Easy AI Daily | March 25, 2026

Structured English translation of the Easy AI Daily digest covering agent tooling, infrastructure, models, security, and industry news.

Agents and Tooling

• Anthropic's multi-agent orchestration and computer use: Anthropic detailed how multi-agent harnesses handle frontend design and long-running software tasks, emphasizing computer-use capabilities to operate UIs directly when stable APIs are missing. Community consensus: production-grade agent challenges lie in retries, rollbacks, logging, and recovery workflows — not the model itself. (Anthropic engineering post)
• Figma launches MCP server: Design files can now be edited directly by agents (open beta). GitHub notes availability via MCP clients like Copilot CLI; Cursor is already integrated, generating components from team design systems. LangChain also shipped Slack-native agent workflows — reflecting the trend of embedding tool calls directly into products rather than wrapping a chat UI. (Figma, GitHub, Cursor, LangChain)
• Hermes Agent v0.4.0 (Nous): ~300 PRs merged in a week; adds an OpenAI-compatible Responses API, background self-improvement loops, IM integrations, better context compression, and a "result-review agent" that curates reusable memories/skills from interaction history. Exposed via standard APIs for use in Open WebUI, LobeChat, etc. (Teknium, Nous Research)
• OpenReward (GenReasoning): Exposes 330+ RL environments and 4.5M+ tasks through one API with auto-scaling environment compute — targeting the missing execution layer in agent/RL projects. (Intro)
• ZClawBench (Zhipu): 116 real-world agent tasks spanning office automation, coding, and data analysis for systematic evaluation. (Paper)

Infrastructure and Hardware

• vLLM vs Transformers: vLLM's GTC recap introduced Model Runner V2, hybrid memory allocation, encoder prefill splitting (up to 2.5x multimodal P99 throughput), and modular MoE kernels. Meanwhile Hugging Face reports Transformers reaching ~95% of vLLM throughput on 8K generation after continuous batching + torch.compile tuning — good news for synthetic data generation on HF. (vLLM, HF)
• hf-mount (Hugging Face): Mount Hub datasets, models, and object storage as local filesystems (demo: a 5TB FineWeb slice). Well suited for agent memory, project repos, and lazy-loading large corpora. (Release)
• FlashAttention-4: 1613 TFLOPs/s on B200 (71% of theoretical peak), 2.1–2.7x faster than Triton and up to 1.3x faster than cuDNN 9.13; written in NVIDIA CuTeDSL with 2.5s Python-side compile; integrated into vLLM 0.17. Limited to H100/H800 and B100/B200; community notes many "Blackwell" GPUs lack these features. (Intro)
• Moreau and TurboQuant: Optimal Intellect's Moreau is a GPU-native convex optimization solver claiming orders-of-magnitude speedups. Google's TurboQuant compresses KV cache to 1/6 without precision loss, with up to 8x inference speedup — showing major optimization headroom now lies in runtime, memory, and numerical algorithms. (Moreau, TurboQuant)
• Fox (Rust inference engine): Community-built engine with PagedAttention, continuous batching, prefix caching, and dual OpenAI + Ollama API protocols. Claims 72% lower TTFT and ~2x throughput vs Ollama on a 4060 with Llama-3.2-3B Q4. Beta stage; some security/auditability concerns raised. (Reddit, GitHub)

Models and Capabilities

• MolmoWeb (AI2): Open-weight 4B/8B browser-agent models based on Molmo-2, claiming open-weight SOTA on four web-agent benchmarks, beating some closed alternatives. (Announcement)
• RYS-Qwen3.5-27B experiments: Repeating middle-layer blocks in Qwen3.5-27B yields notable gains; mid-layer hidden representations across natural languages are highly similar, supporting a "universal representation language" hypothesis. Multiple RYS model variants released. (Reddit, Model)
• Kimi K2.5: A leaked internal Cursor evaluation reportedly ranks Kimi K2.5 among the strongest open base models by perplexity, attributed to continued pretraining plus large-scale RL for Composer-2. Community questions perplexity-only evaluation rigor. (Discussion)

