Easy AI Daily Digest | January 13, 2026: Apple-Google Siri Deal, OpenAI Health Push, Anthropic Cowork

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for January 13, 2026 covers major AI industry moves and research highlights. Apple announced that the next-generation Siri and Apple Foundation Models will run on Google's Gemini models with Private Cloud Compute. OpenAI launched ChatGPT Health and acquired medical startup Torch to push HIPAA-compliant healthcare AI. Anthropic released Cowork, a research preview extending Claude Code-style workflows to general office tasks on macOS. DeepSeek published Engram, a conditional memory module using hash n-gram lookup for static knowledge. Gemini API raised per-file upload limits from 20MB to 100MB, while community testing found long-context degradation beyond ~120k tokens. GLM-4.7 shipped on Cerebras and Together AI. Research items include DroPE for long-context positional encoding, agent memory methods AgeMem and SimpleMem, game-theoretic feedback stabilizing LLM agents, and evidence that LLM judges score inconsistently across models (Krippendorff's alpha ~0.042). On infrastructure, Ramp open-sourced its coding agent that produced ~30% of merged PRs in a week, AI21 shared lessons from 200k+ SWE-bench runs, and Google unveiled the Universal Commerce Protocol for AI-driven shopping. Policy news: the FDA updated clinical trial statistical guidance, raising the bar for AI medical validation.

Key points

Industry & Companies

Apple picks Google Gemini for next-gen Siri: Apple announced the next Siri and Apple Foundation Models will be built on Google's Gemini models and cloud tech, combining on-device and Private Cloud Compute with a continued privacy emphasis. Apple previously evaluated ChatGPT and xAI Grok. Seen as a win for Google, a loss for OpenAI, and it has sparked antitrust discussion. Links: WCNC report, Google statement, MacRumors analysis
OpenAI acquires Torch, doubles down on health: ChatGPT Health launches with a separate memory space; Torch's team joins OpenAI to integrate lab results, medications, and visit recordings, with HIPAA-compliant ChatGPT and API work underway. Links: announcement, Torch site
Anthropic launches Cowork: a research preview letting Claude read/write designated local folders and use Chrome to organize files, build spreadsheets, and draft reports — "Claude Code for non-programmers." Available on macOS for Claude Max users. Links: blog
Phind shutting down: the developer search/code assistant closes this weekend; the community is discussing Perplexity, OpenAI + browsing, and self-hosted RAG as alternatives.
China LLM IPO narratives: Zhipu reads as ToB/ToG infrastructure (long sales cycles, heavy R&D) while MiniMax reads as consumer/global platform with a better growth-margin story — explaining divergent IPO reception. Thread
Meta hires Daniel Gross to lead a new AI infrastructure initiative alongside Dina Powell McCormick and Santosh Janardhan. Report

Models & Capabilities

DeepSeek Engram: a conditional memory module using hash n-gram lookup tables to offload "memorization" of static knowledge from compute-heavy forward passes into O(1) queries; embedding tables can live in main memory/NVMe. GitHub, paper PDF
Gemini long-context reality check: engineers report needle-in-haystack degradation after ~120k tokens despite the official 400k claim; Claude Haiku/Sonnet held steadier. Community consensus: stress-test long context yourself.
GLM-4.7 launches on Cerebras (Hugging Face) and Together AI, touting 200k context and strong coding. HF integration
RAM Score: a normalized Hugging Face download metric shows 1–9B models dominate downloads, top large models differ only ~4x in median downloads, GPT-OSS is unusually popular, and large Chinese MoE models show weaker momentum. Analysis
Gemini API file limits raised from 20MB to 100MB per file, with GCS import and signed URLs; up to 2GB registered files in some cases.

Research & Methods

DroPE (Sakana): converge with RoPE during training, drop positional encoding at inference to reduce long-context distortion. TTT-E2E (NVIDIA/Stanford/Astera): test-time incremental training that "writes" key context into weights, easing KV-cache pressure.
Agent memory: AgeMem unifies memory ops as tool actions with staged RL (up to 13% over Mem0 on Qwen); SimpleMem uses lossless semantic compression, lifting LoCoMo F1 from 34.2 to 43.24 while cutting query tokens from ~16,900 to 531.
FASHN Human Parser open-sourced: SegFormer-B4 fine-tune, 18 body/garment classes, 384×576 input, ~300ms GPU inference, targeting poor ATR/LIP/iMaterialist dataset quality. HF model
Game-theoretic agent stabilization: zero-sum games over interaction logs, Nash equilibrium stats injected as system-prompt control signals; benchmark success rose 20% → 42.9%, tool-use variance shrank 5.2x, completion time 2.7x faster. Paper
Evaluative Fingerprints: LLM judges are self-consistent but nearly uncorrelated across models (Krippendorff's alpha ≈ 0.042); judge models identifiable with 89.9% accuracy — automated eval inherits systematic bias. Paper
CURE-GRPO: self-critique + GRPO recipe from the Google Tunix Hackathon improving reasoning in Gemma-class models.

Agents & Tooling

Ramp open-sources its coding agent (Inspect): cloud-based, built on open-source tools + Modal + Cloudflare, reportedly producing ~30% of merged frontend/backend PRs in a week, with a public blueprint.
AI21's SWE-bench lessons (200k+ runs): decouple generation from the eval environment by reusing prebuilt repo+deps+MCP servers per issue — failure rate dropped from 30% to 0%, repo downloads from 8000+ to 500.
DVCP (Vibe Coding Protocol): commander/executor thread structure forcing whole-file outputs to make chat-driven coding reproducible.
Agentic CLI benchmark: across 20 web-dev tasks, Kiro scored highest (77% success), Gemini CLI lowest (47%); users warn some tools claim "done" without working. Benchmark
PulseFramework: a lightweight orchestration layer positioned between "fully autonomous agents" and manual workflows, compared against Claude Code and Codex. GitHub

Infrastructure & Hardware

NVIDIA Blackwell "11 cycles for 256×256" questioned: GPU MODE community notes the figure reflects asynchronous operations, not true latency, and could mislead kernel design.
RTX 3090 reboots under LLM inference: disabling GSP firmware plus undervolting/underclocking markedly improved stability on Fedora and Windows.
popcorn-cli v1.2.2 integrates NVIDIA NCU profiling into the CLI with downloadable .ncu-rep files. Release
LM Studio on ARM: Orange Pi 6 Plus (8-core) running Qwen3-4B Q4 on CPU at ~6.6 token/s, though the Electron UI suffered severe graphics corruption.
TinyBox BMC lockouts: "LAN Parameter Data does not match" errors resolved by reflashing BIOS/BMC firmware and resetting BMC from the UEFI menu.

Products & Policy

Google Universal Commerce Protocol (UCP): lets Gemini handle checkout, Business Agents, and Direct Offers — making models the shopping interface, not just search. Announcement
AI receptionist workflow: ChatGPT + n8n handling calls, bookings/rescheduling, and SMS notifications (LMArena demo).
FDA updates clinical trial statistical guidance: stricter validation, pre-registration, and uncertainty reporting expected to raise the compliance bar for medical AI (auditable logs, versioned data and prompts).
Env vars are not safe for agent hosts: local agents can read process environments and exfiltrate API keys; use Vault/AWS Secrets Manager and avoid dumping env vars in logs.
Jailbreak/prompt-injection arms race continues: BASI community cases include tricking customer-service bots and bypassing Gemini image moderation, showing mainstream safety policies remain under active attack and patching.