Easy AI Daily | January 13, 2026

A daily roundup of AI industry news, model releases, research, agent tooling, infrastructure notes, and policy updates, curated by the Easy AI community.

Key Points

• 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 processing with Private Cloud Compute while emphasizing privacy. Apple had also evaluated ChatGPT and xAI's Grok. The deal is seen as a win for Google and a loss for OpenAI, sparking antitrust discussion.
• Links: WCNC report, 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, integrating lab results, medications, and visit recordings. The Torch team joins OpenAI; a HIPAA-compliant medical ChatGPT and API are in progress.
• Links: ChatGPT Health launch, Torch site
• Anthropic launches Cowork: A research preview extending Claude Code-style capabilities to all office work—reading/writing local folders, working with Chrome, organizing files, building spreadsheets, and writing reports. Available on macOS for Claude Max users; community discusses its security model and enterprise cost.
• Link: Official blog
• Phind shutting down: The developer search/code assistant announced closure this weekend, prompting discussion of alternatives (Perplexity, OpenAI + browsing, self-hosted RAG) and vendor lock-in risks.
• Zhipu vs MiniMax IPO narratives: Analysis attributes divergent IPO reception to narrative—Zhipu as ToB/ToG infrastructure with long sales cycles, MiniMax as a consumer/global platform with a better growth and margin story.
• Link: ZhihuFrontier thread
• Meta hires Daniel Gross for AI infrastructure: Reports say Gross will lead a new Meta AI infrastructure project alongside Dina Powell McCormick and Santosh Janardhan.
• Link: Report

Models & Capabilities

• DeepSeek Engram: A conditional memory module using hash n-gram lookup tables to offload static-knowledge memorization from expensive forward passes to O(1) lookups, freeing the backbone for reasoning. Embedding tables can live in main memory/NVMe; described as a key component for next-gen sparse models.
• Links: GitHub, Paper PDF
• Gemini long-context in practice: Engineers report needle-in-a-haystack degradation past ~120k tokens despite the official 400k claim, while Claude Haiku/Sonnet held up better; others report 400k+ tasks working fine—results are highly task- and prompt-dependent. Community advice: benchmark yourself, don't trust marketing numbers.
• Link: Discussion
• GLM-4.7 launches on Cerebras (Hugging Face) and Together AI, touting 200k context and strong coding.
• Links: HF integration, Together AI
• RAM Score (Relative Adoption Metric): A proposed normalized Hugging Face download metric. Findings: 1–9B models get the most downloads, but median download gaps among top large models are only ~4x; GPT-OSS is unusually popular; large Chinese MoE models show weaker momentum.
• Link: Analysis thread
• Gemini API file limits raised: Per-file cap up from 20MB to 100MB, with direct Google Cloud Storage import or signed URLs (from other clouds too); up to 2GB registered files in some cases—useful for high-volume RAG.
• Links: Update 1, Update 2

Research & Methods

• DroPE (Sakana): Converge with RoPE during training, then drop positional encodings at inference to reduce long-context distortion. TTT-E2E (NVIDIA/Stanford/Astera): continued incremental training on current context at inference, effectively writing key info into weights and reducing KV cache pressure.
• Links: DroPE paper, TTT-E2E
• Agent memory: AgeMem vs SimpleMem: AgeMem unifies long/short-term memory as tool actions (add/delete/edit, retrieve, summarize) trained with staged RL, up to +13% over Mem0 on Qwen models. SimpleMem uses lossless semantic compression + on-demand retrieval, raising LoCoMo F1 from 34.2 to 43.24 while cutting average query tokens from ~16.9k to 531.
• Links: AgeMem, SimpleMem
• FASHN Human Parser open-sourced: A fashion e-commerce human-parsing model fine-tuned from SegFormer-B4, outputting 18 body/garment masks at 384×576, ~300ms on GPU, addressing poor ATR/LIP/iMaterialist dataset quality.
• Links: PyPI, Hugging Face, Blog
• Game-theoretic feedback stabilizes LLM agents: Converting interaction logs to graphs, constructing a zero-sum attack-defense game, solving for Nash equilibrium, and injecting statistics into system prompts doubled a 44-run benchmark success rate from 20% to 42.9%, reduced tool-use variance 5.2x, and cut completion time 2.7x.
• Links: Paper, Code
• Evaluative Fingerprints: Same-LLM self-ratings are stable, but cross-LLM agreement is near zero (Krippendorff's α≈0.042). The evaluator model can be identified at 89.9% accuracy from scoring patterns—LLM-based evaluation inherently introduces model-specific bias.
• Link: Paper
• CURE-GRPO: A Google Tunix Hackathon writeup combining Self-Critique with GRPO to improve reasoning in Gemma-class models.
• Link: Writeup

