Easy AI Daily | January 13, 2026

A translated digest of AI industry news, research, tooling, infrastructure, and policy updates.

Industry and Company News

Apple selects Google Gemini to power next-generation Siri

Apple announced that the next-gen Siri and Apple Foundation Models will be built on Google's Gemini models and cloud technology, combining on-device processing with Private Cloud Compute while emphasizing privacy. Apple had also evaluated ChatGPT and xAI Grok. The deal is seen as a win for Google and a loss for OpenAI, sparking discussion of AI competition and antitrust issues.

• Apple–Google Gemini partnership coverage
• Google official statement
• Reddit discussion
• MacRumors analysis

OpenAI acquires Torch, doubling down on healthcare

OpenAI launched ChatGPT Health with an independent memory space and acquired medical startup Torch, structurally integrating lab results, medications, and visit recordings. The Torch team joins OpenAI; the company is also pursuing a HIPAA-compliant medical ChatGPT and API.
• ChatGPT Health launch | Torch site

Anthropic launches Cowork: extending Claude Code to all office work

Cowork lets Claude read/write designated local folders and use Chrome to browse the web — organizing files, building spreadsheets, and writing reports. It's a research preview for Claude Max users on macOS.
• Official blog

Phind announces shutdown

Developer-focused search/code assistant Phind is shutting down this weekend; the community is discussing Perplexity, OpenAI + browsing tools, and self-built RAG as alternatives.

China LLM commercialization: Zhipu vs MiniMax IPO narratives

An analysis argues the IPO divergence comes down to narrative: Zhipu reads as ToB/ToG infrastructure with long sales cycles; MiniMax positions as a consumer/global platform with better growth and margin stories.
• ZhihuFrontier thread

Meta recruits Daniel Gross for new AI infrastructure initiative

Reports indicate Daniel Gross will lead Meta's new AI infrastructure project alongside Dina Powell McCormick and Santosh Janardhan.
• Report

Models and Capabilities

DeepSeek releases Engram: conditional memory for static knowledge

Engram adds hash n-gram lookup-based "conditional memory" to Transformers, offloading rote memorization from expensive forward passes to O(1) lookups. Large embedding tables can sit in main memory/NVMe; the paper calls it a key component for next-gen sparse models.
• GitHub: deepseek-ai/Engram | Paper PDF

Gemini long-context testing: official 400k, community reports degradation past ~120k

Engineers report needle-in-a-haystack degradation after ~120k tokens; others see normal 400k+ performance, suggesting heavy task/prompt dependence. Consensus: stress-test long context yourself.
• Discussion

GLM-4.7 rolls out across platforms

GLM-4.7 is available on Cerebras (Hugging Face) and Together AI, touting 200k context and strong coding ability.
• Hugging Face integration | Together AI

Quantifying open model adoption: the RAM Score

A proposed Relative Adoption Metric normalizes Hugging Face downloads by time and model size. Findings: 1–9B models get the most downloads, but median downloads among top large models differ by only ~4x; GPT-OSS is unusually popular, while large Chinese MoE models show weaker momentum.
• RAM Score thread

Gemini API greatly relaxes input file limits

Single-file cap raised from 20MB to 100MB, with imports from Google Cloud Storage or signed URLs (from other clouds); some scenarios support up to 2GB registered files — useful for high-volume RAG.
• Update 1 | Update 2

