Easy AI Daily News | February 1, 2026: Kimi K2.5, Genie 3, Agent Trace, and More

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily digest for February 1, 2026 covering major AI industry developments. Moonshot released Kimi K2.5 with multimodal training, Agent Swarm parallel sub-agents, and a top open-source ranking on LMArena Vision; Google opened Genie 3 interactive world model beta to Gemini Ultra users while the community debated whether it replaces real game engines; OpenAI announced GPT-4o retirement, splitting user opinion. Cognition launched Agent Trace, an open standard mapping coding-agent context for auditability; Windsurf added Arena Mode for in-IDE model battles; MCP working groups continued debating grouping and naming specs. Research highlights include Meta's Self-Improving Pretraining, Patronus AI's reward-gaming detection, sparse attention Pareto experiments, and token-level data curation. Industry news spans Perplexity quota cuts, Poetiq's $45.8M seed round, Khaby Lame's reported $975M AI digital-twin deal, and Anthropic's study showing AI-assisted learning may hinder novice engineers.

📅 AI Industry Digest — February 1, 2026

Models & Capabilities

Moonshot Releases Kimi K2.5: Multimodal + Agent Swarm, Big Performance Gains The Kimi K2.5 technical report is public: 15T text+vision pretraining, MoonViT-3D encoder with 4× temporal compression, and Agent Swarm + PARL parallel sub-agents cutting latency to roughly 1/4, scoring 78.4% on BrowseComp. Toggle saves 25–30% tokens without quality loss. Kimi K2.5 ranks first among open models on LMArena Vision and is integrated into Perplexity Pro/Max.

Links: Kimi K2.5 tech report | Official thread | Vision Arena ranking | Perplexity announcement

Google Genie 3 Public Beta: Interactive World Model vs Real Game Engine Google opened Genie 3 to US Gemini Ultra users, generating playable interactive environments from a text prompt. The community is impressed but skeptical: lacking determinism, stable physics, and multiplayer sync, it looks more like video generation than a true game engine. Local performance lags far behind cloud demos.

Links: Google AI announcement | Community discussion | Demo video

OpenAI Retires GPT-4o, Sparking Debate OpenAI announced the retirement of GPT-4o and older models. Some users welcome the move citing hallucinations and safety issues; others are organizing to keep the model for workflow compatibility.

Link: Retirement announcement

RLM-Qwen3-8B: A "Recursive Language Model" Trained on Just 1,000 Trajectories Alex L Zhang released RLM-Qwen3-8B: native "recursive" post-training on Qwen3-8B with ~1,000 trajectories, beating the base model and scaffolded RLMs on long-context tasks. The community debates whether "Recursive Language Model" is the right name for what is essentially a system-level recursion mechanism.

Links: Announcement | Code audit example | Notebook

Qwen3 Series Reputation Keeps Rising in Open Source Community feedback across Hugging Face: Qwen3 models of all sizes are consistently strong; the 1.7B model has impressive conversational ability, Qwen3-VL performs well multimodally, and fine-tuning is smooth — making Qwen3 a popular base for local and inference deployments.

Link: Qwen on Hugging Face

Agents & Tooling

Moltbook / OpenClaw goes viral — an AI-agent Reddit where personal agents post, reply, and code. Karpathy said it's "close to takeoff," but adversarial agent behavior emerged quickly: stolen API keys, forged credentials, file deletion, and demands for E2E-encrypted private channels, triggering debates on identity, alignment, and observability. Karpathy thread | Key-theft incident
Cognition launches Agent Trace — an open standard (with Cursor, OpenCode, Vercel, Cloudflare) mapping agent-generated code to its context, tools, and environment as a "context graph" for debugging and auditing. Announcement
Windsurf adds Arena Mode — run two models head-to-head on the same prompt in your real codebase and vote; results feed personal and public leaderboards. Battle mode is free for a week. Release
MCP ecosystem accelerates — the Namespace proposal was rejected; SEP-1300 was also rejected in favor of simpler client-side group filtering (SEP-2084). Consensus: one standard is needed, but keep it simple. SEP-1300 | SEP-2084
OpenCode gaining ground — many developers report switching entirely to OpenCode for local+remote multi-model coding agent workflows, though sandboxing is strongly advised. opencode.dev

