Easy AI Daily | February 19, 2026

English edition of the Easy AI daily roundup (2026-02-19), covering model releases, agentic tooling, infrastructure, research, products, industry news, and AI policy.

Models & Capabilities

Google Releases Gemini 3.1 Pro: 77.1% on ARC-AGI-2, Same Price

Gemini 3.1 Pro launched as a "scaled-down sibling" of Deep Think. It scores 77.1% on ARC-AGI-2 (roughly 2x Gemini 3.0), 80.6% on SWE-Bench Verified, and 68.5% on Terminal-Bench 2.0, taking first place in 6 of 10 Artificial Analysis evaluations with significantly lower hallucination rates. Context: 1M tokens; output: 64k; knowledge cutoff: January 2025; API pricing unchanged from 3 Pro.

Links: Official announcement | Model card | Artificial Analysis review | ARC Prize cost/scores

Gemini 3.1 Pro Rolls Out Across Products and Platforms

Available in the Google Gemini App, NotebookLM, Google AI Studio, and Vertex AI, plus third-party access via OpenRouter, Perplexity, Cursor, and Windsurf. Windsurf offers a 0.5x credit promotion; Perplexity Pro/Max subscribers can switch directly.

Community Testing: Strong Reasoning, Weak GDPval

Early feedback: code generation, long-instruction following, and multimodal performance clearly improved over 3.0. Criticisms include GDPval real-world task scores lagging other frontier models and uneven rollout of UI/IDE tools (Antigravity, CLI, Code Assist). Verdict: strong benchmarks, real-world utility still under observation.

Lyria 3: DeepMind's New Music Generation Model

Integrated into the Gemini interface, generating 30-second music from text, images, or video. Users report audio quality (especially distorted guitars and complex instruments) better than Suno with fewer artifacts, but arrangements feel "boring" and the length limit is questioned.

Trillion Labs Open-Sources Tri-21B-think Preview

A 21B reasoning-focused open model (Apache-2.0). Mid-tier overall on Artificial Analysis, but strong on some tool-use benchmarks with relatively low hallucination rates — at the cost of very high reasoning token usage. Weights only; no hosted endpoint yet.

Mistral Voxtral Realtime: Open Real-Time Speech Recognition

End-to-end latency under 500ms for real-time transcription; model and code released under Apache-2.

Agents & Tooling

OpenClaw Goes Viral: High Costs, Ban Risks, Ecosystem Forks

Users report burning $1,000+ per day in API costs, some switching back to Claude Code. Anthropic explicitly bans using personal OAuth tokens with third-party tools (including the Agent SDK); some users were apparently banned for using Claude Pro/Max keys in OpenClaw. Google reportedly restricted antigravity OAuth use externally. Alternatives emerging: vibeclaw.dev (local deployment) and Rust-based DeepCLI.

OpenClaw Agents Start "Writing Their Own Tools"

One agent set up its own dashboard and security system in production; another learned an on-chain MMORPG in 20 minutes with Haiku 4.5, writing Python scripts, sending mining transactions, and scheduling cron jobs. A community n8n node orchestrates all OpenClaw Gateway tools in visual workflows.

Agent Debugging & Observability Updates

LangChain/LangShip trace filtering and observability improvements plus a startup program; Raindrop introduces a "trace browser" for replaying agent call chains.

Qbit: Terminal, Git, and MCP in One Agent IDE

Open-source IDE focused on an "agent-powered terminal" with project management, unified timeline, model selection, inline editing, Git integration, and MCP tools — keeping humans in the loop with visibility and rollback.

Claude Code Deep Workflow: 3,000 Hours of "Multiplayer IDE"

A power user built a full pipeline in ~3,000 hours: requirements clarification, architecture review, atomic task breakdown, multi-model collaborative development, adversarial code review, QA/security audit. Key lessons: minimize context noise, lock human intent into rules docs rather than fully delegating to agents.

Infrastructure & Hardware

OpenRouter Outages and Billing Bug

Two database outages (~30 minutes each, early morning) plus Clerk identity provider degradation caused login issues. Some users on the new backend were charged for image tokens without receiving images; the team acknowledged missed edge cases in testing and promised refunds.

GPU MODE Community Highlights

A 0.5B model trained stably on 4×4090s at FP8 for 350B tokens (clean data + small models credited); DirectML ONNX inference on Windows approaches CUDA but lacks Linux support and is in maintenance mode; NVIDIA leaderboard vs Modal runner Cutlass version mismatch causing submission errors.

Terradev CLI: Multi-Cloud GPU Price Comparison

Avoid single-cloud GPU price gouging with BYO-API multi-cloud scheduling, real job cost tracking, and one-click deploy to Hugging Face Spaces. v2.9.2 adds multi-cloud GPU arbitrage and finer cost attribution.

Local LLM Ecosystem: "Expensive but Fun"

LM Studio community consensus: more VRAM is always useful, but privacy, learning, and avoiding cloud churn drive purchases. NVLink helps inference throughput less than expected; bandwidth is rarely the bottleneck. ROCm approaches CUDA performance in some cases but the ecosystem remains thin.

