Easy AI Daily News — January 13, 2026

A curated digest of AI industry news, model releases, research, tooling, infrastructure, and policy developments.

Industry & Company News

• Apple picks Google Gemini to power next-gen Siri: Apple announced that the next Siri and Apple Foundation Models will be built on Google's Gemini models and cloud tech, combining on-device processing with Private Cloud Compute while emphasizing privacy. Apple also evaluated ChatGPT and xAI Grok. Seen as a major win for Google and a loss for OpenAI, sparking discussion on competition and antitrust implications.
• Coverage | Google statement | MacRumors analysis
• OpenAI acquires Torch, doubles down on health: OpenAI launched ChatGPT Health with a separate memory space and acquired medical startup Torch to structure lab results, medications, and visit recordings into the product. The Torch team joins OpenAI, which is also pursuing HIPAA-compliant medical ChatGPT and APIs.
• ChatGPT Health launch | Torch announcement | torchbio.com
• Anthropic launches Cowork: Extends Claude Code to general office work — Claude can read/write designated local folders, use Chrome, organize files, build spreadsheets, and write reports. Research preview for macOS Claude Max users; community is watching its security model and enterprise deployment costs.
• Official blog | Announcement
• Phind shutting down: The developer search/code assistant announced closure this weekend; the community is discussing alternatives (Perplexity, OpenAI + browser tools, self-built RAG) and reflecting on vendor lock-in.
• China LLM IPO narratives: Zhipu vs MiniMax: A summary argues the divergence in their IPO performance comes down to narrative — Zhipu as ToB/ToG infrastructure with long sales cycles vs MiniMax's consumer-facing, global platform growth story.
• Meta recruits Daniel Gross to lead a new AI infrastructure initiative, working with executives Dina Powell McCormick and Santosh Janardhan.

Models & Capabilities

• DeepSeek releases Engram: A conditional memory module introducing hash n-gram lookup "conditional memory" into Transformers, moving rote memorization from costly forward passes to O(1) lookups. Embedding tables can live in main memory/NVMe with low inference overhead; positioned as a key component for next-gen sparse models.
• GitHub | Paper PDF
• Gemini long-context reality check: Engineers report needle-in-haystack degradation past ~120k tokens despite official 400k claims; Claude Haiku/Sonnet performed more steadily. Community consensus: stress-test long context yourself rather than trust marketing numbers.
• GLM-4.7 rolls out on Cerebras (Hugging Face) and Together AI, touting 200k context and strong coding ability.
• RAM Score metric proposed: Normalizes Hugging Face downloads by time and model size. Findings: 1–9B models get the most downloads, top models differ by only ~4x in median downloads, GPT-OSS is unusually popular, and large Chinese MoE models show weaker momentum.
• Gemini API raises file limits: Single-file cap raised from 20MB to 100MB, with Google Cloud Storage import and signed URL support; up to 2GB registered files in some cases — useful for high-volume RAG.

Research & Methods

• DroPE and test-time training: Sakana's DroPE converges with RoPE during training, then removes positional encodings at inference to reduce long-context distortion. NVIDIA/Stanford/Astera's TTT-E2E performs incremental training on current context at inference, effectively "writing" key info into weights to reduce KV cache pressure.
• Agent memory: AgeMem vs SimpleMem: AgeMem unifies long/short-term memory as tool actions (add/delete/edit, retrieve, summarize) trained via staged RL, improving over Mem0 by up to 13% on Qwen models. SimpleMem uses lossless semantic compression + merge + on-demand retrieval, raising LoCoMo F1 from 34.2 to 43.24 while cutting average query tokens from ~16,900 to 531.
• FASHN Human Parser open-sourced: A SegFormer-B4 fine-tuned model for fashion e-commerce outputting 18 body/garment masks at 384×576, ~300ms on GPU, addressing poor quality in ATR/LIP/iMaterialist datasets.
• PyPI | Hugging Face
• Game-theoretic feedback stabilizes LLM agents: Converting interaction logs into graphs, constructing zero-sum games, and feeding Nash equilibrium statistics into system prompts raised a 44-test benchmark success rate from 20% to 42.9%, reduced tool-use variance 5.2x, and cut completion time 2.7x.
• Paper | Code
• LLM-as-judge inconsistency: "Evaluative Fingerprints" finds individual LLMs are self-consistent but cross-model agreement is near zero (Krippendorff's α ≈ 0.042); judges are identifiable at 89.9% accuracy from scoring patterns. Model choice introduces systematic bias in automated evaluation.
• Paper
• CURE-GRPO: A Google Tunix Hackathon writeup combining Self-Critique with GRPO to improve reasoning in Gemma-class models.

Agents & Tooling

• Ramp open-sources its coding agent blueprint (Inspect): Cloud-based agent using open-source tools, Modal, and Cloudflare; reportedly produced ~30% of merged frontend/backend PRs in one week.
• AI21's SWE-bench lessons (200k+ production runs): Decouple generation from evaluation environments — reuse pre-built repos + dependencies + MCP servers per issue, enabling re-runs without regeneration. Cut failure rates from 30% to 0% and repo downloads from 8000+ to 500.
• DVCP (Vibe Coding Protocol): A commander-thread + executor-thread structure forcing whole-file outputs and multi-thread switching to make LLM coding reproducible rather than ad-hoc.
• Longform article
• Agent CLI benchmark: Across 20 web dev tasks, Kiro scored highest (77% success), Gemini CLI lowest (47%). Users caution some tools claim "done" without actually working.
• Full comparison
• PulseFramework: A new lightweight LLM workflow orchestration layer emphasizing easy model switching and step control.

Infrastructure & Hardware

• NVIDIA Blackwell performance claims questioned: GPU MODE community notes the "256×256 op in 11 cycles" result is asynchronous and shouldn't be treated as true latency; profiling must account for async, queuing, and memory access details.
• RTX 3090 restarts under LLM inference: Users traced crashes on Fedora and Windows to GSP firmware; disabling it and undervolting improved stability.
• popcorn-cli v1.2.2 integrates NVIDIA NCU profiling into the CLI with summary display and downloadable .ncu-rep files.
• LM Studio on ARM SBC: Orange Pi 6 Plus (8-core ARM) running Qwen3-4B Q4 at ~6.6 tokens/s on CPU, though the Electron UI suffers severe graphics glitches.
• TinyBox BMC lockouts: Tinygrad community shares recovery steps — reflash BIOS/BMC firmware and reset BMC from the UEFI menu.

Products & Applications

• Google's Universal Commerce Protocol (UCP): Enables Gemini/AI Mode to complete checkouts, chat with brand Business Agents, and deliver Direct Offers — making the model a shopping interface and transaction initiator.
• Claude Cowork desktop: Hands a folder to Claude for renaming, organizing, extracting tables from screenshots into Excel, and generating reports; macOS/Claude Max only for now.
• AI receptionist workflow: A ChatGPT + n8n demo handling calls, scheduling/rescheduling, and SMS notifications, being hardened for production.

Policy, Governance & Safety

• FDA updates clinical trial statistics guidance: Stricter statistical validation, pre-registration, and uncertainty reporting are expected to raise the bar for AI/ML medical products, requiring auditable reasoning logs, versioned data, and prompts.
• FDA announcement
• Env vars are not safe for agents: A local agent can read process environments and extract API keys stored in env vars. Recommendations: use dedicated secret managers (Vault, AWS Secrets Manager) and avoid dumping environments in logs.
• Jailbreak/prompt-injection research continues: BASI community discussions of bypassing commercial bot guardrails underscore that mainstream model safety policies remain actively probed and patched.

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