Easy AI Daily | January 15, 2026

A digest of AI industry developments curated by the Easy AI Daily newsletter, covering models, agents, infrastructure, research, products, industry moves, and policy/safety.

Models & Capabilities

• OpenAI released GPT-5.2-Codex, positioned as its strongest "long-horizon" coding model via the Responses API, aimed at large-scale refactoring, bug hunting, and security vulnerability discovery. Cursor and GitHub Copilot integrated it on day one. (OpenAI announcement)
• A team ran GPT-5.2-Codex for a full week in Cursor, producing 3 million lines of Rust implementing HTML parsing, CSS layout, rendering, and a JS VM — simple websites already run. The case became a landmark for long-running agents and underscored the need for mandatory human review loops. (Author thread)
• New multimodal models: Zai open-sourced GLM-Image (hybrid autoregressive + diffusion, strong at text rendering); LTX-2 generates up to 20s of 4K video with audio; Google Veo 3.1 added portrait mode, image-to-video, and 1080p/4K upscaling across Gemini, YouTube, and AI Studio. (GLM-Image on GitHub)
• ERNIE-5.0 entered the LMArena Text Arena Top 10 (score 1460, ranked #8) — the first Chinese model to do so, with notable math and professional-domain scores. (Leaderboard)
• Local LLMs on 16GB VRAM: community consensus says ~14B models are the sweet spot; 30B requires aggressive quantization and CPU offload with worse real-world quality. (Reddit discussion)
• TiinyAI's 30W mini PC claims local 120B inference with 80GB memory; the community is skeptical of bandwidth/pricing but sees value for offline, privacy-critical use. (Discussion)
• Math progress: a math-specialized Gemini reportedly proved a new theorem (arXiv:2601.07222); GPT-5.2 Pro improved an upper bound on the Moser worm problem, verified by INRIA mathematicians.

Agents & Tooling

• LangChain launched LangSmith Agent Builder with a file-system view of agents, memory, triggers, skills/MCP/subagents. Official advice: default to a single agent; split only when hitting context, ownership, or decomposition bottlenecks.
• "Skills" are emerging as a universal plugin layer: Phil Schmid's Agent Skills spec for Antigravity enables reusable skill directories across Gemini CLI, Claude Code, and OpenCode.
• Claude Code's /compact loses context: the server keeps only a summary. The community workaround: dump long messages to local files, keep summaries + file references, and retrieve details via local full-text search. (Reddit thread)

Infrastructure & Hardware

• OpenAI partnered with Cerebras, signaling that inference speed is now a product feature, not just an infra metric. (OpenAI announcement)
• GLM-4.7 provider benchmarks (Artificial Analysis): Cerebras hits ~1445 tok/s (TTFT ~1.6s, ~131k context), while GPU providers like Fireworks and Baseten trade some throughput for 200k context.
• Modal shared self-hosted inference guidance claiming it can beat public APIs on cost, and SemiAnalysis profiled how Modal runs a 20,000-GPU fleet with vLLM/FlashInfer. (Modal guide)
• Low-precision stack: NVIDIA's TransformerEngine FP8 primer, NVFP4 training discussions, and Helion 0.2.10 with flex attention kernels and SM oversubscription. (FP8 primer)
• NVLink 6 skepticism: GPU MODE members want real benchmarks for "72 GPUs as one card" claims, citing B200 instability and NCCL multi-node hangs.

Research & Methods

• DroPE: remove RoPE entirely and fine-tune for better long-context performance.
• Engram (DeepSeek/PKU): hash-based O(1) sparse memory table separating memory from computation.
• Ministral3 report: layer pruning, PCA rotations, online DPO recipes for shrinking small models.
• UniversalRAG: route by modality first, then retrieve at the right granularity (paragraphs, documents, image/video clips), with gains across 10 multimodal benchmarks.
• VPBench: merely changing marker colors in images can dramatically shuffle VLM leaderboard rankings.
• Spectral Sphere Optimizer (SSO): spectral-norm-constrained training compatible with muP; outperforms AdamW/Muon on 1.7B dense and 8B MoE models in Megatron.
• SlopCodeBench: shows code agents struggle with early architecture decisions and cleanup refactoring in multi-stage large projects. (GitHub)
• Context-as-environment (arXiv:2512.24601): teaching models to actively prune and reorganize context slows long-context performance decay.

Products & Applications

• Google open-sourced the Universal Commerce Protocol (UCP) for AI shopping agents — browsing, cart, payments, with Agent2Agent workflows, AP2 payments, and MCP integration. Adoption and long-term maintenance are open questions. (UCP repo)
• Loggr: an offline health journal on Apple Silicon using quantized Qwen2.5-VL (MLX) for handwritten OCR, sub-100ms NLP pipeline; community suggests PaddleOCR or MiMo-VL-7B-RL for messy handwriting.
• Manus × Similarweb backlash: users burned 2,500–5,000 credits in seconds without warning, fueling calls for spending caps and cost previews.
• Code-jp: a free, open-source VS Code fork for local AI coding with Ollama and LM Studio support.

Industry & Business

• Ahmad Al-Dahle, former Meta Llama lead, became Airbnb's CTO — read as a sign that open-model experience is now a top-tier credential.
• Thinking Machines Lab named Soumith Chintala CTO; Barret Zoph, Luke Metz, and Sam Schoenholz returned to OpenAI.
• Diffraqtion raised a $4.2M pre-seed for programmable quantum-optics components targeting optical-layer "inference-optimized" imaging.
• OpenRouter launched awesome-openrouter and openrouter-apps repos to grow its multi-model ecosystem.

Policy, Governance & Safety

• Chutes migrated to TEE-based inference for verifiable privacy; some OpenRouter models (e.g., R1 0528) are temporarily offline during the transition. (Announcement)
• Jailbreak communities target Grok image safety, Gemini 3.0 Pro restrictions, and Llama 3.2 policies; Google AI Studio logs jailbreak data for training, so many payloads expire quickly.
• Copyright concerns resurfaced around LLM extraction research showing models can reproduce novel characters and plots (arXiv:2402.14873), with fears of legal overreach from misinterpretation.

📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant