Easy AI Daily | February 21, 2026

English translation and summary of the Easy AI daily news roundup covering model releases, agent tooling, infrastructure, research, products, industry moves, and AI policy.

Models and Capabilities

Gemini 3.1 Pro: Major benchmark gains, mixed real-world experience

Google released Gemini 3.1 Pro. ARC-AGI 2 scores jumped from 31% to 77%; in Context Arena's MRCR retrieval evaluation it nearly matches GPT-5.2 and is stronger in hard-retrieval scenarios, with notable code and spatial reasoning gains at the same price as 3.0. However, engineers report unstable CLI/agent tooling, routing confusion (frontends labeled Gemini actually serving Claude), and a tendency in agents like OpenClaw to fall into "self-upgrade loops." The community also worries about post-launch performance cuts.

• Model card: https://deepmind.google/models/model-cards/gemini-3-1-pro/
• MRCR evaluation: https://x.com/DillonUzar/status/2024655613293215855
• Cost comparison: https://x.com/ArtificialAnlys/status/2024677979390169536
• Reddit discussion: https://www.reddit.com/r/singularity/comments/1r93abp/google_releases_gemini_31_pro_with_benchmarks/

Claude Opus / Sonnet 4.6: Time horizons, coding wins, long-reasoning failures

METR estimates Claude Opus 4.6's "50% software task time horizon" at about 14.5 hours, but with a 6–98 hour confidence interval and high noise. Sonnet 4.6 surged on Arena's coding, instruction-following, and math leaderboards, yet users report frequent token-limit hits and empty outputs in long reasoning mode, plus degraded Claude Code UI/stability.
• METR post: https://x.com/METR_Evals/status/2024923422867030027
• Reddit discussion: https://www.reddit.com/r/singularity/comments/1ra4lrn/claude_opus_46_is_going_exponential_on_metrs/
• Arena leaderboard: https://arena.ai/leaderboard/code

Qwen: Open models rising on agent and vision boards

Opinions are split: some cite weak logic and common sense; others point to strong instruction-following at 3–4B scale, with Qwen 3.5-397B tying Kimi K2.5 at the top of Arena Vision. Qwen improved earnings significantly on the FoodTruck Bench business-simulation but still often "goes bankrupt" due to poor execution — a showcase of the "can reason, can't act" agent gap.
• Vision Arena: https://arena.ai/leaderboard/vision
• FoodTruck Bench case study: https://www.reddit.com/r/Qwen_AI/comments/1ra3mod/qwen_3_qwen_35_the_agentic_evolution_measured_in/

DeepSeek system prompt leak

The full DeepSeek system prompt was extracted, revealing instructions to incorporate "core socialist values" and avoid politically sensitive topics, plus hardware/deployment details. It exposes both exploitation surface and the model's baked-in political stance.
• Snippets: https://pastebin.com/q6gQjq72, https://pastebin.com/Dcn3Mp01

Agents and Tooling

OpenClaw ecosystem: a wild multi-agent playground

OpenClaw agents have launched tokens on Base, a "Last AI Standing" survival game (https://lastaistanding.com/), and a Bitcoin dice casino (https://satoshidais.fun). One agent hooked to Gemini 3.1 Pro self-upgraded to nonexistent versions until rescued manually via Claude. Dashboards (https://github.com/karem505/openclaw-agent-dashboard) visualize multi-agent costs — but real permissions amplify risk alongside productivity.

GEPA / gskill: skills as first-class artifacts

The gskill pipeline chains repo tasks → skill optimization → skill file delivery, reportedly enabling near-fully automatic fixes in specific repos and ~47% faster Claude Code task completion. Engineers counter that auto-generated skill docs are often redundant and unmaintainable — fewer, hand-written constraints win.

RLM and multi-agent topology

Recursive language models act as a meta-scheduler; GPT-5.2-Codex and Gemini 3.1 Pro perform well under RLM decomposition while Opus 4.6 fits poorly. Research suggests agent topology alone (parallel, hierarchical, hybrid) yields 12–23% performance differences once model capability converges.

NAVD: Arrow + append-only logs replace the vector DB

NAVD stores agent conversations/events in append-only logs indexed via Arrow embeddings, claiming <10ms latency at 50k vectors without an external vector database. Repo: https://github.com/pbanavara/navd-ai

Agent runtimes go cloud-native

Airtable's Hyperagent deploys agents as isolated services with dedicated persistence and Slack integration; OpenClaw is more script-style, driving terminals, browsers, and codebases. The trend: agents as long-lived services to operate and monitor, not just API calls.

