Easy AI Daily | 2026-02-11

Key points

Models & Capabilities

• Qwen-Image-2.0 (Alibaba): A 7B unified text-to-image and image-editing model with native 2K resolution and ~1K token prompts. High-quality typography, Chinese calligraphy, infographics, and multi-panel comic character consistency. Shrink from 20B to 7B enables faster inference and easier local deployment; the community awaits open weights.
• Official blog | Reddit discussion
• Seedance 2.0 (ByteDance): Went viral as the "Will Smith spaghetti" meme-ender. Community tests show natural motion, improved body/cloth physics, high-quality anime fight scenes and realistic 1v1 basketball. Currently ~15 seconds per clip; users want longer durations.
• Petapixel coverage | Anime fight compilation
• Kimi K2.5 & Agent Swarm (Moonshot): Can dispatch up to ~100 sub-agents and 1,500 tool calls per task, claiming ~4.5x speedup via parallel execution. One user built a full storyboard-to-video pipeline combining K2.5-generated Excel storyboards with Seedance 2.
• Agent Swarm announcement | Baseten TTFT/TPS benchmarks
• Claude Opus 4.6 (Anthropic): claude-opus-4.6-thinking ranks #1 on Arena text and code leaderboards; reported to one-shot complex UI design and long-form writing better than 4.5, though slower and more token-hungry. Safety researchers criticize that Anthropic's high-risk capability threshold relies mainly on internal staff surveys, lacking hard quantitative evaluations.
• Arena leaderboard | RSP safety criticism
• OpenAI Deep Research on GPT-5.2: Underlying model upgraded, adding external data connectors and progress controls—positioning it as a long-running research agent.
• OpenAI announcement
• Gemini 3 Pro checkpoint spotted in A/B testing, while paid users complain about hallucinations and degraded coding vs. Copilot/Cursor; others prefer its personality and citations.
• TestingCatalog scoop
• Underrated open multimodal models: GLM-OCR (document recognition), MiniCPM-o-4.5 (GPT-4o-like on phones), InternS1 (scientific charts, papers) — all free for commercial use.
• Model roundup
• GLM-4.7-Flash-GGUF became one of the most-downloaded models on Unsloth, warming local LLM adoption.
• IsoDDE (DeepMind/Isomorphic Labs): Claims roughly doubled performance over prior structure-prediction baselines (vs AlphaFold 3), with antibody binding-site and affinity predictions potentially exceeding physics simulations. Architecture details remain sparse.
• Technical report

Agents & Tooling

• OpenAI Responses API: Adds server-side context compression, hosted networked containers for running code, and first-class Skills — signaling "research/computer-use agents" as a formal product.
• LangChain deepagents v0.4: Pluggable sandbox backends (Modal/Daytona/Runloop), improved summarization/compression, and OpenAI Responses as default — favoring "sandbox as tool, resumable agents."
• Coding agent UX: VS Code/Copilot adds worktrees, MCP apps, slash commands; multi-model workflows (Opus 4.6 + GPT-5.3-Codex + Gemini 3 Pro) with cross-review sub-agents are becoming standard. GPT-5.3-Codex is seen as token-efficient for large workflows.
• Claude Code hidden flag --sdk-url: Turns the local CLI into a WebSocket client for browser/mobile frontends — an unofficial "remote IDE protocol."
• Electric SQL "Configurancy": Argues agents need configuration-first design, budgeted re-runnable workflows, and guardrails — AI coding as controllable CI/CD rather than one-shot scripts.
• Configurancy post
• LMArena: Now supports PDFs in prompts; new Academic Partnerships Program funds up to $50,000 per project for evaluation research.

