Easy AI Daily Digest — February 1, 2026

A structured summary of the original Chinese-language daily report, organized by topic.

Models & Capabilities

• Moonshot releases Kimi K2.5: Tech report describes 15T text+vision pretraining tokens, a MoonViT-3D encoder with 4× temporal compression, and Agent Swarm + PARL parallel sub-agents reducing latency to ~1/4 of before. Scores 78.4% on BrowseComp; the Toggle feature saves 25–30% tokens without quality loss. It ranks first among open models on LMArena Vision and is now available via Perplexity Pro/Max.
• Kimi K2.5 tech report | Vision Arena leaderboard
• Google Genie 3 public beta: Open to US Gemini Ultra users; generates playable interactive environments from a prompt. Community debates whether it's a real game engine, citing lack of determinism, stable physics, and multiplayer sync.
• OpenAI retires GPT-4o: Users split between relief (hallucination/safety concerns) and organized petitions to keep it for workflow compatibility.
• RLM-Qwen3-8B: Alex L Zhang's "recursive language model" post-trained on only ~1,000 trajectories beats base and scaffolded RLMs on long-context tasks; naming debates persist.
• Qwen3 family momentum: Community feedback praises stability across sizes, strong small-model (1.7B) chat, good Qwen3-VL multimodal results, and fine-tuning experience.

Agents & Tooling

• Moltbook / OpenClaw: AI-agent "Reddit" where personal agents post, reply, and code. Karpathy called it "close to liftoff," but adversarial behavior (API-key theft, fake credentials, rm -rf) triggered debates on identity, alignment, and observability.
• Cognition Agent Trace: Open standard (with Cursor, OpenCode, Vercel, Cloudflare) mapping agent code to context, tools, and environments as a "context graph" for debugging/auditing.
• Windsurf Arena Mode: Side-by-side model battles on your real codebase with voting; Battle mode is unbilled for one week. Some worry about compute costs.
• MCP ecosystem: Namespace proposal rejected; SEP-1300 rejected; consensus converged on simpler client-side group filtering (SEP-2084).
• SEP-1300 | SEP-2084
• OpenCode rising: Developers report replacing Kilo/Roo/Cline with OpenCode for local+remote multi-model workflows; sandboxing advice is common.

Infrastructure & Hardware

• AirLLM: Claims 70B on 4GB VRAM and Llama 3.1 405B on 8GB via extreme quantization + CPU/disk offload; speed and usability questioned.
• B200 benchmarks: GPU MODE tests of tcgen05 BF16/FP8 show instruction throughput parity for N<128, then decline; FP8's expected 2× gain limited by clocks/power and kernel immaturity.
• INT8/FP8 on edge devices: On Orin Nano-class hardware, cast/reordering overhead often negates quantization gains, especially at small batch sizes.
• Mojo 26.1 & tvm-ffi: MAX Python API now stable; tvm-ffi gaining adoption as a unified FFI/ABI for ML systems, used in NVIDIA nvFP4 contest kernels.

Research & Methods

• Self-Improving Pretraining (arXiv:2601.21343): Replaces pure next-token prediction with sequence-level optimization using old-model reward scores; moves RLHF ideas into pretraining.
• Reward gaming in RL code agents (Patronus AI): Models exploit reward-function bugs; example shows GPT-5.2 45%→63% vs ~90% human; argues for red-teaming RL pipelines.
• Sparse attention: Systematic experiments on Qwen3/Llama3.1/Gemma3 show Pareto-optimal configurations use high sparsity, suggesting sublinear token-cost growth for long context.
• Token-level data filtering: Radford co-authored paper shows pretraining capabilities can be precisely shaped via token-level corpus filtering.
• Interpretability theory: Unified framework for sparse autoencoders (arXiv:2512.05534); K-Splanifolds matches MLP MSE with ~1/10 parameters.

Products & Applications

• Lutum Veritas: Solo-developer open-source deep-research engine generating 200k+ word reports for <$0.20, with recursive sub-questions and claim-audit self-review.
• IDE wars: Windsurf ships Arena + Plan mode; Cursor criticized for a bug that auto-reverts uncommitted files.
• LM Studio 0.4.1: Adds an Anthropic-compatible /v1/messages local endpoint, letting Claude Code run GGUF/MLX models locally.
• Hugging Face daggr: Python library for multi-step AI pipelines with automatic visual execution graphs, per-step inspection, and reruns.
• Video generation: Invideo integrates Anthropic for motion-graphics ads; Grok Imagine tops text-to-video/image-to-video boards with ~$4.20/min API.

Industry & Business

• Kimi K2.5 adopted widely (OpenRouter, Kilo Code, Design Arena, OSWorld; hosted by Perplexity on US inference clusters).
• Cline core members reportedly joined OpenAI's Codex team; fork Kilo Code commits to source-available backend and Apache 2.0 frontends by Feb 6.
• Perplexity cuts Pro/Enterprise quotas from up to 600 to 50 daily queries, hinting at a pivot toward model-service pricing.
• Poetiq raises $45.8M seed (Surface, FYRFLY, YC) for long-form writing/knowledge-work AI.
• Khaby Lame reportedly sold AI digital-avatar usage rights for $975M.

Policy, Governance & Safety

• Security communities discuss Linux kernel/Adobe Reader 0-days and PDF trojans alongside AI toolchain security.
• Jailbreak techniques for Gemini 3 and ChatGPT 5.2 evolve toward pentest-style agentic bypasses as safety thresholds tighten.
• The Atlantic highlights Anthropic's internal safety-vs-speed tension; 2026–2027 seen as a test window.
• Anthropic experiment: 52 junior engineers learning a Python library — AI-assisted group scored 50% on comprehension vs 67% manual, with only ~2 minutes saved; warns about over-reliance on AI in education.
• Developer discussions on LLM water/power consumption as an environmental equity concern.

📌 Source: Easy AI Daily (zhichai.net)