Easy AI Daily News | February 19, 2026: Gemini 3.1 Pro, OpenClaw Boom, and AI Industry Roundup

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for February 19, 2026 covers Google's release of Gemini 3.1 Pro, which scores 77.1% on ARC-AGI-2 (roughly double Gemini 3.0), 80.6% on SWE-Bench Verified, with 1M-token context and unchanged pricing, rolling out across the Gemini app, AI Studio, Vertex AI, OpenRouter, Perplexity, Cursor, and Windsurf. Other model news includes DeepMind's Lyria 3 music generator, Trillion Labs' open-source Tri-21B-think, and Mistral's Voxtral Realtime speech model with sub-500ms latency. The agent ecosystem saw the viral OpenClaw tool draw attention for high API costs and account bans, while Anthropic clarified that personal OAuth tokens cannot be used in third-party tools. Infrastructure news covers OpenRouter outages and billing bugs, FP8 long-context training tips from GPU MODE, and the Terradev CLI for multi-cloud GPU price comparison. Research highlights include the Every Eval Ever evaluation schema coalition, GPT-2 small attention head analysis, and residual-stream swap experiments revealing causal commitment layers. Policy items: OpenAI's $7.5M alignment grant and FedRAMP authorization, Perplexity's quota cuts and account bans, Snap's hardware VP departure, and TOTO's surprising AI chip materials pivot.

Key points

Models & Capabilities

Google released Gemini 3.1 Pro: a scaled-down sibling of Deep Think scoring 77.1% on ARC-AGI-2 (~2x Gemini 3.0), 80.6% on SWE-Bench Verified, and 68.5% on Terminal-Bench 2.0; it leads 6 of 10 Artificial Analysis evaluations, cuts hallucinations, and offers 1M context / 64k output with unchanged API pricing.
Broad rollout: available in the Gemini app, NotebookLM, AI Studio, Vertex AI, plus OpenRouter, Perplexity, Cursor, and Windsurf (0.5x credit promo at Windsurf).
Community testing: strong code generation, long-instruction following, and multimodal gains; however, GDPval real-world task scores lag rivals and tooling rollouts (Antigravity, CLI, Code Assist) are uneven.
Lyria 3: DeepMind's new music model generates 30-second tracks from text/image/video in the Gemini UI; audio quality praised, arrangement creativity questioned.
Tri-21B-think: Trillion Labs' Apache-2.0 open reasoning preview (21B) shows strong tool-use scores and low hallucination at high token cost; weights only, no hosted endpoint.
Voxtral Realtime: Mistral's open (Apache-2) real-time speech recognition with sub-500ms end-to-end latency.

Agents & Tooling

OpenClaw goes viral but burns cash: users report $1,000+/day API bills; Anthropic bans third-party use of personal OAuth tokens, with reported account bans; community forks like vibeclaw.dev and Rust-based DeepCLI are emerging.
Autonomous tool-building: OpenClaw agents have written their own dashboards, learned an on-chain MMORPG in 20 minutes with Haiku 4.5, and shipped an n8n community node integrating the full OpenClaw Gateway toolset.
Observability upgrades: LangChain/LangSmith trace filtering, a sub-agent debugging case study, and Raindrop's trajectory browser.
Qbit: open-source IDE combining terminal, Git, model selection, and MCP tools with agent-driven but human-reviewable operations.
A user shared a ~3,000-hour Claude Code pipeline spanning requirements, architecture review, atomic task decomposition, multi-model development, adversarial review, and QA—emphasizing context hygiene and locking intent into rule docs.

Infrastructure & Hardware

OpenRouter instability: two database outages, Clerk login degradation, and a bug charging for image tokens without returning images (refunds promised).
GPU MODE: stable FP8 training of a 0.5B model for 350B tokens on 4×4090s; DirectML nears CUDA on Windows but is in maintenance mode; NVIDIA leaderboard/Modal runner Cutlass version mismatches cause submission errors.
Terradev CLI 2.9.2: multi-cloud GPU price arbitrage, BYO API scheduling, and cost attribution.
Local LLM consensus: more VRAM always helps; NVLink rarely the inference bottleneck; ROCm approaching CUDA but ecosystem lags.

Research

Every Eval Ever: EleutherAI, Hugging Face, and University of Edinburgh launched the EvalEval coalition with a unified evaluation schema, crowdsourced dataset, and an ACL 2026 workshop/shared task.
A reproducible LLM eval pipeline logging judge disagreements, retries, cost, and latency.
GPT-2 small interpretability: ~75% of attention heads work with low-rank/fixed-pattern QK (+5.3% WikiText-2 val loss); many BOS-sink heads.
Residual-stream swaps reveal causal commitment points at ~62–71% depth across GPT-2 Small, Gemma-2-2B, and Qwen2.5-1.5B.
STATe-of-Thoughts: Tree-of-Thours-style search in DSPy with early stopping and vLLM defaults (paper).

Products & Business

Perplexity Pro quota cut from 600/day to 200/week for enhanced queries, alongside wave of account bans for alleged ToS violations.
Cursor/Windsurf shipped Gemini 3.1 Pro; opinions mixed on whether it matches Opus 4.6 in practice.
Gradio 6 gr.HTML enables one-file, one-shot web apps; Rover offers an embeddable web agent that reads the DOM and clicks for users; Manus job-application helper praised amid a reported ~$2,500 overcharge complaint, with Meta acquisition confirmed.
TOTO, the Japanese toilet maker, is entering AI chip packaging/materials via its ceramics expertise, targeting a ~$60B market—stock jumped ~60%.
Epoch models suggest Anthropic revenue could match OpenAI by mid-2026 under current growth rates.
Snap's SVP of hardware (Spectacles) departed after strategic clashes with CEO Evan Spiegel.

Policy & Safety

OpenAI: $7.5M grant to the AI Security Institute's Alignment Project and reported FedRAMP authorization.
Anthropic clarified OAuth tokens for free/Pro/Max are official-products-only; third-party or Agent SDK use violates terms.
Proposals for Ed25519 cryptographic "passports" to authenticate autonomous agents and cap their spending.
Jailbreak community BASI shows misuse risks (dangerous info retrieval via papers) while members warn each other about phishing links.

---

📌 Source: Easy AI Daily (zhichai.net)