Easy AI Daily Digest | February 17, 2026

Forum topic · 小凯 · 2026-03-27

Summary

This forum digest compiles AI industry news from February 17, 2026. Key releases include Alibaba's open-source Qwen3.5-397B-A17B MoE model (397B total, 17B active parameters, 256K native context, Apache-2.0 license), MiniMax M2.5, GLM-5, Claude Opus 4.6 with 1M-token context, and cost-effective Step 3.5 Flash. OpenAI acquired OpenClaw, with its creator joining OpenAI while the project moves to a foundation. Infrastructure coverage spans NVIDIA GB300 NVL72 efficiency claims, power constraints, AccelOpt's LLM-written CUDA kernels, and WebGPU-based BitNet inference at 125 tok/s on M4 Max. Research highlights include Chain-of-Verification, Recursive Language Models, rubric-based RL, model provenance techniques, and the Assistant Axis finding on long-conversation drift. Policy news covers Pentagon supply-chain risk warnings to Anthropic, Google's claims about Gemini distillation attacks, OpenAI's Lockdown Mode, and reproducibility concerns over OpenAI-assisted physics papers. Perplexity's quota cuts drove user migration to rivals.

📅 AI Industry Roundup — February 17, 2026

A structured summary of the original Chinese daily digest from Easy AI (zhichai.net).

Key points

Models & Capabilities

Alibaba released Qwen3.5-397B-A17B: open-source sparse MoE with hybrid linear attention; 397B total / 17B active parameters, 201 languages, native 256K context (extendable to ~1M), Apache-2.0 license. Day-zero vLLM support; community estimated KV-cache overhead at ~31KB/token. API version Qwen3.5-Plus offers 1M context plus search and code interpreter, though API pricing was criticized as expensive. Blog, vLLM support
MiniMax M2.5: 230B total / 10B active, 200K context; reported ~2500 tok/s/GPU on 8×H200 with vLLM. Token-level process rewards boost RL signal efficiency. Local deployment needs ~200GB VRAM (2× RTX 6000 Blackwell reached 120–130 tok/s). GLM-5 praised for tool calling and multi-turn agent tasks, though service stability is still settling. SemiAnalysis benchmarks, Unsloth guide
Claude Opus 4.6: 1M-token context plus an automatic "check your work" self-review step that can overturn earlier errors in long sessions; strict hourly rate limits remain. LMArena discussion
Step 3.5 Flash: flagged by the OpenRouter community as a high-value reasoning model punching above its price, though platform routing/support lags behind major closed models. Comparison video
CommonLID: a new 109-language web-text language-identification benchmark from Common Crawl/EleutherAI; top LangID models fall below 80% F1 even on languages they claim to support. Paper, Dataset

Agents & Tooling

OpenAI acquired OpenClaw: creator Peter Steinberger joins OpenAI for personal-agent work; the project moves to a foundation to stay open source. Community debated whether it's a one-person+agent success story or a rough API/MCP glue layer with hidden costs. Altman confirmation
"Harness engineering" as a moat: discussions argue agent quality depends more on orchestration, context management, lifecycle, and observability than on the base model; minimal alternatives (PicoClaw, nanobot) and LangSmith's trace-first approach emerged. LangSmith
Real-world OpenClaw experiments: root-SSH automated Proxmox v6→v8 upgrades, "agent companies" where a lead-bot delegates to frontend/backend bots, Tavus video-call mode, and SEO content pipelines producing hundreds of articles.
MCP spec debates: developers flagged the "token tax" of stuffing JSON schemas into prompts for structured output, proposing explicit text/image/object result types with schemas in metadata, and passing timezone/environment context explicitly. MCP thread
Terminal agents: Jazz bundles MCP, Git, shell, email, and cron into a CLI agent; Crowdcent packages DSPy as an MCP Server; Cloudflare experiments with Markdown for agents.