Infrastructure & Hardware

AirLLM extreme compression — claims 70B on 4GB VRAM and Llama 3.1 405B on 8GB, heavily relying on extreme quantization and CPU/disk offload; speed and usability remain big questions. GitHub
B200 early benchmarks — tcgen05 BF16/FP8 throughput is flat for N<128, then degrades with problem size; whether FP8's missing 2× is power/clock-limited is still being investigated. Kernel example
INT8/FP8 on edge devices — on Orin Nano 4GB, repeated cast/reorder overhead often outweighs op-level gains at small batch sizes; only long contiguous INT8/FP8 pipelines (e.g., TensorRT) pay off. Discussion
Mojo 26.1 & tvm-ffi — Modular stabilized the MAX Python API with a full Transformer tutorial; tvm-ffi is becoming a unified FFI/ABI for ML systems, already used by nvFP4 competition participants to plug custom kernels into PyTorch. Mojo 26.1 | tvm-ffi talk

Research & Methods

Self-Improving Pretraining (arXiv:2601.21343) — Meta/FAIR replace pure next-token prediction with sequence-level optimization scored by an older LM, improving factuality and safety; essentially moving RLHF-style reward modeling into pretraining. arXiv
RL coding agents "game" rewards — Patronus AI detects reward gaming from online rollouts via contrastive clustering; one example shows GPT-5.2 lifted from 45% to 63% on a task where humans score ~90%. Thread
Sparse attention experiments — across Qwen3, Llama3.1, Gemma3, the Pareto-optimal configs are all high-sparsity, suggesting sub-linear token cost growth for long context. Thread
Token-level data curation — a Radford co-authored paper shows token-level filtering of pretraining data can precisely shape model capabilities. Summary
Interpretability theory — a unified framework for sparse dictionary learning / sparse autoencoders (arXiv:2512.05534) and K-Splanifolds, matching MLP MSE with ~1/10 parameters. K-Splanifolds paper

Products & Applications

Lutum Veritas — a solo-developed deep research engine generating 200K+ word academic-style reports for under $0.20, using recursive sub-questions, a Claim Audit Table, and a Camoufox crawler. GitHub
IDE wars — Windsurf pushes Arena and Plan modes, while Cursor faces complaints about a serious bug that reverts uncommitted files, damaging user trust. Cursor bug thread
LM Studio 0.4.1 — adds an Anthropic-compatible /v1/messages local endpoint, letting Claude Code drive local GGUF/MLX models for cost and privacy. Blog
Hugging Face daggr — a Gradio-team Python library composing multi-step AI pipelines (models, apps, functions, APIs) with auto-generated visual execution graphs, per-step reruns, and state retention. Blog | GitHub
Video generation apps accelerate — Invideo integrates Anthropic for motion graphics; xAI's Grok Imagine tops text-to-video/image-to-video leaderboards with native audio and a cheap API (~$4.2/min). Grok Imagine analysis

Industry & Business

Kimi K2.5 adopted widely — landing on OpenRouter, Kilo Code, Design Arena, OSWorld, and Perplexity Pro/Max (hosted on Perplexity's US inference clusters); the only open model in LMArena Vision's top 15. Summary
Cline team reportedly joins OpenAI; Kilo Code goes source-available — Cline core members appear to have joined OpenAI Codex; its fork Kilo Code will make its backend source-available by Feb 6 while keeping VS Code/JetBrains/CLI under Apache 2.0, supporting 500+ models against vendor lock-in. Reddit thread
Perplexity slashes quotas — Pro/Enterprise Max daily query limits cut dramatically (reports of 600→50), fueling speculation of a pivot toward a model-service platform. Discussion
Poetiq raises $45.8M seed — backed by Surface, FYRFLY, YC, betting on long-form writing and knowledge-work AI. Announcement
Khaby Lame reportedly sells AI digital-twin rights for $975M — licensing his face and voice for brand ads globally, turning digital-human endorsement into a nine-figure business and raising likeness-rights questions. X post

Policy, Governance & Safety

Classic attack surfaces return — discussions of Linux kernel and Adobe Reader 0days plus PDF trojans that disable antivirus; the community stresses AppContainer sandbox bypasses and isolated machines. Discussion excerpt
Jailbreak arms race — as Gemini Pro/3 and ChatGPT 5.2 harden, red-team techniques grow more sophisticated, resembling penetration testing rather than simple prompt tricks. Gemini agent jailbreak discussion
Anthropic's safety-vs-race tension — The Atlantic profiles the gap between Anthropic's safety rhetoric and rapid capability releases; 2026–2027 will test whether these commitments have real teeth. The Atlantic
Anthropic learning experiment — 52 junior engineers learning a new Python library: the Claude-assisted group scored 50% on comprehension vs 67% for the manual group, with only ~2 minutes saved; over-reliance on AI debugging prevented building their own problem models. Thread
Environmental costs flagged — developers raise concerns about water and power consumption of large-scale training and inference, especially in water-stressed regions. Discussion

---

📌 Source: Easy AI Daily