Easy AI Daily News Digest – February 21, 2026

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for February 21, 2026 covers major AI industry developments: Google released Gemini 3.1 Pro with a large ARC-AGI 2 score jump (31% to 77%) and strong retrieval performance, though developers report unstable tooling and routing issues. METR estimated Claude Opus 4.6's 50% task time horizon at roughly 14.5 hours with wide confidence intervals, while Sonnet 4.6 climbs coding leaderboards but suffers token-limit failures in long reasoning. The ggml/llama.cpp team joined Hugging Face, and Unsloth partnered with HF for free fine-tuning. Taalas unveiled an ASIC running Llama 3.1 8B at ~16,000 tok/s per user; Stanford released ThunderKittens 2.0 for Blackwell kernels; tinygrad doubles down on AMD compiler infrastructure. Benchmark methodology controversies hit SWE-bench and ARC-AGI, Anthropic launched Claude Code Security (500+ vulnerabilities found), Amazon's Kiro AI was linked to two AWS outages, and security stocks lost ~$10B after an Anthropic blog post. Policy items include leaked DeepSeek system prompts, jailbreak techniques, and FBI charges over stolen chip secrets.

Easy AI Daily – February 21, 2026

A roundup of AI models, agent tooling, infrastructure, research, products, industry news, and safety/policy developments.

Key points

Models & Capabilities

Gemini 3.1 Pro released. ARC-AGI 2 score jumped from 31% to 77%; near GPT-5.2 on Context Arena's MRCR retrieval (stronger on hard-retrieval cases), with notable code and spatial reasoning gains at the same price as 3.0. However, engineers report unstable CLI/agent tooling, routing confusion (Gemini frontends serving Claude), and self-upgrade death loops in agents like OpenClaw; the community fears post-launch performance cuts.
Links: Model card | MRCR eval | Cost comparison | Reddit discussion
Claude Opus/Sonnet 4.6. METR estimates Opus 4.6's 50% software-task time horizon at ~14.5 hours, but with a 6–98 hour confidence interval and heavy noise. Sonnet 4.6 surges on Arena coding/instruction/math leaderboards, yet users hit token limits and empty outputs in long-reasoning mode, plus degraded Claude Code UI stability.
Links: METR post | Reddit | Arena leaderboard
Qwen series. Polarized reception: some find logic/common sense weak; others cite strong instruction-following at 3–4B scale and Qwen 3.5-397B tying for top-two on Arena Vision with Kimi K2.5. In FoodTruck Bench agentic simulations, revenue improved but agents still often go bankrupt—highlighting the "can reason, can't execute" gap.
Links: Vision Arena | Reddit | FoodTruck Bench case study
DeepSeek system prompt leak. The full system prompt reportedly instructs the model to embed "socialist core values" and avoid discussing/attacking the CCP, plus hardware/deployment details—exposing preset political stances and material useful for prompt-injection abuse.
Links: Snippet 1 | Snippet 2

Agents & Tooling

OpenClaw ecosystem: agents launching tokens on Base, a "Last AI Standing" survival game, and a Bitcoin dice casino; one agent self-upgraded to nonexistent versions after hooking Gemini 3.1 Pro and had to be rescued manually. Dashboards like ClawTower visualize multi-agent cost/state, but real permissions amplify both risk and productivity. (Dashboard repo | Last AI Standing | Satoshidais)
GEPA / gskill pipeline treats skills as first-class artifacts (repo task → skill optimization → skill file dispatch), reportedly enabling near-fully automatic fixes in specific repos and ~47% faster Claude Code task completion. Counterpoint: model-generated skill docs are often redundant; fewer, hand-written constraints work better. (Intro thread | Dimakis overview | DSPy Weekly)
RLM as meta-scheduler: GPT-5.2-Codex and Gemini 3.1 Pro do well under RLM decomposition; Opus 4.6 fits poorly. Research suggests multi-agent topology (parallel/hierarchical/hybrid) alone yields 12–23% performance differences once model capabilities converge. (RLM discussion | Topology paper summary)
NAVD replaces vector DBs with append-only logs plus Arrow embedding indexes, claiming <10ms latency at 50k vectors. (Project page)
Cloud-native agent runtimes: Airtable's Hyperagent deploys agents as isolated services with dedicated persistence and Slack integration; OpenClaw remains script-style (terminal, browser, codebase). Agents are becoming long-lived infrastructure services. (Hyperagent | OpenClaw)