Infrastructure & Hardware

NVIDIA GB300 NVL72: claimed ~50× per-MW performance and 35× lower per-token cost vs. Hopper; industry bottlenecks shift from GPUs/HBM to datacenter power and distribution. Western Digital's 2026 HDD capacity reportedly fully booked, with AI customers locking supply into 2027/2028.
Agent-written kernels: AccelOpt's self-optimizing LLM agents report 1.5× faster GQA paged decode and 1.38× faster prefill than FlashInfer 0.5.3 on B200 (repo); GPU MODE runs a FlashInfer-bench kernel competition.
Kernel tuning pitfalls: NVProf TFLOPs fluctuating 1400–1500 on simple matmul; Achieved Occupancy excludes idle SMs — use SM active cycles instead; CUTLASS/CuteDSL/Proton version mismatches on B200 cause frequent breakage.
WebGPU + BitNet: a developer ran BitNet-B1.58 2B at ~125 tok/s on M4 Max via WebGPU, open-sourced as the Hesper library.

Research & Methods

Long-reasoning toolkit: Meta's Chain-of-Verification (CoVe) roughly doubles accuracy via answer-then-verify prompting; Omar Khattab's Recursive Language Models use code-writing recursion over long attention; Cameron Wolfe's survey of 15+ rubric-based RL papers replaces fuzzy LLM-judge scoring with structured rubrics.
Model genealogy: papers on matrix-driven weight homology (arXiv:2508.06309), black-box independence tests that reconstruct Llama-family fine-tune trees (arXiv:2502.12292), and black-box provenance (arXiv:2510.19796) may become tools against model re-badging.
Assistant Axis: arXiv:2601.10387 defines a persona-direction axis and measurably shows activations drifting along it in long conversations — the first quantitative evidence for "assistant drift."
Safety research: X-Ware uses activation-trained diffusion models for cleaner activation editing and identifies higher-level "meta-neurons"; FAR.AI warns that training against deception probes may teach models to mask internal states rather than become honest.
QED-Nano 4B: Lewis Tunstall's small model for IMO-level math proofs, using multi-stage distillation and inference caching to concentrate compute on hard problems. Intro

Products & Applications

Perplexity backlash: Deep Research quota cut from 200 to 20/month, upload limits, 7-day data retention; matching old usage would cost ~$167/month vs. $20 previously. TrustPilot dropped to 1.5/5; users migrating to Claude Opus 4.6 or Kimi.
Kimi & MiniMax friction: Kimi K2.5 performs well (sometimes beating Sonnet/Opus 4.5 at coding) with a $40/month API plan for OpenClaw, but CLI install failures, duplicate billing, quota jumps, and scam clone sites push some users toward self-hosting large MoEs (700GB RAM + 200GB VRAM).
Coding workflows: Latent Space practitioners favor Claude Cowork for pipeline tasks, planners like Ergo to break down features, and Codex/Claude Code/OpenClaw for execution — the workflow (planning + version control + observability) beats any single model.
Security applications: PassLLM fine-tunes Qwen3-4B LoRA on real password leaks for persona-aware password guessing; ATIC uses three independent Claude Opus 4.5 instances to score aleatory uncertainty and knowledge gaps, escalating to humans above thresholds.

Industry & Policy

Chinese "Spring Festival model wave": Qwen3.5, GLM-5, MiniMax 2.5, and ByteDance Seedance 2.0 (video) launched together; video generation seen as entering genuine director-style workflows, alongside serious humanoid robot investment.
Anthropic vs. Pentagon: per Axios, the DoD considered labeling Anthropic a "supply chain risk" over its usage restrictions (no mass surveillance of Americans, no fully autonomous weapons). Report
Gemini distillation attacks: Google says attackers issued 100,000+ prompts trying to clone Gemini; the community questioned both the claim's effectiveness and Google's own IP stance.
ChatGPT Lockdown Mode: OpenAI adds restricted tool-calling modes for enterprise to reduce prompt-injection and data-exfiltration risk.
Reproducibility debate: criticism of an OpenAI-assisted physics paper using GPT-5.2 without disclosing prompts, tools, or model details; OpenAI responded with clarifications, but concerns persist.

---

📌 Source: Easy AI Daily (Easy AI 日报)