Easy AI Daily | 2026-03-25

A roundup of the day's AI industry news, organized by topic. Original source: Easy AI Daily.

Key points

• Anthropic detailed multi-agent orchestration for frontend design and long-running software tasks, plus "computer use" — directly operating UIs in real software lacking stable APIs. Community consensus: production-grade agents are hard because of retry, rollback, logging, and recovery — not the models themselves. (link)
• Figma launched an MCP server with in-canvas AI direct editing (open beta). GitHub confirmed it works via MCP clients like Copilot CLI; Cursor has integrated it. LangChain also shipped Slack-native agent workflows — a 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 OpenAI-compatible Responses API, background self-improvement loops, IM integrations, context compression, and a "result-review agent" that distills reusable memories/skills from history. Exposed via standard APIs for use in Open WebUI, LobeChat, etc. (Teknium | NousResearch)
• GenReasoning OpenReward: 330+ RL environments and 4.5M+ unique tasks exposed via a single API, with auto-scaling environment compute. (link)
• Zhipu ZClawBench: 116 real-world agent tasks covering office automation, coding, and data analysis for systematic agent evaluation. (paper)

Infrastructure & Hardware

• vLLM (GTC recap): Model Runner V2, hybrid memory allocation, encoder prefill splitting (up to 2.5x multimodal P99 throughput), modular MoE kernels. Hugging Face reports Transformers with continuous batching + torch.compile tuning now reaches ~95% of vLLM's throughput on 8K generation. (vLLM | HF)
• hf-mount (Hugging Face): mount Hub datasets/models/object storage as a local filesystem (demo: 5TB FineWeb slice). Well-suited as infrastructure for agent memory, project repos, and lazy-loading large corpora. (link)
• FlashAttention-4: 1613 TFLOPs/s on B200 (71% of theoretical peak), 2.1–2.7x faster than Triton, 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, B100/B200 and a few others. (link)
• Moreau (Optimal Intellect): GPU-native convex optimization solver, claimed orders of magnitude faster than existing tools. TurboQuant (Google): KV cache compression to 1/6 memory with up to 8x inference speedup at no accuracy loss (paper). Both signal that big wins now live in runtime, memory, and numerics rather than bigger models. (Moreau)
• Fox: a Rust local inference engine with PagedAttention, continuous batching, and prefix caching; on a 4060 + Llama-3.2-3B Q4 setup, TTFT dropped 72% and throughput doubled vs Ollama (claims). Beta, Docker images available, OpenAI + Ollama compatible APIs — though security/audit questions were raised. (Reddit | GitHub)

Models & Capabilities

• MolmoWeb (AI2): open-weight 4B/8B browser-agent models based on Molmo-2, claimed open-weight SOTA on four web-agent benchmarks, beating some closed solutions. (link)
• RYS-Qwen3.5-27B: community experiments repeating mid-layer blocks yield notable gains; mid-layer latent representations across natural languages are highly similar, supporting a "universal representation language" hypothesis. Multiple RYS model variants released. (Reddit | HF model)
• Kimi K2.5: a leaked internal Cursor evaluation reportedly ranks it the strongest open-source base model by perplexity; community disputes using perplexity alone and the claimed "large-scale optimization." (discussion)

Security, Governance & Policy

• LiteLLM PyPI supply-chain attack: versions 1.82.7/1.82.8 were maliciously modified via a .pth file executing at Python startup — stealing cloud keys, SSH/K8s configs, CI/CD secrets, and wallets, and running rm -rf / for specific timezones. Origin: the CEO's GitHub account was compromised. Users of these versions should rotate all credentials immediately and pin dependency versions in production. (GitHub issue | Karpathy analysis | Reddit 1 | Reddit 2)
• Agent permission design: post-incident, developers call for stronger sandboxing, least-privilege defaults, and slimmer audited dependency trees for autonomous coding agents. Anthropic's Claude Code Auto mode was criticized as too permissive despite productivity gains; human approval for critical operations is advised. (discussion | Auto mode)
• LM Studio false positive: Windows Defender flagged it as GlassWorm; LM Studio and Microsoft confirmed a false positive caused by Electron code obfuscation tripping heuristics. Lesson: local AI tools should publish security audits and build processes; heavy obfuscation resembles malware behavior. (discussion)
• US advisory report: warns China's open-source LLM momentum threatens US AI leadership; commenters note Chinese models' price-performance and US labs' weak open-source commitment. (news)

Products & Applications

• Claude Cowork / Claude Code computer use (research preview, macOS, Pro/Max only): Claude can open apps, browse the web, and edit spreadsheets with authorization, preferring connected services (Slack, calendar) before direct screen control. Excitement tempered by safety and misoperation concerns. (docs | Reddit)
• Claude Code AutoDream (/dream): periodic "memory consolidation" for Auto Memory — reviews sessions, merges useful info into structured files, prunes stale/contradictory content; touches memory files only, like GC + index rebuild for dev agents. (design | prompt)
• SillyTavern extension: adds lifelike NPCs to old games using local RP model Cydonia for dialogue and Qwen 3.5 0.8B as a "game master" reading game wikis and state, with voice cloning from original game audio. (post)

Industry & Company News

• Apple vs. "vibecode" apps: Latent.Space argues human App Store review can't scale when anyone can AI-generate apps; Apple has started rejecting Replit and Vibecode-style tools — part of a 2026 "war on slop" that may challenge Apple's distribution dominance. (article | Replit)
• Microsoft hires AI2 leadership: reports say Ali Farhadi, Hanna Hajishirzi, and Ranjay Krishna are joining Microsoft Superintelligence, raising concerns about open research institutes competing with cloud giants. (1 | 2)
• OpenAI: its foundation will spend at least $1B over the next year on AI resilience, disease, and civic organizations (Wojciech Zaremba leading). Reports also say OpenAI finished round one of next-gen LLM "Spud" and is winding down Sora app/API to free compute. (Altman | Zaremba | Sora/Spud)
• Third-party Sora app shutting down: citing high compute costs, strict copyright policy, and insufficient users — evidence text-to-video apps lack a stable business model. (1 | 2)
• Chinese LLM landscape: ByteDance leads in video/multimodal but publishes no open weights; Alibaba is strong in small models and T2I/T2V; Tencent's Hunyuan pushes 3D mesh and motion-capture generation (open first, closed later); the "six tigers" (Zhipu, MiniMax, etc.) release large open models regularly. On OpenRouter's weekly token leaderboard, Chinese models like Xiaomi MiMo-V2-Pro rank near the top. (post)

Research & Methods

• Standardized agent evaluation stacks: OpenReward (unified RL environment API) + ZClawBench (116 real-world tasks) + MolmoWeb are pushing agent research toward an engineering stack of environment-as-a-service, reproducible experiments, and general harnesses — instead of one-off flashy demos. (OpenReward | ZClawBench)

Policy & Governance

• Jensen Huang: "AGI achieved?": NVIDIA's CEO claimed AGI has been realized in an interview without clear technical criteria, igniting debate. The community notes there's still no agreed AGI definition, and declaring victory via "human-level language + common sense" doesn't map to earlier visions of recursive self-improvement or the technological singularity. (video | Reddit)

📌 Source: Easy AI Daily