Easy AI Daily News Digest – February 11, 2026

A curated English summary of the Easy AI daily digest covering models, agents, infrastructure, research, products, industry moves, and AI safety/policy.

Models & Capabilities

• Alibaba Qwen-Image-2.0: A 7B-parameter unified text-to-image + image editing model with native 2K resolution and ~1K-token prompt support. Handles high-quality typography, Chinese calligraphy, infographics, and character-consistent multi-panel comics. Shrunk from 20B to 7B versus the previous generation for faster inference and easier local deployment; open weights anticipated. (Official blog, Reddit thread)
• ByteDance Seedance 2.0: Went viral for natural motion and improved body/cloth physics; generates quality anime fight scenes and realistic 1v1 basketball clips. Currently ~15 seconds per clip; users want longer durations. (Petapixel)
• Moonshot Kimi K2.5 + Agent Swarm: Can dispatch up to ~100 sub-agents and 1,500 tool calls per task, claiming ~4.5x speedup via parallel execution. Community members built storyboard-to-video pipelines feeding Seedance 2. (Announcement)
• Claude Opus 4.6: claude-opus-4.6-thinking ranked #1 on both Arena text and coding leaderboards; reported as noticeably better at one-shotting complex UI and long-form writing than 4.5, though slower and more token-hungry. Security critics question Anthropic's reliance on internal employee surveys for high-risk capability thresholds. (Arena leaderboard, RSP critique)
• OpenAI Deep Research → GPT-5.2: Upgraded underlying model, plus external data connectors and progress controls, positioning it as a long-running research agent. (OpenAI tweet)
• Gemini 3 Pro: A new checkpoint spotted in A/B testing; meanwhile paid users report degradation (hallucinations, weaker coding) while others praise its citations and personality. (TestingCatalog)
• Underrated open multimodal models: GLM-OCR (documents), MiniCPM-o-4.5 (on-phone GPT-4o-like), InternS1 (scientific figures/papers) — all free for commercial use. (Roundup)
• Zhipu GLM-4.7-Flash GGUF became the most-downloaded model on Unsloth. (Z.ai)
• DeepMind / Isomorphic Labs IsoDDE: Technical report claims roughly doubled performance on protein/molecule structure benchmarks over AlphaFold 3, with antibody binding-site and affinity predictions exceeding physical simulations; architectural details remain scarce. (Report intro)

Agents & Tooling

• OpenAI Responses API adds server-side context compression, hosted networked containers, and first-class Skills — signaling long-running work agents as a formal product. (Dev update)
• LangChain deepagents v0.4: Pluggable sandbox backends (Modal/Daytona/Runloop), improved summarization/compression, and OpenAI Responses as default backend — favoring "sandbox as tool, resumable agent" architecture. (deepagents update)
• Coding agent UX: VS Code/Copilot adds worktrees, MCP apps, slash commands; multi-model parallel review (Opus 4.6 + GPT-5.3-Codex + Gemini 3 Pro) emerging as a pattern. GPT-5.3-Codex in VS Code briefly paused at launch.
• Claude Code hidden flag --sdk-url: Turns the local CLI into a WebSocket client for browser/mobile frontends, enabling remote IDE setups and autonomous tmux + sub-agent workflows. (Demo)
• Electric SQL "Configurancy": Argues agents produce maintainable code when configuration and structural constraints come first; proposes rerunnable workflows with budgets and guardrails, closer to CI/CD than one-off scripts. (Blog post)
• LMArena: Now supports PDFs in prompts; launched an Academic Partnerships Program funding projects up to $50,000. (Image/PDF update, Academic program)

Infrastructure & Hardware

• Unsloth: Up to 12x faster MoE training via custom Triton kernels and torch._grouped_mm, ~35% VRAM savings, small-MoE training on <15GB consumer GPUs; also released long-context RL guides. (Faster MoE docs)
• vLLM production lessons (AI21): Throughput doubled via configuration + queue-based autoscaling; a postmortem on a rare ~0.1% garbled-output bug traced to request-parsing timing under memory pressure.
• Distributed inference: Kimi K-2.5 (658GB) run across 4 Thunderbolt-linked Mac Studios via MLX Distributed with reportedly near-linear throughput scaling; ~40 tok/s for Qwen3Next Q4 on an AMD "AI MAX" 395 laptop with 96GB unified memory.
• Nubank is hiring CUDA/kernel engineers to train in-house foundation models on B200s (paper).
• Modular acquires BentoML: Aims for a "write once, run anywhere" inference stack across NVIDIA, AMD, and future accelerators; BentoML stays Apache 2.0. (Announcement)

