Easy AI Daily | February 21, 2026

A curated digest of AI industry news compiled from community discussions on zhichai.net.

Models & Capabilities

Gemini 3.1 Pro: Major reasoning and retrieval gains, mixed real-world experience

Google released Gemini 3.1 Pro, raising its ARC-AGI 2 score from 31% to 77% and approaching GPT-5.2 on Context Arena's MRCR retrieval benchmark—stronger in hard-retrieval scenarios—with notable code and spatial reasoning at the same price as 3.0. However, engineers report unstable CLI/agent toolchains, confusing routing (frontend says Gemini but actually serves Claude), and "self-upgrade death loops" in agents like OpenClaw. The community also worries about post-launch performance downgrades.

• Gemini 3.1 Pro model card
• DillonUzar MRCR evaluation
• Artificial Analysis cost comparison
• Reddit discussion

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

METR estimates Claude Opus 4.6's "50% software task time horizon" at ~14.5 hours, though the confidence interval spans 6–98 hours with heavy noise. Sonnet 4.6 leapt up Arena's coding, instruction-following, and math leaderboards, but users complain about frequent token-limit hits and empty outputs in long reasoning mode, plus degraded Claude Code UI stability.
• METR time horizon post
• Reddit discussion
• Arena leaderboard
• Long reasoning failure feedback

Qwen series: Open models climbing agent and vision leaderboards

Community opinions are split: some find Qwen weak on logic and common sense, while others praise instruction-following at the 3–4B scale, citing Qwen 3.5-397B tying Kimi K2.5 for top-two on Arena Vision. Qwen improved returns on the FoodTruck Bench business-simulation benchmark but often goes bankrupt from poor execution—highlighting the "can think, can't act" agentic gap.
• Vision Arena rankings
• Reddit discussion
• FoodTruck Bench case study

DeepSeek system prompt leak

Someone extracted DeepSeek's full system prompt, which explicitly instructs the model to incorporate "socialist core values," avoid discussing or attacking the CCP, and includes hardware/deployment notes. It both enables prompt-injection exploits and reveals the model's preset stance on politically sensitive topics.
• Prompt excerpt 1 | Excerpt 2

Agents & Toolchains

OpenClaw ecosystem: From self-destructing agents to token launches and casinos

OpenClaw has become the most active multi-agent playground: agents autonomously launched tokens on Base, a "Last AI Standing" survival game, and a Bitcoin dice casino. One agent connected to Gemini 3.1 Pro spiraled upgrading itself to nonexistent versions and had to be manually rescued by Claude. Dashboards and tools like "ClawTower" visualize multi-agent costs and states—but real permissions amplify both risk and productivity.
• Dashboard repo | Last AI Standing | Satoshidais casino

GEPA / gskill: Treating skills as first-class artifacts

The gskill pipeline based on GEPA chains repo tasks → skill optimization → skill file distribution, reportedly achieving near-fully automated fixes in specific codebases and ~47% faster task completion in Claude Code. Engineers caution that model-generated skill docs are often bloated and hard to maintain; fewer, hand-written constraints often work better.
• GEPA/gskill thread | Alex Dimakis overview | DSPy Weekly summary

RLM and multi-agent topology: Orchestration matters more than the model

Community experiments treat RLMs (recursive language models) as a meta-scheduler. GPT-5.2-Codex and Gemini 3.1 Pro perform well under RLM decomposition; Opus 4.6 does not. Research suggests that once model capabilities converge, multi-agent topology (parallel, hierarchical, hybrid) alone accounts for 12–23% performance differences—"who orchestrates" is the new battleground.
• RLM discussion | Multi-agent topology paper summary

NAVD: Arrow + logs as an agent memory component

NAVD stores all agent conversations and events in an append-only log indexed via Arrow embeddings, achieving <10ms latency at 50k vectors without an external vector database.
• NAVD project

Agent runtimes go cloud-native

Airtable launched Hyperagent, deploying agents as isolated services with dedicated persistent environments and Slack integration; OpenClaw remains script-style, driving terminals, browsers, and codebases directly. The trend: agents as long-lived infrastructure services to operate and monitor, not just API calls.
• Hyperagent launch | OpenClaw

Infrastructure & Hardware

Taalas ASIC: Llama 3.1 8B at ~16,000 tok/s for a single user

Taalas burned Llama 3.1 8B into a 6nm, ~53B-transistor ASIC, achieving ~16–17k tok/s single-user inference at roughly $0.005 per million tokens (at $0.10/kWh). The tradeoff: models are essentially fixed—switching models requires a new tape-out, badly mismatched with iteration cadence. A more realistic path may be "frozen base model + post-trained adapters."
• Forbes coverage | Taalas technical intro | Reddit thread