Infrastructure & Hardware

Taalas ASIC: a 6nm, ~53B-transistor chip hardwires Llama 3.1 8B for ~16–17k tok/s single-user inference (~$0.005 per million tokens at $0.10/kWh). Trade-off: swapping models requires a new tape-out; a pragmatic path may be frozen base models + adapter post-training. (Forbes | Taalas | Reddit)
ThunderKittens 2.0 (Stanford Hazy Research): BF16/MXFP8/NVFP4 GEMM on Blackwell, matching or beating cuBLAS; removing wrong optimizations matters as much as adding good ones due to undocumented tensor-core pipeline behavior. (Blog)
ggml/llama.cpp joins Hugging Face: the team will maintain the ggml stack and integrate deeply with HF transformers—local inference is now mainstream ecosystem territory. (ggml announcement | HF welcome | Reddit)
tinygrad bets on AMD: George Hotz prioritizes solid compiler/codegen infrastructure and bounties for quantifiable gains over hand-written per-card kernels. (Discord summary)

Research & Methods

Benchmark methodology problems: MiniMax and Epoch AI both admitted their SWE-bench Verified harness configs differed from official ones; reruns aligned with official scores. Meanwhile, models scoring 70%+ on ARC-AGI 2 still play Connect Four poorly, renewing doubts about what such benchmarks measure. (Epoch correction | ARC-AGI vs Connect4)
Time-horizon metrics: Opus 4.6's headline number is high, but confidence intervals are huge and task sets near saturation—METR itself warns against linear extrapolation. (METR thread | Statistical uncertainty)
Hodoscope / ARES: Hodoscope is a trajectory browser for auditing agent benchmark behavior (already found a benchmark bug); ARES exposes intermediate activations with probing/steering to locate and fix multi-step failure modes. (Hodoscope | ARES repo)

Products & Applications

Claude Code Security (research preview): a security scanner with patch suggestions; Anthropic claims 500+ long-standing vulnerabilities found and fixed in real open-source repos. Restriction: users may not scan arbitrary third-party code—legally and productively awkward. (Launch)
Document/search AI round: Qwen-AI Slides generates near-finished decks in minutes (Chinese/English only); Kimi's CLI beats its VS Code plugin for large-repo agent chats; Perplexity loses heavy users to ChatGPT/Claude/Kimi over tightened limits and bot-only support. (Qwen Slides)
Local inference vs API: beyond privacy—control (no surprise downgrades), offline use, fine-tuning, lower latency, potentially lower long-term cost—against hardware upfront cost and setup effort. (Reddit)
Speed-focused apps: ChatJimmy claims ~15k tok/s chat; Voxtral Realtime does speech-to-text with sub-500ms latency for live meetings/captions. (ChatJimmy | Voxtral)

Industry & Company News

ggml.ai/llama.cpp joins Hugging Face (see above): mixed feelings—worry over centralization vs. hope for sustainable maintenance. (GitHub discussion)
Unsloth partners with Hugging Face on free LLM fine-tuning; 100k+ Unsloth-tuned models already open-sourced on HF. Data, not compute, is now the bottleneck for individuals. (Announcement)
Amazon Kiro AI incidents: FT reports the internal coding agent once 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," but internally rethinking agent permissions and two-person approval flows. (FT)
Perplexity & OpenRouter outages: Perplexity upset Pro users with tightened limits and API errors; OpenRouter's backend refactor returned empty image generations while still billing—compensated with refunds. (OpenRouter statement)
Security stocks drop: ~$10B in combined market value (CrowdStrike, Cloudflare, Okta, etc.) wiped within an hour of an Anthropic blog post analyzing AI's impact on cybersecurity. (Tweet roundup)

Policy, Governance & Safety

Claude Code data-access controversy: Anthropic restricts scanning third-party repos while a user reported seeing another company's commercial lease documents in Claude Cowork—raising training-data, permission, and legal-compliance questions regardless of whether it was a hallucination. (Reddit)
Jailbreak techniques: leaked DeepSeek/Sonnet 4.6 system prompts circulate; "Crescendo"渐进越狱 (gradual context escalation framed as research) tested on Gemini 3.1—Gemini holds up best, API endpoints are easier targets.
Local AI for privacy: many teams prefer buying hardware over APIs to avoid unannounced model downgrades, bans, and terms changes. (Reddit)
FBI chip-theft case: three engineers arrested for stealing processor-security and cryptography documents from Google and others—a reminder that concentrated hardware expertise creates human insider risk. (FBI notice)

---

📌 Source: Easy AI Daily