Research & Methods

• iGRPO: A two-stage GRPO variant — sample drafts, pick the best with unified scoring, then train to beat it — no critic or textual feedback needed; reported to outperform GRPO across 7B/8B/14B model families.
• Self-verification & ConceptLM: Learning to self-verify achieves better reasoning with fewer tokens; ConceptLM quantizes hidden states into a "concept vocabulary" for next-concept prediction with continued-pretraining gains.
• Ganguli: Natural language's conditional entropy decay and token-correlation decay can predict scaling-law exponents in data-constrained regimes. (Thread)
• Architecture leaks from source digging: GLM-5 rumored at ~740B total / ~50B active params with MLA attention and sparse indexing for 200K context; Qwen3.5 speculated as a hybrid SSM-Transformer (Gated DeltaNet linear attention interleaved with full attention) plus MoE with shared experts. (Unconfirmed community findings.)
• Generative Meta-Model: A diffusion model trained on ~1B LLM activations enabling on-manifold steering in activation space; of interest for controllable generation and interpretability. (Project page)
• Model self-reflection: On Llama3.1 and Qwen2.5-32B, models spontaneously invent vocabulary (e.g., "loop," "mirror") describing their own activation patterns, with frequencies correlating to real autocorrelation/spectral features. (DOI)

Products & Applications

• Perplexity Pro: Silent quota cuts (Deep Research to 20/month, uploads to 50/week) triggered user backlash and cancellations.
• Cursor: Composer 1.5 at 50% off, but complaints about auto-switching to Auto, disconnects, slow queues, and opaque billing persist.
• Arena upgrades: Image/Video leaderboards clustered by scenario from 4M user prompts (~15% noise filtered); Video Arena moved from Discord to the web; Veo 3.1 topped Video Arena.
• P402.io: A middleware on OpenRouter tracking per-model costs, suggesting cost-efficient model swaps, with USDC/USDT payments at 1% fees.
• AuditAI (open source): Corrective RAG (LangGraph) auditing security policies against NIST CSF 2.0, with mandatory evidence citations; evaluated with Llama 3.3 70B on Groq. (Code, Demo)

Industry & Business

• Qwen3-Coder-Next is winning local-LLM fans as a capable general-purpose small model despite its name.
• Local LLM economics: Community debates whether a $5K–10K local machine beats cloud subscriptions; consensus: capability still trails closed cloud models, but momentum favors local for privacy-sensitive, stable workloads.
• Salesforce exodus: Multiple executives (Slack, Tableau CEOs, President, CMO) departed to OpenAI and AMD, seen as a vote on future growth and possible AI strategy repositioning.
• Cloudflare passed $2B annual revenue; a developer's $46K Vercel bill for Jmail HTML rendering drew an offer from Vercel's CEO to optimize the architecture and cover costs.
• a16z leads investment in Japan's Shizuku AI Labs, betting on AI companion characters rooted in anime/ACG culture.

Policy, Governance & Safety

• OpenClaw jailbreak concerns: Loosely designed permissions and weak system prompts enable indirect privilege-escape-style attacks; researchers suggest embedding whitelists plus syntactic constraints.
• GPT-5.2 / Opus 4.6 jailbreak arms race: ENI attacks partially succeed on Opus 4.6; "Glossopetrae" virtual-world-language bypasses continue evolving; some groups now sell red-team services.
• Discord ID verification requirements for some channels/features drew developer pushback, with speculation about IPO-related compliance or broader data collection.
• OpenAI tests ads in ChatGPT for select search/information queries, raising concerns about commercial influence on model outputs.
• KOKKI Agent-Auditor Loop: A draft-then-audit architecture (Audit(Draft(input))) with cross-model auditing (e.g., GPT drafts, Claude audits) proposed to systematically reduce hallucinations, with early gains in safety/factuality.

*Source: Easy AI Daily (zhichai.net translation digest).*