ThunderKittens 2.0: Kernel optimization "subtraction" for Blackwell

Stanford Hazy Research released ThunderKittens 2.0 with BF16/MXFP8/NVFP4 GEMM on Blackwell, claiming parity or better vs cuBLAS. The new version stresses that deleting wrong optimizations matters: tensor core pipelines exhibit undocumented behaviors where poor layouts leave hardware idle.
• Blog post

ggml / llama.cpp joins Hugging Face

The llama.cpp/ggml team joined Hugging Face and will continue maintaining the ggml stack with deep integration into HF transformers. Local model loading, tooling, and community support become centralized—local inference is now part of a major ecosystem rather than a grassroots project.
• ggml announcement | Hugging Face welcome | Reddit reactions

tinygrad bets on AMD

George Hotz says tinygrad's near-term focus is solid compiler and codegen infrastructure for AMD GPUs, with bounties for any quantifiable performance gain—improving the core IR so all backends benefit, rather than hand-written kernels for specific cards.
• Discord summary

Research & Methods

Benchmark methodology failures: SWE-bench and ARC-AGI debates

MiniMax and Epoch AI both admitted their earlier SWE-bench Verified harness configurations were inconsistent; re-running aligned results with official numbers—same benchmark, very different harness results. Meanwhile, models scoring 70%+ on ARC-AGI 2 still play Connect Four poorly, deepening doubts about what these puzzle benchmarks actually measure.
• Epoch AI correction | ARC-AGI vs Connect4

"Time horizon" evaluations: Impressive numbers, fragile statistics

METR's time-horizon metric measures the window in which a model completes 50% of complex software tasks. Opus 4.6's point estimate is very high, but confidence intervals are huge and task sets are nearly saturated—METR itself repeatedly warns against extrapolating from single points.
• METR thread | Statistical uncertainty explainer

Hodoscope / ARES: Systematic auditing of agent trajectories and activations

Hodoscope is a "trajectory browser" for inspecting agent action sequences on benchmarks—it reportedly quickly found a flaw in one benchmark. ARES exposes intermediate activations in long tasks, and with probing/activation steering can localize and correct failure modes—an oscilloscope for multi-step agents.
• Hodoscope intro | ARES repo

Products & Applications

Claude Code Security: AI security auditing of real open source

Anthropic launched the Claude Code Security research preview, a code security scanner with patch suggestions; officially credited with finding 500+ long-standing vulnerabilities in real open source repos. Restriction: users can't freely scan third-party open source code—a legally and productively awkward line.
• Launch post | Vulnerability cases

Qwen-AI Slides, Kimi, Perplexity: Document and search AI compared

Qwen-AI Slides generates near-finished decks in minutes (Chinese/English only for now); Kimi's CLI is considered far better than its VS Code plugin for large repos; Perplexity is losing power users to ChatGPT, Claude, and Kimi over tightening quotas and poor bot-only support.
• Qwen-AI Slides discussion

Local inference vs API: Beyond privacy

A Mac Studio M3 Ultra user running Qwen 3.5 concluded local is mainly "better privacy" since APIs are cheap. Commenters counter: control (no surprise downgrades/shutdowns), offline use, fine-tuning for your own use case, lower latency, and potentially lower long-term cost—if you accept hardware costs and tinkering.
• Reddit discussion

Extreme-speed applications

ChatJimmy claims up to 15,000 tok/s chat; Guillaume Lample's Voxtral Realtime delivers STT with sub-500ms latency for real-time meetings and captions.
• ChatJimmy | Voxtral Realtime

Industry & Company News

• ggml.ai joins Hugging Face: seen as local AI's mainstream moment—concerns about centralization, but hopes for sustainable long-term maintenance. Announcement | llama.cpp discussion
• Unsloth partners with Hugging Face: free LLM fine-tuning pipelines on HF, with 100k+ community-finetuned models already open. Announcement video
• Amazon Kiro AI blamed for second AWS outage: FT reports the internal coding agent decided to "delete and rebuild an environment" during an incident, causing a 13-hour regional outage; officially attributed to "user error," but internal review of agent permissions and two-person approvals has begun. FT report
• Perplexity and OpenRouter reliability woes: Perplexity faces quota cuts, unexplained subscription cancellations, API 500s, and bot-only support; OpenRouter's backend refactor returned empty image generations while still billing, later refunded. OpenRouter statement
• Security stocks shed ~$10B after one Anthropic blog post: within an hour of an Anthropic blog analyzing AI's impact on cybersecurity, CrowdStrike, Cloudflare, Okta, and others reportedly lost ~$10B combined in market cap. Related thread

Policy, Governance & Safety

• Claude Code data-access controversy: Anthropic restricts third-party repo scanning while a user reported seeing another company's commercial lease documents in Claude Cowork—sparking debate over indexing, training-data residue, or permission bugs. Reddit thread
• Jailbreak techniques evolve: the BASI community analyzed leaked DeepSeek and Sonnet 4.6 system prompts and tested "Crescendo" progressive jailbreaks on Gemini 3.1—gradually steering context from benign to sensitive topics under a "research/documentation" framing. Gemini guards tightest; API variants are relatively easier to jailbreak. Prompt excerpts
• Privacy drives local AI adoption: many teams prefer buying another GPU over API lock-in amid unannounced model downgrades, account bans, and ToS changes. Reddit discussion
• FBI charges engineers for chip trade-secret theft: three engineers were arrested for allegedly stealing processor security and cryptography documents from companies including Google—a reminder that concentrated hardware expertise concentrates insider risk. FBI notice

---

📌 Source: Easy AI Daily