Infrastructure & Hardware

• Unsloth 12x MoE training speedup: Custom Triton kernels + torch._grouped_mm, ~35% less VRAM, small MoE trainable on <15GB consumer GPUs; also released long-context RL guides.
• Faster MoE docs
• vLLM production lessons (AI21): Config + queue-based autoscaling doubled burst throughput; a postmortem traced a ~0.1% garbled-output bug to request-classification timing under memory pressure.
• Distributed inference: Kimi K-2.5 (658GB) run across 4 Thunderbolt-linked Mac Studios via MLX Distributed with reportedly linear scaling; ~40 tok/s Qwen3Next Q4 on an AMD H395 "AI MAX" laptop with 96GB unified memory.
• Nubank is hiring CUDA/kernel engineers to train foundation models on B200s (team includes the Liger Kernel author; paper).
• Modular acquires BentoML: Combines BentoML deployment with MAX/Mojo for write-once inference across NVIDIA, AMD, and future accelerators; BentoML stays Apache 2.0.
• Announcement

Research & Methods

• iGRPO: Two-stage RL without a critic — sample drafts, pick the best via unified scoring, then train the model to beat its own best draft. Reported to outperform GRPO across 7B/8B/14B families.
• Self-verification & ConceptLM: Self-check-and-correct achieves better reasoning with fewer tokens; ConceptLM quantizes hidden states into a "concept vocabulary" for next-concept prediction, gaining further points via continued pretraining.
• Ganguli scaling theory: Conditional-entropy decay and token-correlation behavior of natural language can predict scaling-law exponents in data-constrained regimes.
• Architecture leaks from source digging: GLM-5 reportedly ~740B total / ~50B active params with DeepSeek-V3-like MLA attention and sparse indexing for 200K context; Qwen3.5 speculated as a hybrid SSM-Transformer (Gated DeltaNet + full attention, MoE with expert sharing).
• Generative Meta-Model (Radford et al.): A diffusion model trained on 1B LLM activations enables on-manifold steering in activation space — promising for controllable generation and interpretability.
• Project page
• Model self-reflection study: Llama3.1 and Qwen2.5-32B spontaneously invent vocabulary (e.g., "loop," "mirror") for their own activation patterns, correlating with real autocorrelation/spectral features. Paper

Products & Applications

• Perplexity Pro quietly cut Deep Research to 20 runs/month and uploads to 50/week, sparking refund waves and backlash over bot-heavy support.
• Cursor: Composer 1.5 at 50% off, but users complain about forced Auto-model switches, disconnects, slow queues, and opaque Auto billing.
• Arena upgrades: Image/Video leaderboards now cluster 4M prompts into scenario-based rankings (~15% noise removed); Video Arena moved from Discord to web; Veo 3.1 tops Video Arena.
• P402.io: An OpenRouter middleware for automatic model selection and cost accounting (e.g., Opus 4.6 vs Sonnet 4.5 cost-performance), with 1% fees via USDC/USDT.
• AuditAI (open source): Corrective RAG (CRAG) via LangGraph auditing security policies against NIST CSF 2.0, with semantic routing and mandatory evidence citations; evaluated with Llama 3.3 70B + Groq.
• Backend code | Web demo

Industry & Business

• Qwen3-Coder-Next is praised in local-LLM circles as a top all-around small model despite the "Coder" name.
• Local LLM economics: Community debates whether $5k–10k local rigs beat cloud subscriptions long-term; consensus: still behind closed models, but compelling for privacy-sensitive, stable workloads.
• Salesforce exodus: Slack/Tableau CEOs, the president, and CMO have left for OpenAI and AMD — read as a talent vote on AI strategy.
• Cloudflare tops $2B annual revenue; meanwhile a developer's $46K Vercel HTML-rendering bill went viral, with the Vercel CEO offering to cover costs and optimize the architecture.
• a16z leads investment in Japan's Shizuku AI Labs, betting on anime-style companion AI agents.

Policy, Governance & Safety

• OpenClaw indirect jailbreaks: Loose permissions and weak system prompts let agents bypass boundaries "from the outside"; mitigations proposed include embedding whitelists plus grammatical constraints.
• GPT-5.2 / Opus 4.6 jailbreak arms race: ENI attacks still partially succeed on 4.6; "Glossopetrae"-style fictional-world-language bypasses keep evolving; red-team services are being commercialized.
• Discord ID verification requirements draw developer backlash; some suspect pre-IPO compliance, others fear mass data collection.
• OpenAI tests ads in ChatGPT for some search/informational queries, raising concerns about commercial influence on model outputs.
• KOKKI Agent-Auditor Loop (v15.5): Splits drafting and auditing agents (Audit(Draft(input))), with cross-model auditing (e.g., GPT drafts, Claude audits) reported to cut hallucinations more reliably than single-model self-review.

📌 Source: Easy AI Daily