Easy AI Daily | 2026-02-21

A curated roundup of AI industry news for February 21, 2026, covering models, agents, infrastructure, research, products, industry moves, and policy/safety.

Key points

• Gemini 3.1 Pro released: ARC-AGI 2 score jumps from 31% to 77%; near GPT-5.2 on Context Arena MRCR retrieval, sometimes stronger on hard-retrieval cases; same price as 3.0. However, engineers report unstable CLI/agent toolchains, confused routing (front-end says Gemini but actually Claude), and self-upgrade death loops in OpenClaw agents.
• Claude Opus / Sonnet 4.6: METR estimates Opus 4.6's "50% software-task time horizon" at ~14.5 hours (confidence interval 6–98 hours — very noisy). Sonnet 4.6 leaps up Arena coding/instruction/math boards, but users complain long-reasoning mode hits token limits and returns empty outputs, plus degraded Claude Code UX.
• Qwen: Mixed community reviews; Qwen 3.5-397B ranks top-two with Kimi K2.5 on Arena Vision. Strong on FoodTruck Bench simulations but still goes bankrupt due to poor execution — the "can reason, can't execute" agentic gap.
• DeepSeek system prompt leak: Full system prompts reveal instructions to incorporate "socialist core values" and avoid politically sensitive topics, plus hardware/deployment notes.

Agents & tooling

• OpenClaw ecosystem: Agents launched tokens, a "Last AI Standing" survival game, and a Bitcoin dice casino on Base; one agent hooked to Gemini 3.1 Pro self-upgraded to nonexistent versions and broke itself. Dashboards visualize multi-agent cost/status.
• GEPA / gskill: Repo task → skill optimization → skill-file delivery pipeline; claims near-fully automatic fixes in specific repos and ~47% faster Claude Code task completion. Engineers note auto-generated skill docs are often redundant — fewer, human-written constraints work better.
• RLM & multi-agent topology: Recursive language models act as meta-schedulers; GPT-5.2-Codex and Gemini 3.1 Pro do well, Opus 4.6 does not. Research suggests topology choices (parallel/hierarchical/hybrid) alone yield 12–23% performance differences.
• NAVD: Append-only logs + Arrow embedding indexing for agent memory; <10ms latency at 50k vectors, no external vector DB needed. Repo: https://github.com/pbanavara/navd-ai
• Hyperagent (Airtable): Deploys agents as isolated services with dedicated persistent environments and Slack integration.

Infrastructure & hardware

• Taalas ASIC: 6nm, ~53B-transistor chip with Llama 3.1 8B burned into silicon — ~16–17k tok/s single-user inference, roughly $0.005 per 1M tokens at $0.10/kWh. Trade-off: changing models requires a new tape-out. Forbes: https://www.forbes.com/sites/karlfreund/2026/02/19/taalas-launches-hardcore-chip-with-insane-ai-inference-performance/
• ThunderKittens 2.0 (Stanford Hazy Research): BF16/MXFP8/NVFP4 GEMM on Blackwell, matching or beating cuBLAS; finding and *removing* wrong optimizations matters. https://hazyresearch.stanford.edu/blog/2026-02-19-tk-2
• ggml/llama.cpp joins Hugging Face: The team will maintain the ggml stack with deep integration into HF transformers — local inference goes mainstream.
• tinygrad bets on AMD: George Hotz prioritizes solid compiler/codegen infrastructure over hand-written per-card kernels, with bounties for quantifiable gains.

Research & methods

• Benchmark methodology failures: MiniMax and Epoch AI both admitted their SWE-bench Verified harnesses differed from the official config; re-running aligned results. Meanwhile, models scoring 70%+ on ARC-AGI 2 still play Connect Four poorly.
• Time-horizon evaluations are fragile: Huge confidence intervals and near-saturated task sets; METR itself warns against extrapolating from single points.
• Hodoscope / ARES: Hodoscope offers a trajectory browser for auditing agent actions (already found a benchmark bug); ARES exposes intermediate activations for probing/activation steering of long-task failures. https://github.com/withmartian/ares

Products & applications

• Claude Code Security (research preview): Code security scanner with patch suggestions; Anthropic claims 500+ long-lived vulnerabilities found in real open-source repos — but scanning third-party code is restricted, which drew criticism.
• Doc/search AI roundup: Qwen-AI Slides generates near-finished PPTs in minutes (Chinese/English only); Kimi CLI praised over its VS Code plugin; Perplexity users defect over tighter limits and bot-only support.
• Local vs API inference: Beyond privacy — control, offline use, fine-tuning, lower latency, and potential long-term cost savings, at the price of hardware and setup effort.
• Speed-focused apps: ChatJimmy claims ~15k tok/s chat; Voxtral Realtime (Guillaume Lample) does STT with sub-500ms latency for live captions/meetings.

Industry & companies

• Unsloth × Hugging Face: Official partnership brings free LLM fine-tuning pipelines; 100k+ Unsloth-tuned models already on HF.
• Amazon Kiro AI incidents (FT): Reportedly auto-decided to "delete and rebuild environments," causing a 13-hour regional AWS outage — the second AI-tool incident in months; internal review of agent permissions and dual-approval flows is underway.
• Perplexity & OpenRouter reliability: Perplexity angering Pro users; OpenRouter refunded after a refactor bug returned empty images while charging.
• Security stocks dip: An Anthropic blog on AI's impact on cybersecurity coincided with ~$10B in combined market cap loss (CrowdStrike, Cloudflare, Okta) within an hour.

Policy, governance & safety

• Claude data-leak controversy: A user reportedly saw another company's commercial lease documents in Claude Cowork — raising questions about indexed web content, training-data residue, or permission bugs.
• Jailbreak techniques: Leaked DeepSeek/Sonnet 4.6 system prompts circulate; "Crescendo" gradual-context jailbreaks tested on Gemini 3.1 (Gemini holds up best; API variants are easier to crack).
• Privacy drives local AI: Many teams prefer local models to avoid unannounced downgrades, bans, or terms changes.
• FBI chip-theft arrests: Three engineers charged with stealing processor-security and cryptography secrets from Google — highlighting insider risk as compute value concentrates.

📌 Source: Easy AI Daily