Agents & Tooling

• Ramp open-sources its coding agent "Inspect": Runs in the cloud using open-source tools, Modal, and Cloudflare; reportedly produced ~30% of merged frontend/backend PRs in one week. Build blueprints are open-sourced.
• Links: Intro, Build details
• AI21's SWE-bench lessons (200k+ production runs): Decouple generation from the evaluation environment—reuse prebuilt repos, dependencies, and MCP servers per issue so failed unit tests can be re-run without regeneration. Failure rate dropped from 30% to 0%; repo downloads from 8000+ to 500.
• Link: Thread
• DVCP (Vibe Coding Protocol): A commander-thread + executor-thread structure forcing whole-file outputs instead of fragmented patches, turning chat-based coding into a reproducible batch refactoring pipeline.
• Link: Longform
• Agentic CLI benchmark: Across 20 web dev tasks, Kiro led (77% success) and Gemini CLI trailed (47%). Users warn some tools claim "done" without working—try multiple tools.
• Links: Benchmark, Reddit
• PulseFramework: A new lightweight LLM workflow orchestration layer emphasizing model swapping, tool integration, and step control, compared in Cursor circles with Claude Code and Codex.
• Link: GitHub

Infrastructure & Hardware

• NVIDIA Blackwell microbenchmark questioned: GPU MODE members argue the "256×256 op in 11 cycles" result reflects asynchronous operations, not true latency, and could mislead kernel design; profiling must account for async, queuing, and memory access.
• Link: Discussion PR
• RTX 3090 rebooting under LLM inference: LM Studio users traced crashes on Fedora and Windows to GSP firmware; disabling it and undervolting/underclocking improved stability.
• popcorn-cli v1.2.2 integrates NCU: View NVIDIA NCU profiling summaries in the terminal and download .ncu-rep files for shareable kernel performance comparisons.
• Links: Release, Docs
• LM Studio on ARM SBC: On an Orange Pi 6 Plus (8-core ARM, Ubuntu), Qwen3-4B Q4 ran at ~6.6 tok/s on CPU, though the Electron UI suffered severe graphics glitches.
• TinyBox BMC lockouts: Tinygrad users hit "LAN Parameter Data does not match" errors; the fix is reflashing BIOS/BMC firmware and resetting the BMC from the UEFI menu, with config backups.

Products & Applications

• Google Universal Commerce Protocol (UCP): Lets Gemini act as a shopping interface—testing one-click checkout, Business Agents for brand chat, and Direct Offers.
• Link: Announcement
• Claude Cowork for non-technical users: Hand a folder to Claude for renaming, sorting, extracting tables from screenshots into Excel, and generating reports from meeting notes; browser actions included.
• Links: Blog, Reddit
• AI receptionist workflow: A ChatGPT + n8n demo handling calls, bookings/rescheduling, and SMS notifications, shared on LMArena and seeking productionization.

Policy, Governance & Safety

• FDA updates clinical trial statistical guidance: Stricter statistical validation, pre-registration, and uncertainty reporting expected to raise the bar for AI/ML medical tools; teams may need auditable reasoning logs and versioned data/prompts.
• Link: FDA press release
• Env vars aren't safe for agent secrets: Local agents can read process environments and harvest API keys; use secret managers (Vault/AWS Secrets Manager) and avoid dumping environments into logs.
• Link: Reddit
• Jailbreak/prompt-injection research continues: BASI community discussions cover prompt injection against commercial bots, bypassing Gemini image moderation and Grok safety boundaries—showing mainstream safety policies remain under continuous attack and hardening.

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