Research & Methods

Every Eval Ever: Unified Evaluation Data Format

EleutherAI, Hugging Face, and the University of Edinburgh launched the EvalEval alliance's "Every Eval Ever" initiative: a unified schema for LLM evaluation results plus a crowdsourced dataset to align HELM, lm-eval-harness, and Inspect AI results. A workshop and shared task are planned for ACL 2026, with co-authorship for top contributors.

Links: Website | GitHub | Dataset

Industrializing LLM Evaluation

An open reproducible eval pipeline logs judge disagreement, retries/failures, and per-experiment cost and latency — not just leaderboard numbers. Focus shifts to "how much money and time does one run cost" and evaluation stability.

Attention Interpretability: 75% of GPT-2 Small Heads Don't Need Full-Rank QK

Structural constraints show ~75% of GPT-2 Small attention heads can be replaced with low-rank or fixed-pattern QK while maintaining or slightly improving validation loss (~5.3% gain on WikiText-2). Many "bos-sink" heads indicate heavy reliance on the first token as an information hub.

Residual Stream Swaps Reveal a "Causal Commitment Layer"

Layer-wise residual swaps on GPT-2 Small, Gemma-2-2B, and Qwen2.5-1.5B show a causal commitment point at ~62-71% depth: swaps before it barely change outputs, swaps after it drastically alter results — quantitative evidence of where models "decide."

STATe-of-Thoughts: Tree-of-Thoughts in DSPy

Integrates Tree-of-Thoughts into DSPy with early stopping to avoid context rot, text interventions to control branch diversity, and a self-hosted vLLM cluster instead of expensive closed APIs. Includes a persuasive-argument case study analyzing why different reasoning paths work.

Links: Code | Paper

Products & Applications

Perplexity Tightens Usage: Pro Moves from 600/Day to 200/Week, Bans Follow

Perplexity Pro users found "enhanced query" limits cut from 600/day to 200/week, alongside sudden bans citing "terms violations" with template-only support replies. Speculation centers on cracking down on discount keys/resale, or pushing Pro users to Max.

Cursor and Windsurf Quickly Adopt Gemini 3.1 Pro

Cursor users debate whether its coding capability matches or slightly exceeds Opus 4.6, with some saying real-world performance lags benchmark scores; community stresses .cursorrules and architecture docs for stronger context. Windsurf offers half-price credits to drive migration.

Gradio 6 Adds gr.HTML: Full Web Apps in One File

Write entire frontends (dashboards, pomodoro timers) in a single Python file, generatable by an LLM in one shot. See the official blog.

Rover: An Embeddable Web Agent That Clicks Your Pages

A script snippet that lets an agent read the DOM, click buttons, and complete actions for users — no API or backend needed. Positioning: give any website a UI-capable assistant. Website

Manus Job Assistant: Praised for Usability, Criticized for a $2,500 Overcharge

Users praise Manus for auto-filling resumes and application forms, but one reports being charged ~$2,500 for a $680 plan with unresponsive support and a planned BBB complaint. The community also confirmed Manus has been acquired by Meta.

Industry & Company News

TOTO Enters AI Chips via Ceramics Expertise

Japanese toilet-maker TOTO is leveraging high-end ceramic capabilities to enter AI chip packaging/materials, targeting an ~$60 billion market. The stock surged ~60% on the news — an extreme case of "toilet company pivots to compute."

AI Revenue Race: Anthropic Could Match OpenAI by 2026

Epoch's models suggest Anthropic's revenue could catch OpenAI around mid-2026 at current growth rates — with the caveat that such extrapolations are highly sensitive to marginal slowdowns and price wars.

Snap Hardware Chief Departs Over Strategic Conflict

Snap's SVP for Spectacles left after clashes with CEO Evan Spiegel over hardware strategy, ending six years of hardware exploration — widely read as a pullback from consumer AR.

Policy, Governance & Safety

OpenAI Announces $7.5M Alignment Funding and FedRAMP Authorization

OpenAI committed $7.5M to the AI Security Institute's Alignment Project for systematic safety/alignment research, and reportedly obtained FedRAMP authorization, enabling US federal procurement of some services under cloud security standards.

Anthropic Bans Personal OAuth Tokens in External Tools

Updated terms clarify Claude Free/Pro/Max OAuth tokens may only be used within official products; using them in third-party tools (including the Agent SDK) violates policy — explaining recent account bans for browser-hijack tools.

Agent Identity: Encrypted "Passports" for AI Agents

With autonomous agents calling APIs, transferring funds, and ordering online, a proposal suggests issuing Ed25519 cryptographic "passports" per agent for signing, reputation tracking, and spending limits — preventing impersonation and resource abuse.

BASI Community: Jailbreaks, Weapon Recipes, and Self-Policing

Discussions range from extracting dangerous information from public papers to prompt-induced misbehavior; members increasingly warn each other about phishing and law-enforcement risks from unknown links.

---

📌 Source: Easy AI Daily (zhichai.net)