Easy AI Daily | February 19, 2026

A digest of AI industry news covering models, agents, infrastructure, research, products, and policy.

Models & Capabilities

Google releases Gemini 3.1 Pro: 77.1% on ARC-AGI-2 at unchanged pricing

• Gemini 3.1 Pro, described as a "scaled-down sibling" of Deep Think, 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
• Tops 6 of 10 evaluations on Artificial Analysis, with significantly reduced hallucination rates
• 1M context window, 64k output, knowledge cutoff January 2025, API pricing unchanged from Gemini 3 Pro
• Links: release post | DeepMind model card

Broad platform availability

• Live in Gemini App, NotebookLM, Google AI Studio, and Vertex AI; available via OpenRouter, Perplexity, Cursor, and Windsurf (which offered 0.5x credit promotion)

Community reception: strong reasoning, weaker GDPval

• Feedback highlights improved code generation, long-instruction following, and multimodal performance vs 3.0
• Criticisms: GDPval real-world task scores trail other frontier models; uneven rollout of UI/IDE tooling (Antigravity, CLI, Code Assist)
• Reddit discussion

Other model releases

• Lyria 3: DeepMind's new music model in the Gemini interface, generates 30-second tracks from text, images, or video. Better audio quality than Suno per users, but compositions seen as less creative
• Tri-21B-think: Trillion Labs' 21B open reasoning model (Apache-2.0); moderate overall scores but strong on some tool-use benchmarks with low hallucination, at high token usage
• Voxtral Realtime: Mistral's open-source (Apache-2) real-time speech recognition model with claimed end-to-end latency under 500ms

Agents & Tooling

OpenClaw goes viral amid cost and ban issues

• Users report burning $1,000+/day in API costs; some switched back to Claude Code
• Anthropic prohibits using personal OAuth tokens in third-party tools including the Agent SDK; some users of Claude Pro/Max keys in OpenClaw were banned
• Community alternatives emerging: local deployment via vibeclaw.dev, Rust-based DeepCLI
• Notable experiments: agents writing their own tools, an n8n integration (n8n-nodes-openclaw), and an agent learning an on-chain MMORPG in 20 minutes with Haiku 4.5

Other tooling news

• LangChain/LangSmith: new trace filtering and observability features plus a startup program; Raindrop shipped a "trace browser" for replaying agent call chains
• Qbit: open-source IDE combining terminal, Git, and MCP tools with agent-driven operations and human review/rollback
• A user shared a ~3,000-hour Claude Code pipeline covering requirements, architecture review, task decomposition, multi-model development, adversarial review, and QA/security audits — key lesson: minimize context noise and lock human intent into rules docs

Infrastructure & Hardware

• OpenRouter: two database outages, Clerk login issues, and a billing bug charging image_tokens without returning images; refunds promised
• GPU MODE: stable FP8 training of a 0.5B model for 350B tokens on 4×4090s; DirectML接近 CUDA on Windows but is in maintenance mode; NVIDIA leaderboard/Modal runner Cutlass version mismatch causing submission errors
• Terradev CLI 2.9.2: multi-cloud GPU price comparison, BYO API scheduling, and cost attribution (PyPI)
• Local LLM community consensus: more VRAM always helps; privacy and learning drive adoption; NVLink offers limited inference gains; ROCm approaches CUDA performance but the ecosystem remains weak

Research & Methods

• Every Eval Ever: EleutherAI, Hugging Face, and University of Edinburgh launched the EvalEval consortium defining a unified schema for LLM eval results, with a crowdsourced dataset and an ACL 2026 workshop (site, GitHub)
• Reproducible eval pipeline: open-source pipeline logging judge disagreements, retries, cost, and latency per experiment
• Attention interpretability: ~75% of GPT-2 Small attention heads can use low-rank or fixed-pattern QK while maintaining validation loss (~5.3% improvement on WikiText-2); many "bos-sink" heads found
• Causal commitment layers: residual stream swap experiments on GPT-2 Small, Gemma-2-2B, and Qwen2.5-1.5B show representations "lock in" around 62–71% depth (preprint)
• STATe-of-Thoughts: Tree-of-Thoughts in DSPy with early stopping and branch diversity control (paper)

Products & Applications

• Perplexity: Pro enhanced queries cut from 600/day to 200/week; wave of account bans for "terms violations" with template-only support responses; suspected crackdown on discounted keys
• Cursor & Windsurf: both integrated Gemini 3.1 Pro; debate over whether benchmark scores translate to real-world coding ability
• Gradio 6: new gr.HTML component enables full web apps in a single Python file (blog)
• Rover: embeddable web agent that reads DOM and clicks UI elements via a script tag (site)
• Manus job-hunting assistant: praised for auto-filling applications, but a user reported ~$2,500 charged for a $680 plan; community confirmed Meta acquisition

Industry & Business

• TOTO (Japanese toilet maker) is entering AI chip packaging/materials leveraging ceramics expertise, targeting a ~$60B market; shares surged ~60% on the news
• Epoch AI models suggest Anthropic could match OpenAI revenue by mid-2026 if current growth rates hold
• Snap: hardware SVP leading Spectacles/Specs departed after strategy conflicts with CEO Evan Spiegel, signaling a pullback on consumer AR hardware

Policy, Governance & Safety

• OpenAI: $7.5M funding for the AI Security Institute's Alignment Project; reportedly received FedRAMP authorization for US federal procurement
• Anthropic: clarified that personal OAuth tokens for Claude Free/Pro/Max may only be used in official products; third-party tool use (including Agent SDK) violates terms
• Agent identity: a proposal for Ed25519 cryptographic "passports" for AI agents to enable signing, reputation tracking, and spending limits (proposal)
• Jailbreaking community (BASI): discussions ranging from extracting dangerous information via public papers to prompt-based manipulation of DeepSeek; members increasingly warning each other against phishing links

📌 Source: Easy AI Daily