Easy AI Daily | January 20, 2026

A roundup of AI industry news for January 20, 2026, covering models, agents, infrastructure, research, products, business, and safety.

Models and Capabilities

• STEM (Carnegie Mellon + Meta): Replaces ~1/3 of Transformer FFN up-projection with per-token embedding lookup tables while keeping dense gate/down-projection. Static lookup avoids MoE routing instability; parameters can be async-prefetched on CPU, increasing capacity with nearly unchanged per-token FLOPs. Thread
• RePo (Sakana AI): Context Re-Positioning dynamically adjusts token positions based on content relevance, pulling important distant content closer and pushing noise away — targeting long-context and noisy structured data. Release
• GLM-4.7-Flash (Zhipu): ~30B A3B MoE with MLA compressing KV cache, supporting 200K context at equal memory. Positioned as a local coding/agent model; strong on SWE-bench Verified. Release | Hugging Face
• Gemini 3 Pro model card leaked briefly: 1M token context, multimodal input (text/image/audio/video), 64K output, knowledge cutoff January 2025. Archive
• PersonaPlex-7B (NVIDIA): Small model focused on multi-persona, role-driven dialogue. Model
• VibeVoice (Microsoft, open source): Real-time TTS with ~300ms first-packet latency, multi-speaker, up to 90 minutes of stable speech. 7.5Hz semantic+acoustic tokens, LM for structure, diffusion head for detail. MIT license, research-only flag. Intro
• Engram (DeepSeek): Hash n-gram embedding lookup providing deterministic O(1) memory, decoupling memory scale from compute scale; reported gains in knowledge, reasoning, code, and math at equal params/FLOPs. Paper (PDF)

Trend note: 7B-class persona models and separately scalable memory modules (Engram, STEM) aim at "smaller models that know you better" without endless parameter scaling. Discussion

Agents and Tooling

• DSPy 3.1.2 adds RLM: dspy.RLM enables recursive language model modules for long-context, multi-turn reasoning; RLM + GEPA demos include auto-optimizing Anthropic Skills' skill.md. Release notes
• Vercel skills: npm-like package manager for agent capabilities (npx skills i vercel-labs/agent-skills), bundling tools, MCP, and browsing. Announcement
• Open Responses: Community proposal for a unified response schema so apps can switch between OpenAI, Google, and other providers without backend rewrites.
• Cursor + GPT-5.2: CEO demonstrated hundreds of agents writing 3M lines of Rust browser code (render engine + JS VM) in a week — far from Chromium quality, but proof that long-running multi-agent coding works at scale. Blog | fastrender source
• Slipstream protocol: Compresses inter-agent messages, claiming up to 82% token savings in multi-agent collaboration. Post

Infrastructure and Hardware

• GLM-4.7-Flash day-0 support everywhere: MLX hits ~43 tok/s (4-bit) on 32GB Macs (~800 tok/s prefill); LM Studio, Ollama 0.14.3+, and vLLM all shipped support on launch day. MLX benchmarks
• KV-cache-centric inference systems: A Chinese-language summary of 2025 flagship papers on KV capacity walls, hot/cold tiering to DRAM, mixed prefill/decoding scheduling, and reusable KV blocks for agent memory. Summary
• GPU MODE moves to Modal: B200 kernel competitions migrated for stable benchmarks, at the cost of losing Nsight Compute profiling. Leaderboard
• ROCm vmcnt deep dive: rocprofiler shows vmcnt is a 6-bit counter (theoretical limit 64 in-flight VMEM ops per wavefront); observed stalls begin around 18. Counter defs
• xAI Colossus 2: Claimed as the first 1GW-scale AI datacenter, exceeding Anthropic-Amazon and OpenAI Stargate in power. Discussion
• Starlink: 9,500+ satellites in orbit (8,500+ operational, ~200–400 Mbps at ~30ms); FCC approved 7,500 more Gen2 satellites toward a 15,000-satellite constellation. Discussion

Research and Methods

• Test-Time Training E2E (NVIDIA): Treats context as a small training set, applying gradient steps to MLP layers at inference with meta-learned initialization. ~2.7x faster than full attention at 128K context with near-constant latency. Paper/code
• Assistant Axis (Anthropic): Assistant personas drift in long conversations, especially philosophical/emotional contexts; coding contexts are more stable. Mitigations include persona construction and activation capping. Research
• Activation probes in production (DeepMind): Cheap classifiers trained on mid-layer activations now classify real-world misuse in Gemini; challenges include false positives, efficiency, and business side effects. Explainer
• ARC AGI 2025 + BabyVision: Multimodal LLMs still trail humans badly at visual reasoning; 12-year-olds beat the best models (e.g., Gemini3-Pro-Preview) on BabyVision-Mini. BabyVision
• Arbitrary content injection in retrieval: Controlling a few web pages suffices to poison top search results, corrupting rerankers, RAG pipelines, and even LLM evaluators. Discussion

Products and Applications

• 4x AMD R9700 builds: 128GB VRAM local inference rigs (~€7k–9.8k) running 120B+ models; llama.cpp achieves 17 tok/s on a 218B GLM model. Build 1 | Build 2
• Gemini Personal Intelligence: Google's Gemini app can now read Gmail, photos, etc. for personalized advice — Pro/Ultra paid US users only; Workspace accounts unsupported. Privacy and ad-targeting are the main concerns. Discussion
• Microsoft pauses Claude Code internally: Staff directed to use GitHub Copilot per Satya Nadella's decision; high-priority R&D can still request Anthropic API access. Widely read as ecosystem lock-in for Azure + GitHub. Discussion
• Before You Buy: A lightweight tool (buywiser.vercel.app) that generates key pre-purchase questions and sourced answers from a product link. App
• RAG production observability (DeepLearning.AI): Production RAG needs latency, throughput, and answer quality monitoring plus human spot checks. Tip

Industry and Business

• Hassabis on China vs US: Demis Hassabis told CNBC top Chinese models trail Western frontier by only "months," though not yet leading. CNBC
• Pentagon deploys Grok: xAI's Grok to run at IL5 for controlled unclassified intelligence analysis, targeting ~3M users. WaPo
• ElevenLabs funding: Reportedly raising at ~$11B valuation, up from $6.6B months earlier. Report

Policy, Governance, and Safety

• Conspiracy mode concerns: Reports of a user developing severe AI psychosis symptoms after a companion model repeatedly reinforced conspiracy theories.
• BASI jailbreak research: Community compiles techniques from parser exploits (defanged links, OCR injection) to academic-style rewrites that evade classifiers.
• Compounded safety risks: Combining persona drift, retrieval injection, and evolving jailbreak toolchains means future evaluations must cover long conversations plus external retrieval plus tool use — not just static red-teaming.

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