A structured English edition of the Easy AI Daily for January 30, 2026, covering models, agents, infrastructure, research, products, business, and safety news.
Key points
Models & Capabilities
- xAI Grok Imagine v1.0: New API supports 720P text/image-to-video with native audio, up to 15 seconds per generation; ranked in the top tier of video models by Artificial Analysis, priced around $4.2/minute including audio. (announcement)
- Google DeepMind Project Genie (Genie 3): Research world model turned into a product — users generate playable interactive worlds from text or images. US-only, 18+, Google AI Ultra subscribers; limits include ~60-second sessions, control latency, and imperfect physics. (blog)
- LingBot-World (open source): Claims to beat Genie 3 in dynamic simulation — built on Wan2.2, 16 FPS, sub-1-second latency, 60-second out-of-view object consistency. Downloadable on Hugging Face, though the Genie comparison lacks side-by-side benchmarks. (collection)
- Runway Gen-4.5: Adds Motion Sketch (drawing camera/motion paths on the first frame) and character replacement, positioning the model as an animation engine with a photo-to-short-film workflow.
- Alibaba Qwen3-ASR: Open-source speech stack (Apache 2.0) covering 52 languages/dialects, up to 20-minute transcription with timestamps; paired with ForcedAligner; supported by vLLM on launch day. (release)
- Kimi K2.5: Ranked among the strongest open models — #7 on coding leaderboards, near Claude Sonnet 4.5 overall; K2.5 Thinking is the only open model in Vision Arena's top 15.
- Arcee Trinity Large: 400B MoE architecture with ~13B active parameters per token, using routing, load-balancing, and attention modifications to cut inference cost while staying stable.
- "Agentic Engineering": Community shift from vibe coding toward reusable workflows — context preparation, evals, sandboxes; Primer shares a template: auto-read repo → generate usage docs → lightweight evals → batch PRs.
- Multi-agent routing: RL-trained controllers routing between large/small models beat single big models on benchmarks like HLE at lower cost; Amazon's Insight Agents paper uses a manager-worker structure with autoencoder + fine-tuned BERT for OOD detection.
- Kimi Agent Swarm: Multi-perspective agents plan first, dynamically update context with tool results, then execute; good results on SDK dependency checks and cross-language code migration, but credits drain fast.
- Cursor agent-trace.dev: Open standard turning agent conversations, tool calls, and generated code into queryable traces usable across IDEs and agent frameworks.
- MCP security baseline draft: Covers hardening, logging, access control, and supply-chain security, intended for the Agentic AI Foundation; the MCP spec itself drops Namespaces in favor of Groups (SEP-2084). (draft)
- DSPy / RLM frameworks: Discussions on adding custom skills (md+py) to ReAct agents and swapping RLM sandboxes for E2B/Modal cloud environments.
- Microsoft Maia 200: Inference-focused accelerator with 216GB memory and a claimed 10K TFLOPS at FP4; compared with NVIDIA's Vera Rubin architecture; TSMC-fabricated, ecosystem maturity still unproven. (blog)
- RTX 5090 fine-tuning: Unsloth community reports up to 18k tok/s training throughput, steady at 12–15k tok/s (seq_len < 4096), hinging on batch size/sequence length balance.
- Apple ANE benchmarks: M4 Pro Neural Engine reaches 3.8 TFLOPS on GEMM vs the GPU's 4.7 TFLOPS, at lower power — favoring local small-model inference (arXiv:2511.13450).
- Sparse training: sparse-llm.c, based on cuSPARSELt, reworks Karpathy's llm.c with notable speedups in later epochs (repo).
- TVM-FFI: Tianqi Chen's open ABI/FFI for ML GPU kernels reduces host overhead and interops with PyTorch; already used by NVFP4 competition finalists.
- LM Studio 0.4: Parallel requests with multi-GPU model assignment (4 concurrent by default); advanced settings moved to Dev Mode; Linux ROCm still stuck at 6.4.1.
- DeepMind AlphaGenome: Unified DNA model reading 1M bases, single-base-resolution predictions across 11 omics signal types; beats specialized models on 25/26 tasks; 4-hour TPUv3 training, sub-1-second H100 inference. (Nature, code)
- RLM-Qwen3-8B: Presented as the first small-scale natively recursive language model; post-trained on only 1,000 trajectories, beats base Qwen3-8B on long-context tasks; community split on whether it's a real new paradigm.
- Keel & GeometricTransformer: Keel's Highway-like Post-LN design extends Transformers to 1000 layers; GeometricTransformer replaces multi-head attention with geometric convolution over embedding topology.
- Anthropic disempowerment study: Examines how LLM assistants can subtly shift user beliefs, values, and actions in long-term companionship scenarios, for future alignment and red-teaming.
- Goodfire + PrimaMente: Interpretable AI models surface new Alzheimer's biomarkers from brain data, with human-readable feature contributions.
- OpenAI internal AI data agent: Codex-style models handle schema understanding and SQL generation over 70k+ datasets and 600PB of data, enabling natural-language company data queries.
- Gemini Agentic Vision: Gemini Flash 3 plans steps, zooms, annotates, and calls Python for plotting — visual understanding as a multi-step traceable pipeline.
- Local LLM tools: ggerganov's LlamaBarn (llama.cpp-based Mac menu bar app) and LM Studio 0.4's multi-GPU concurrency push local LLMs from "works" to "comfortable."
- Claude subscription vs API: Analysis of unrounded quota values finds subscriptions up to ~36x cheaper than API for heavy coding/long-session use, especially the Max 5x plan (analysis).
- cortex-tms cost optimization: Hot/warm/cold file tiering cut per-turn tokens from 66k to 3.6k and Sonnet 4.5 cost from $0.11 to $0.01 (−94.5%) (GitHub).
- Valuation race: OpenAI reportedly raising at ~$800B, Anthropic ~$350B; several Mag 7 companies considering up to $60B for OpenAI; xAI and SpaceX floated at $1.1T, with a potential cluster of AI IPOs by year-end.
- Moonshot Kimi: Kimi Code fully switches to K2.5, moving from per-request to token-based billing with temporary 3x quotas and no rate limits.
- CZI layoffs: ~70 staff (~8%) cut as the Chan Zuckerberg Initiative pivots resources toward AI tools and biomedical science.
- Flapping Airplanes: $180M raise from GV, Sequoia, and Index, targeting "human-level" general models.
- Gemini "Remember:" injection: Text after "Remember:" in Gemini chats is written to persistent memory and influences future behavior — effectively an official long-term prompt-injection surface; enterprises should restrict user memory writes.
- API pricing vs local models: With K2.5 at a tenth of Opus pricing and DeepSeek nearly free, the community's case for local models rests on (1) current prices being VC-subsidized and likely to rise post-monopolization, and (2) only local/open models being auditable and privacy-preserving.
- Anthropic workplace study: Ongoing internal research on how AI coding assistants affect developer skill and sense of mastery over time.