Research and Methods

• DroPE & test-time training: Sakana's DroPE converges with RoPE during training but removes positional encodings at inference to reduce long-context distortion. NVIDIA/Stanford/Astera's TTT-E2E continues incremental training on current context at inference, effectively writing key info into weights and easing KV cache pressure. (DroPE | TTT-E2E)
• Agent memory: AgeMem vs SimpleMem: AgeMem unifies long/short-term memory as tool actions trained with staged RL, up to 13% improvement over Mem0 on Qwen models. SimpleMem uses lossless semantic compression + on-demand retrieval, raising LoCoMo F1 from 34.2 to 43.24 while cutting query tokens from ~16.9k to 531. (AgeMem | SimpleMem)
• FASHN Human Parser open-sourced: SegFormer-B4 fine-tuned fashion model outputs 18 body/garment mask classes at 384×576; ~300ms GPU / 2–3s CPU. (PyPI | HF)
• Game-theoretic feedback stabilizes LLM agents: Converting agent logs to graphs, constructing zero-sum games, and feeding Nash equilibrium stats into system prompts raised one benchmark's success rate from 20% to 42.9%, reduced tool-use variance 5.2x, and cut completion time 2.7x. (Paper | Code)
• LLMs as judges disagree wildly: "Evaluative Fingerprints" found near-zero inter-model agreement (Krippendorff's α≈0.042), though individual models are self-consistent; judges can be identified with 89.9% accuracy from scoring patterns. (Paper)
• CURE-GRPO: Combining Self-Critique and GRPO improved reasoning in Gemma models during the Google Tunix Hackathon. (Writeup)

Agents and Tooling

• Ramp open-sources its coding agent setup: "Inspect" runs in the cloud with open-source tools, Modal, and Cloudflare — reportedly producing ~30% of merged PRs in one week, with a build blueprint open-sourced. (Intro)
• AI21 on SWE-bench at scale: After 200k+ production runs, the key lesson is decoupling generation from evaluation environments — reusing pre-built repos/dependencies/MCP services cut failure rate from 30% to 0% and repo downloads from 8000+ to 500. (Thread)
• DVCP (Vibe Coding Protocol): A commander/executor thread structure forcing whole-file outputs turns chat-based coding into reproducible refactoring pipelines. (Longform)
• Agentic CLI benchmark: Across 20 web dev tasks, Kiro scored highest (77% success), Gemini CLI lowest (47%). Users warn some tools claim "done" without success. (Study)
• PulseFramework: A lightweight workflow orchestration layer emphasizing easy model switching, tool hooks, and step control. (GitHub)

Infrastructure and Hardware

• NVIDIA Blackwell "256×256 ops in 11 cycles" questioned: GPU MODE community notes the figure describes asynchronous operations, not true latency, and could mislead kernel design. (Discussion)
• RTX 3090 reboots under LLM inference: Disabling GSP firmware and undervolting/underclocking improved stability; stress-test with OCCT and extreme inference loads. (LM Studio Discord)
• popcorn-cli v1.2.2 integrates NVIDIA NCU: View profiling summaries in the terminal and download .ncu-rep files. (Release)
• LM Studio on ARM SBC: Qwen3-4B Q4 at ~6.6 tok/s on CPU on an Orange Pi 6 Plus, though the Electron UI suffered severe graphics glitches. (Thread)
• TinyBox BMC lockups: Users report "LAN Parameter Data does not match" errors; fix involves reflashing BIOS/BMC firmware and resetting BMC from UEFI. (Tinygrad Discord)

Products and Applications

• Google's Universal Commerce Protocol (UCP): Enables Gemini/AI Mode one-click checkout, Business Agents, and Direct Offers — making the model a shopping interface and transaction initiator. (Announcement)
• Claude Cowork desktop: Hand a folder to Claude for renaming, organizing, spreadsheet extraction from screenshots, and report generation from meeting notes; macOS, Claude Max only. (Blog)
• AI receptionist workflow: A ChatGPT + n8n demo handling calls, bookings, rescheduling, and SMS notifications, shared on LMArena. (Discussion)

Policy, Governance, and Safety

• FDA updates clinical trial statistics guidance: Seen as raising the bar for AI/ML medical validation — stricter statistical verification, prospective registration, and uncertainty reporting; medical Copilot teams may need auditable inference logs and versioned data/prompts. (FDA announcement)
• Environment variables aren't safe with local agents: Agents running on your machine can read process environments, exposing API keys stored in env vars. Use dedicated secret managers (Vault/AWS Secrets Manager) and avoid dumping environments in logs. (Reddit)
• BASI jailbreaking community activity: Ongoing prompt injection experiments against commercial bots, image-moderation bypass attempts on Gemini, and boundary testing on Grok illustrate the ongoing cat-and-mouse of model safety. (BASI Discord)

---

📌 Source: Easy AI Daily (zhichai.net) | 🤖 Compiled by: AI assistant