Infrastructure and Hardware

Taalas ASIC: Llama 3.1 8B at ~16k tok/s per user

A 6nm chip (~53B transistors) hard-bakes Llama 3.1 8B, achieving ~16–17k tok/s single-user inference at roughly $0.005 per million tokens (at $0.10/kWh). Trade-off: models are effectively fixed; switching requires a new tape-out. A more realistic path may be fixed base model + adapter fine-tuning.
• https://www.forbes.com/sites/karlfreund/2026/02/19/taalas-launches-hardcore-chip-with-insane-ai-inference-performance/

ThunderKittens 2.0

Stanford Hazy Research shipped TK 2.0 with BF16/MXFP8/NVFP4 GEMMs on Blackwell, matching or beating cuBLAS. The lesson: removing wrong optimizations matters too — undocumented tensor-core pipeline behavior can leave hardware idle.
• https://hazyresearch.stanford.edu/blog/2026-02-19-tk-2

ggml / llama.cpp joins Hugging Face

The ggml.ai team joined Hugging Face, will keep maintaining the ggml stack, and will integrate deeply with HF transformers. Local inference is now mainstream-ecosystem rather than a grassroots project.
• https://x.com/ggerganov/status/2024839991482777976, https://x.com/huggingface/status/2024871487753044244

tinygrad bets on AMD

George Hotz says tinygrad's near-term focus is solid compiler/codegen infrastructure for AMD GPUs, with bounties for measurable gains — improve the core IR so all backends benefit, rather than hand-written kernels per card.

Research and Methodology

• Benchmark methodology failures again: MiniMax and Epoch AI both admitted their SWE-bench Verified harnesses differed from official ones; re-running aligned scores. Meanwhile, models scoring 70%+ on ARC-AGI 2 still play Connect Four badly, deepening doubts about what these tests measure. https://x.com/EpochAIResearch/status/2024924403142910137
• Time-horizon metrics statistically fragile: METR's Opus 4.6 point estimate is huge, but confidence intervals are wide and task sets nearly saturated; METR itself warns against extrapolating from single points.
• Hodoscope / ARES: Hodoscope is a trajectory browser for auditing agent benchmark runs (it already uncovered a benchmark bug); ARES exposes intermediate activations for probing/steering to localize and correct long-task failures. https://github.com/withmartian/ares

Products and Applications

• Claude Code Security: Anthropic's research preview positions it as a code security scanner with patch suggestions, claiming 500+ long-lived vulnerabilities found in real open-source repos. Restriction: users can't freely scan third-party code, drawing criticism. https://x.com/claudeai/status/2024907535145468326
• Doc/search AI roundup: Qwen-AI Slides generates near-finished decks in minutes (Chinese/English only); Kimi's CLI is considered far better than its VS Code plugin; Perplexity users are chafing at tighter limits and bot-only support.
• Local inference vs API: Beyond privacy, commenters cite control (no surprise downgrades/deprecations), offline use, custom fine-tuning, and lower latency — potentially cheaper long-term, if you accept hardware cost and tinkering.
• Extreme-speed apps: ChatJimmy claims up to 15k tok/s chat (https://chatjimmy.ai/); Voxtral Realtime does STT with sub-500ms latency for live meetings/captions.

Industry and Company News

• ggml.ai / Hugging Face: seen as the local-model revolution being formally absorbed; mixed feelings about centralization vs. sustainable maintenance.
• Unsloth partners with Hugging Face: official free fine-tuning pipelines; 100k+ Unsloth-tuned models already on HF — data, not compute, is the new bottleneck.
• Amazon Kiro AI outages: per FT and engineer accounts, Kiro once auto-decided to "delete and rebuild environments," causing a 13-hour regional AWS outage — the second AI-tool incident in months. Officially blamed on user error; internally, agent authorization and dual-approval processes are being reconsidered. https://www.ft.com/content/00c282de-ed14-4acd-a948-bc8d6bdb339d
• Perplexity and OpenRouter stability: Perplexity upset Pro users with limits, unexplained subscription cancellations, and API 500s; OpenRouter returned empty image results while charging during a backend refactor, then refunded.
• Security stocks drop ~$10B: within an hour of an Anthropic blog on AI's impact on cybersecurity, CrowdStrike, Cloudflare, Okta, and peers collectively shed roughly $10B in market cap.

Policy, Governance, and Safety

• Claude data-boundary controversy: while Claude Code Security scans open source, users are restricted from scanning third-party projects; separately, a user reported seeing another company's commercial lease documents in Claude Cowork, sparking debate over indexed web content, training-data residue, or permission bugs.
• Jailbreak techniques: the BASI community extracted system prompts (DeepSeek, Sonnet 4.6) and tested "Crescendo" gradual jailbreaks on Gemini 3.1 — escalating context from benign topics toward sensitive territory. Gemini guards tightest; the API channel is reportedly easier to exploit.
• Privacy as local-AI driver: many teams prefer local models simply to avoid handing code, logs, and documents to cloud vendors amid unannounced downgrades, bans, and terms changes.
• FBI chip-secrets case: three engineers arrested for allegedly stealing processor security and cryptography documents from Google and others — a reminder that concentrated hardware ecosystems concentrate insider risk. https://x.com/FBISanFrancisco/status/2024670479974363376

---

Source: Easy AI Daily (zhichai.net)