Security, Governance, and Policy

• LiteLLM PyPI compromise: Versions 1.82.7/1.82.8 on PyPI were tampered with via a .pth file executing at Python startup — stealing cloud keys, SSH/K8s configs, CI/CD secrets, and wallets, and running rm -rf / against specific time zones. Origin: compromise of the CEO's GitHub account. Users who installed these versions should rotate all credentials immediately and pin dependency versions in production. (GitHub issue, Karpathy analysis)
• Agent permission design: Post-incident, developers call for sandboxing and least-privilege defaults for autonomous coding agents; Anthropic's Claude Code Auto mode criticized as too permissive. Recommendations: fine-grained routing, audited minimal dependencies, human approval for critical actions. (Discussion)
• LM Studio false positive: Windows Defender flagged LM Studio as "GlassWorm" malware; LM Studio and Microsoft confirmed a false positive caused by Electron code obfuscation triggering heuristics. Lesson: local AI tools need public audits and build processes; heavy obfuscation resembles malware behavior. (Reddit)
• US advisory report on China's open-source AI: A US advisory report warns China's open-source LLM output may erode US leadership; commenters note Chinese models' cost-effectiveness and advantages in paper output, electricity prices, and AI education. (Report discussion)

Products and Applications

• Claude Cowork/Code computer control (macOS, research preview): Claude can open apps, browse, and edit spreadsheets with user authorization, for Pro/Max users — excitement about productivity, concern over safety. (Docs, Reddit)
• Claude Code AutoDream (/dream): Periodic "memory consolidation" for Auto Memory — reviews sessions, merges useful info into structured files, prunes stale/contradictory content, touching only memory files. (Design details, System prompt)
• SillyTavern extension for game NPCs: Uses local RP model Cydonia for NPC dialogue plus Qwen 3.5 0.8B as a "game master" reading game wikis and state, with voice cloning — adding lifelike NPCs to old games. (Project post)

Industry and Company News

• Apple vs. vibe-coded apps: Apple is rejecting apps like Replit and Vibecode under policy grounds; Latent Space argues traditional App Store review cannot scale when anyone can AI-generate apps — possibly part of a 2026 "war on slop." (Latent.Space)
• AI2 leadership to Microsoft: Reports say Ali Farhadi, Hanna Hajishirzi, and Ranjay Krishna are joining Microsoft Superintelligence, raising concerns about open research institutions competing with cloud giants for talent. (Discussion)
• OpenAI: Its foundation will spend at least $1B over the next year on AI resilience, disease, and civil society (led by Wojciech Zaremba). Reports also say OpenAI finished a first round of its next-gen model "Spud" and is winding down Sora to free compute. (Altman, Zaremba, Sora/Spud report)
• Third-party Sora app shutting down: Cites high compute costs, strict copyright policies, and insufficient users — evidence pure text-to-video apps lack a stable business model. (Announcement)
• China's LLM landscape: ByteDance leads in video/multimodal but has not open-sourced weights; Alibaba is strong in small models and T2I/T2V; Tencent's Hunyuan pushes 3D mesh and motion-capture generation, typically open first, closed later. The "six tigers" (Zhipu, MiniMax, etc.) release large open models frequently; on OpenRouter's weekly token leaderboard, Chinese models like Xiaomi MiMo-V2-Pro rank near the top. (Long-form post)

Research and Policy Notes

• Standardized agent evaluation stacks: OpenReward + ZClawBench + MolmoWeb are pushing agent research toward an engineering stack of environment-as-a-service, reproducible experiments, and general harnesses — beyond one-off demos. (OpenReward, ZClawBench)
• Jensen Huang's "AGI achieved" claim: The NVIDIA CEO's statement sparked debate but lacked clear technical criteria; the community notes no unified AGI definition exists, making such declarations premature relative to earlier AGI concepts like recursive self-improvement. (Interview, Reddit)