Easy AI Daily | 2026-01-08

A structured roundup of the day's AI industry news, originally published in Chinese on zhichai.net.

Key points

Models & Capabilities

• NousCoder-14B: Nous Research built on Qwen3-14B using the Atropos framework, running RL for 4 days on 48 B200 GPUs. Pass@1 on olympiad programming benchmarks rose 7.08% to 67.87%. RL environments, benchmarks, and training scripts are fully open-sourced. Blog
• DeepSeek-R1 report expanded from 22 to 86 pages, adding evaluation prompts, synthetic data generation, training harness, RL details, and distillation analysis. Community read: gains come from trajectory exploration + verifiable rewards and behavior shaping, not just better data. arXiv
• Open video/image generation: Black Forest Labs released quantized FLUX.2 [dev] 32B (up to 10 reference images, 4MP); LTX-2 claims the top open text-to-video leaderboard spot; fal launched OmniHuman 1.5 720P avatar video and a Qwen-Image-Edit multi-view camera-control LoRA.
• Gemini 3 mixed reception: Similarweb shows Gemini above 20% of AI chat traffic; ChatGPT below 65%; Grok above 3%. Reddit users report Gemini 3 Pro issues: poor multi-turn memory, repetition, refused web searches, and the 1M-token context being unusable in practice.
• FP4 quantization: NVFP4 forward passes run in PyTorch (no kernel fusion yet, TPS currently lower than bf16); NVIDIA promotes MXFP4 on Blackwell. Meaningful speedups await mature kernel paths.
• Fuzzy-Pattern Tsetlin Machine rewrite: 32M+ predictions/sec on MNIST (Ryzen 7950X3D), 10x training and 34x inference speedup at ~98% accuracy, plus character-level Shakespeare generation. Tsetlin.jl

Agents & Tooling

• LangChain DeepAgents “Ralph Mode”: agents refresh context in a loop and persist state to the filesystem for long-running tasks; seen as a lightweight “agent harness” paradigm.
• Cursor context rewrite: dynamic context discovery from files, tools, and history instead of global embed+retrieve — 46.9% token reduction and near-unlimited sessions via conversation transcripts written to disk.
• MCP as integration layer: HuggingChat + HF MCP server for chatting with papers; Claude Code driving a Reachy Mini robot via MCP; official docs on server instructions and mTLS for enterprise readiness.
• Coding-agent shell safety: community allowlists/blocklists for agent shell commands (e.g., blocking git push/reset and publish commands).
• Supertonic: Unsloth-community CLI storing only deltas vs. base models — lossless post-training compression for distributing many task models on one base.
• Local IDE stack: a VS Code variant built for local LLMs with LM Studio/Ollama support and rewritten context management. Repo

Infrastructure & Hardware

• GPU prices rising: TrendForce reports NVIDIA/AMD price hikes from 2026, with RTX 5090 possibly reaching $5,000.
• NVIDIA RTX open-source optimizations: GPU-side sampling, concurrent QKV, MMVQ kernels, faster loading, native MXFP4 on Blackwell. Blog
• CuTeDSL / FlashAttention: warp-specialized GEMM via TMA+MMA on Blackwell; CuTeDSL flex attention in flash-attention is ~30% faster on H100 forward. PR
• AMD ecosystem: Helion compiler stack being brought to ROCm; Iris (Triton-based multi-GPU framework) recruiting interns for multi-GPU programming, RDMA, and low-level comms. Iris
• OpenRouter/Qwen TPS drop: users report slower Qwen3-Next-80B since Dec 28; attributed to routing to the cheapest provider (GMICloud). Users can pin faster providers per-provider latency in the dashboard.

Research & Methods

• Karpathy's nanochat scaling laws: a low-cost pipeline fitting Chinchilla-like exponents on small models, benchmarked with CORE scores vs GPT-2/3 — ~$100 and 4 hours on 8×H100.
• LM Arena trust declining: Teknium called it pay-to-climb; SurgeHQ published “LM Arena is a plague on AI,” urging task-specific reproducible evaluation instead.
• CodeClash: a new iterative, adversarial, long-horizon coding benchmark with a released training set.
• Retrieval vs long context: even with long contexts, massive corpora need sublinear indices; old single-shot RAG is what's outdated — replaced by multi-round search-read-search pipelines like Baleen.

Products & Applications

• ChatGPT Health launched: a dedicated health space that can connect medical records and Apple Health/Peloton data; health conversations excluded from training, encrypted and isolated; waitlist first, not a doctor replacement.
• Contradictory LLM diagnosis results: Nature-family papers report 90% vs 52.1% accuracy for ChatGPT diagnosis — differences driven by task setup and datasets; single-number “doctor-level” claims are risky. 90% paper | 52.1% paper
• Liquid AI × AMD: LFM2-2.6B-Transcript summarizes long meeting transcripts locally on CPU/GPU/NPU with <3GB peak memory.
• Voice AI companions: Tolan hit 200K MAU; Razer teases Project AVA, a 5.5-inch AI companion device, at CES 2026.
• Local music generation: ACE-Step runs on 8GB VRAM via CPU offload + 8-bit quant, generating 4-minute songs in ~20 seconds, quality near Suno without subscriptions.
• Local coding on 16GB VRAM: Dyad + Oobabooga self-hosted setup with zero cloud dependency and API costs.

Industry & Business

• Market share: Gemini exceeded 20% of AI chat traffic in 2025; ChatGPT fell from 80%+ to below 65%; Grok and DeepSeek ~3%. Google leverages ecosystem bundling (Drive, Docs, YouTube, 2TB storage).
• Funding rumors: xAI reportedly raised ~$20B; Anthropic reportedly seeking $10B at a $350B valuation.
• Discord IPO: Bloomberg reports a confidential IPO filing with Goldman Sachs and JPMorgan.
• Open-source landscape: Chinese models (Qwen, DeepSeek) lead HF usage; Stanford NLP calls Qwen the “overwhelming winner” among open-weight models; Korea is producing high-profile open models with government support.

Policy, Governance & Safety

• ChatGPT Health privacy debate: health chats excluded from base-model training but allowed for “service improvement and research,” raising regulatory questions.
• Sora child-safety concerns: reports of cameo misuse involving children's photos prompted calls for stronger abuse detection, age verification, and takedown flows.
• OpenRouter account theft: stolen accounts, changed emails, charged cards — plus revelations that some providers can see real user IPs; community advises disposable cards, 2FA, and checking provider IP/log policies.
• Multilingual jailbreaks: safety guardrails tuned on English weaken in Swahili, Navajo, Russian, and other languages — red-teaming must cover the language long tail.
• LLM medical use caution: accuracy figures are highly task-dependent; real-world deployment needs granular task definitions, thresholds, clinical trials, and post-hoc risk controls.

📌 Source: Easy AI Daily (zhichai.net)