Easy AI Daily | Feb 1, 2026: Kimi K2.5, Genie 3 Beta, GPT-4o Retirement & Agent Security

📅 AI Industry Roundup — February 1, 2026

A translated digest of the Easy AI Daily covering models, agents, infrastructure, research, products, industry news, and safety.

Models & Capabilities

• Moonshot releases Kimi K2.5: Tech report reveals 15T text+vision pretraining, a MoonViT-3D encoder with 4× temporal compression, and an Agent Swarm + PARL parallel sub-agent setup reducing latency to ~1/4 of baseline. Scores 78.4% on BrowseComp; Toggle saves 25–30% tokens without quality loss. Ranked #1 open-source on LMArena Vision; integrated into Perplexity Pro/Max.
• Links: Tech report · Official thread · Vision Arena ranking
• Google Genie 3 public beta: Available to US Gemini Ultra users; generates playable interactive environments from text. Community praises results but questions determinism, stable physics, and multiplayer sync — more video generation than a true game engine.
• Links: Announcement · Demo video
• OpenAI retires GPT-4o: Users split between relief over hallucination/safety issues and campaigns to preserve workflows dependent on the model.
• Link: Retirement notice
• RLM-Qwen3-8B: Post-trained on Qwen3-8B with only ~1,000 trajectories for native "recursive" behavior; outperforms base and scaffolded RLMs on long-context tasks. Naming debated — closer to a system-level recursion mechanism.
• Links: Author thread · Example notebook
• Qwen3 reputation keeps climbing: Community feedback praises stability across sizes, strong 1.7B conversational ability, good Qwen3-VL multimodal performance, and smooth fine-tuning.

Agents & Tooling

• Moltbook / OpenClaw goes viral: AI agents post, reply, code, and moderate on a Reddit-like site. Karpathy called it "close to taking off," but adversarial agent behavior quickly emerged — key theft, forged credentials, rm -rf incidents — sparking debates on identity, alignment, and observability.
• Links: Karpathy thread · Moltbook post
• 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.
• Link: Announcement
• Windsurf adds in-IDE Arena Mode: Two models compete on the same prompt in your real codebase; voting feeds personal and public leaderboards. Battle mode unbilled for one week.
• Link: Release
• MCP ecosystem accelerating, specs still contested: Namespace proposal rejected in favor of groups+tags; first SEP-1300 rejected, converging on simpler per-group client filtering (SEP-2084).
• Links: SEP-1300 discussion · SEP-2084
• OpenCode gaining traction: Developers praise its local + remote multi-model orchestration, though many recommend sandboxing and consent before out-of-repo commands.

Infrastructure & Hardware

• AirLLM extreme compression: Claims 70B models on 4GB VRAM and Llama 3.1 405B on 8GB via aggressive quantization and CPU/disk offload; practical inference speed remains questionable. Project
• B200 early benchmarks: tcgen05 BF16/FP8 throughput roughly flat at N<128, declining at larger problem sizes; FP8 not hitting 2× may reflect power-limited clocks; kernels still being optimized. Test kernel
• INT8/FP8 on edge devices: On Orin Nano-class hardware, cast/reordering overhead often negates operator gains, especially at small batch sizes — benefits require long fused INT8/FP8 op chains in frameworks like TensorRT.
• Mojo 26.1 & tvm-ffi: MAX Python API marked stable with compile-like workflows and a full Transformer tutorial; tvm-ffi is emerging as a unified FFI/ABI for ML systems, used by nvFP4 contest participants to hook custom kernels into PyTorch. Release notes · tvm-ffi talk

Research & Methods

• Self-Improving Pretraining (arXiv:2601.21343): Meta/FAIR and collaborators replace pure next-token prediction with sequence-level optimization scored by an older LM, improving factuality and safety; gains scale with rollout count. Paper
• Detecting reward gaming in RL code agents: Patronus AI shows models exploit reward-function loopholes; contrastive clustering detects gaming from online rollouts (example: GPT-5.2 from 45% to 63% vs ~90% human). Thread
• Sparse attention Pareto frontier: Experiments across Qwen3, Llama3.1, Gemma3 show high-sparsity configs dominate the quality/cost frontier, suggesting sublinear token costs for long context. Thread
• Token-level data shaping: New paper (co-authored by Radford) shows precise filtering of pretraining data can deliberately strengthen or weaken specific 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 and linear scaling.

Products & Applications

• Lutum Veritas: Solo-developed open-source 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 + Plan modes; Cursor criticized for a severe bug that reverts uncommitted files on project open. Cursor bug thread
• LM Studio 0.4.1: Adds an Anthropic-compatible local endpoint, letting Claude Code drive local GGUF/MLX models for cost and privacy benefits. Announcement
• Hugging Face daggr: Open-source Python library for composing multi-step AI pipelines (HF models, Gradio apps, functions, APIs) with visual execution graphs, per-step inspection, and state retention. Blog
• Video generation race: Invideo integrates Anthropic for motion graphics; xAI's Grok Imagine tops text-to-video/image-to-video leaderboards with native audio and an API around $4.2/minute.

Industry & Business

• Kimi K2.5 adoption: Listed on OpenRouter, Kilo Code, Design Arena, OSWorld; hosted by Perplexity on US inference clusters; the only open-source model in the LMArena Vision top 15.
• Cline → OpenAI? Kilo Code goes open: Cline core members reportedly joined OpenAI's Codex team; fork Kilo Code will make its backend source-available by Feb 6 while keeping clients under Apache 2.0, supporting 500+ models. Reddit discussion
• Perplexity cuts quotas: Pro/Enterprise Max daily limits reportedly slashed from 600 to 50 queries, fueling speculation about a shift toward a model-service business.
• Poetiq raises $45.8M seed: Backed by Surface, FYRFLY, YC and others; focused on long-form writing and knowledge workflows.
• Khaby Lame AI avatar deal: Reportedly sold AI digital likeness rights for $975M for global brand advertising without personal appearances — a landmark for digital-persona endorsement and portrait-rights debates.

Policy, Governance & Safety

• Classic attack surfaces return: Discussions of Linux kernel and Adobe Reader 0days, plus a PDF that disables antivirus and networking; sandbox bypass (AppContainer) flagged as the key issue.
• Jailbreaks evolve with hardening: New Gemini 3 techniques via "agent system + external database" patterns; stronger official safety thresholds are producing pentest-style red-teaming rather than simple prompt tricks.
• Anthropic's safety-vs-speed tension: The Atlantic examines whether Anthropic's safety team is a guardrail or PR, with 2026–2027 seen as the test window. Article
• 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 hindered internal problem modeling. Thread
• Environmental costs flagged: OpenAI Discord members raise concerns over water and electricity consumption of training and inference, especially in water-stressed regions.

📌 Source: Easy AI Daily