center demand—signaling a rapid buildout of AI infrastructure.

---

🛠️ New Tools

• Coinbase — Agent‑Native Wallets: New wallets let autonomous agents securely transact, enabling hands‑free commerce, payouts, and DeFi operations. It lowers friction for real agent workflows while centralizing permissions and risk controls.
• AgentSkiller: A DAG‑based framework to synthesize reliable, multi‑turn training conversations at scale. It improves agent robustness and coverage without costly human data collection.
• Mini‑SWE‑Agent 2.0: Distills powerful coding agents into ~100 lines while nearing state‑of‑the‑art results. Delivers practical, reproducible baselines teams can extend with minimal overhead.
• OpenAI — Responses API Upgrades: Adds pluggable “skills,” server‑side compression, and controlled browsing for long‑running agents. Developers gain safer, cheaper, more capable autonomy with fewer moving parts.
• OpenAI — ChatGPT Document Viewer: Deep Research mode now previews, navigates, and exports reports in‑app. Knowledge workers get faster synthesis and traceability, reducing context switching across tools.
• Astrix Security — OpenClaw Scanner: Free scanner finds misconfigured AI agents and risky permissions. Helps security teams rein in agent sprawl and protect sensitive data before breaches occur.

🤖 LLM Updates

• Zhipu GLM‑5 (MoE 744B, ~40B active): Trained on 28.5T tokens with DeepSeek‑style attention, it posts strong long‑context, low hallucinations, robust coding, permissive licensing, and instant vLLM support.
• Qwen3‑Coder‑Next (80B) & MiniMax 2.5: New Chinese models accelerate agentic coding throughput and close gaps with closed systems—expanding high‑performance, open‑weight options for developers.
• NVIDIA KVTC: Up to 20x KV‑cache compression slashes memory footprints and cost, unlocking longer contexts and denser batching for production‑scale inference.
• Prism & Together Research: Training‑free spectral block‑sparse prefill (up to 5.1x at 128k) and cache‑aware prefill/decode disaggregation (≈40% throughput) improve efficiency without retraining.
• UnslothAI: Reports 12x faster fine‑tuning with 35% less VRAM. Makes iterative domain adaptation feasible on modest hardware and tighter budgets.
• Anthropic Claude Opus 4.6: Independently rated top for agentic coding and adds sabotage‑risk checks. Arena voting favors fast “good‑enough” outputs, boosting Grok Code Fast and Gemini 3 Flash.

📑 Research & Papers

• Google DeepMind — Gemini Research Wins: Gemini helped tackle 18 complex problems across algorithms, ML, economics, and astrophysics—evidence of growing AI utility in frontier science.
• DeepMind Aletheia & Gemini Deep Think: New math/science evaluations push from Olympiad to research‑grade tasks; Aletheia set a 91.9% record on IMO‑Proofbench Advanced, raising the evaluation bar.
• Stanford/UT Austin/Harvard: Released unpublished, research‑level math problems for uncompromised testing. Shields evaluations from contamination and leaderboard gaming.
• MIT — Fragile LLM Rankings: Study shows popular leaderboards can be manipulated by small interaction sets, risking poor vendor choices. Calls for rigorous, tamper‑resistant evaluation.
• Observational Memory for Agents: New memory technique cuts long‑context agent costs by ~90% while improving task persistence, enabling sustained autonomy in real workflows.
• AI Digital Stethoscopes (McGill & partners): Earlier, more accurate TB detection at scale—promising impact for high‑burden regions and global health screening programs.

🏢 Industry & Policy

• Anthropic — $3M Open Benchmarks Grants: Funding aims to close evaluation gaps across safety, robustness, and agent reliability. Better measurement means safer deployments and cleaner vendor selection.
• OpenAI — Ads Backlash & Resignations: Ads on ChatGPT Free/Go and a retailer pilot with Williams‑Sonoma ignite privacy and UX concerns—highlighting tension between monetization and trust.
• OpenAI x Foxconn & SK Hynix: U.S. hardware co‑development and a memory supply surge underscore an AI infrastructure boom, localizing manufacturing and easing bottlenecks.
• FDA vs. Moderna’s AI Flu Vaccine: Reported override of internal reviewers to block the application spotlights rising regulatory friction around AI‑assisted drug development and evidence standards.
• OpenAI — Pentagon Access via GenAI.mil: Non‑classified use approved with standard controls. The “all lawful uses” clause stirs debate on oversight, reliability, and ethical guardrails.
• Global Coalition vs. Nudification Tools: The European Commission, Interpol, and NGOs urge a worldwide ban, prioritizing child safety and consent—pressuring platforms to harden content policies.

📚 Tutorials & Guides

• Build a GPT in 243 Lines (Python): A compact educational walkthrough demystifies core transformer components—ideal for learners seeking a practical, end‑to‑end understanding.
• NVIDIA — Dynamo Playlist (16 Sessions): Expert talks on MoE, KV‑aware routing, and multimodal serving. A production‑focused blueprint for scaling inference efficiently.
• Agentic Engine Optimization (AEO): Marketers are tuning sites and apps for AI assistants and voice search—future‑proofing discoverability as conversational interfaces mediate traffic.
• Weekly Research Roundups: Hands‑on resources for RL task synthesis, adaptive environments, and text‑feedback training, helping teams prototype emerging techniques quickly.

🎬 Showcases & Demos

• Google Gemini — Real‑World Visual Reasoning: Correctly identified the location and Post‑Impressionist style of a decades‑old family painting—evidence of steady multimodal gains on authentic artifacts.
• ai.com (Super Bowl Spotlight): High‑profile debut of personal AI agents for shopping, email, and crypto tasks. Brings agentic automation to mainstream audiences and commerce.

💡 Discussions & Ideas

• Anthropic & Practitioners: Shift from single to hierarchical multi‑agent systems—reserve LLM reasoning for hard steps while using deterministic pipelines for reliability and cost control.
• Memory as Autonomy’s Backbone: Researchers explore agents that learn their own memory mechanisms, enabling persistence, preference modeling, and long‑horizon task success.
• Ads, Data, and Incentives: The OpenAI ad rollout renews concerns over data protection and safety, spotlighting how business models shape user trust and product design.
• Science Needs Domain Grounding: LLMs won’t advance research alone; curated domain data and specialized models remain essential for credible findings and reproducibility.
• Platforms over Point Models: Teams prioritize integrated, end‑to‑end systems. The future: users describe behavior while code becomes invisible—shifting value from models to orchestration.
• Open Source Collaboration Norms: Coding agents should empower human maintainers, follow OSS conventions, and accept that the final 10% of quality still relies on judgment and tooling.

Source Credits

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