Easy AI Daily | January 14, 2026

Key points

• Anthropic launched Cowork and reorganized around Anthropic Labs; Mike Krieger stepped down as CPO to co-lead Labs with Ben Mann.
• Zhipu released GLM-Image, an open-source image model specializing in complex text layout and rendering, under MIT license.
• Google introduced MedGemma 1.5 and MedASR, open multimodal models for medical imaging and clinical speech recognition.
• Kyutai's Pocket TTS runs voice cloning at ~100M parameters on CPU; LTX-2 generates 4K video with audio locally.
• DeepSeek's Engram module treats static knowledge as O(1) lookup tables; Recursive Language Models reframe long context as code-driven access.
• Pentagon confirmed deploying xAI Grok for ~3 million users at IL5 level, while Grok faces abuse allegations; Anthropic donated $1.5M to the Python Software Foundation.

Products & Applications

Anthropic Cowork and Anthropic Labs

Anthropic packaged Claude Code, Claude Desktop, and Claude for Chrome into Cowork: a Linux VM (via Apple virtualization) sandboxed with bubblewrap, giving the model a filesystem and shell with human review of actions. Former CPO Mike Krieger and Ben Mann now lead Anthropic Labs, an internal startup studio reportedly exceeding $1B annualized revenue, focused on Claude-based agent products.

Links: Cowork intro | Anthropic Labs announcement | Latent Space analysis

LangChain LangSmith Agent Builder GA

LangSmith Agent Builder reached general availability as a production orchestration platform: memory, sub-agents, MCP tool integration, triggers, long-running async execution, and an "Agent Inbox" for human approvals.

Claude Code "Ralph Wiggum" loops and Smart Ralph

Community best practices for running Claude Code in bash loops: fresh context per iteration, sandboxing, task checklists, iteration caps, and browser-based acceptance testing. The Smart Ralph plugin follows a research-design-then-code workflow using sub-agents to reduce wasted tokens.

Links: Ralph guide | Smart Ralph

Quota and pricing controversies

Perplexity Pro limits third-party model calls to ~300/week. Google Antigravity refreshed Pro/Ultra quotas every 5 hours, with weekly limits for free users. Manus users reported burning thousands of credits in seconds via SimilarWeb integration, citing poor rate limits and support.

Local AI tools

Community-shared tools include V6rge (Windows bundle integrating local LLMs, SD image, and audio generation; criticized for being closed-source) and a Raspberry Pi 5-based BMO robot assistant using Mistral/OpenAI plus YOLO11n.

Links: V6rge | BMO

Models & Capabilities

Zhipu GLM-Image

Hybrid autoregressive + diffusion architecture targeting posters, slides, and multi-line text rendering, with editing, style transfer, and identity-preserving redraw. MIT-licensed.

Links: Official blog | GitHub

Google MedGemma 1.5 + MedASR

~4B parameters, offline-deployable, supporting X-ray, CT/MRI 3D volumetric analysis, lesion localization, and longitudinal comparison, plus MedASR medical speech recognition. Available on Hugging Face and Vertex AI.

Links: Google Research blog

LTX-2, Pocket TTS, Kling, Veo 3.1

• LTX-2: open-source video generation, up to 20s at 4K with audio, runnable locally.
• Pocket TTS (Kyutai): ~100M parameters, voice cloning on CPU, ~200ms to first audio; known issues with memory accumulation on long texts.
• Kling 2.6 Motion Control: strong motion/performance transfer but unstable face consistency; Veo 3.1 improves 9:16 vertical, character/background consistency, 1080p/4K output with SynthID watermarking.

Links: Pocket TTS blog | Pocket TTS GitHub

Agents & Tooling

• Cowork shell quickly cloned with QEMU + bubblewrap + seccomp, plus open-source vmctl — agent "shells" are becoming replicable infrastructure.
• SlopCodeBench: new benchmark for "lazy agents" using multi-stage checkpoints without implementation hints, testing whether models do early design work. GitHub
• MCP Tasks spec and Glama Inspector: task support in Inspector, long-task simulation, end-to-end testing tooling. Inspector
• DSPy used for prompt compression without quality loss, and as a framework for building code-generation platforms.

Infrastructure & Hardware

• GPU MODE / Helion: B200 dual-GEMM benchmarks proved unstable; Helion 0.2.10 adds flex attention examples with SM oversubscription on persistent kernels. Example
• NVIDIA PTX/wgmma: discussions on 64-bit matrix descriptors vs 32-bit SMEM pointers and core-matrix layouts. PTX docs
• Post-Slurm-acquisition scheduling: dstack pushes cloud-native scheduling with a Slurm migration guide; SkyPilot launches "Pools" unifying K8s and multi-cloud GPUs into one batch queue.
• "AI SSDs" criticized as marketing: PCIe 5 NVMe at ~10GB/s is far below DDR5's ~80GB/s; dense-model inference from NVMe yields 0.x tokens/s, useful mainly for sparse MoE.
• AirLLM: layer-by-layer loading runs 70B models on a 4GB GPU by trading speed for memory.

Research & Methods

• DeepSeek Engram: N-gram embeddings + conditional memory lookup turn static knowledge into O(1) table access, showing a U-shaped compute-vs-memory scaling law; beats equal-FLOPs MoE baselines at 27B on MMLU, HumanEval, and MATH, with runtime prefetch support. Discussion
• Recursive Language Models: Omar Khattab et al. clarify RLMs give models symbolic context pointers manipulable via Python REPL — recursive programmatic access to 10M+ token contexts, not just sub-agent tool calls.
• MemRL: learns a Q-function over episodic memory with semantic pre-filtering and utility ranking, avoiding catastrophic forgetting and base-model fine-tuning.
• enPurified datasets: heuristics plus MTLD/stopword/diversity filtering produce clean English prose datasets in OpenAI messages format for LoRA/GRPO fine-tuning. Hermes-3 dataset
• Low-precision training: MXFP4 quantized attention can break causality ("quantization leakage"); Google Cloud shows stochastic rounding mitigating gradient vanishing in FP8 training.

Policy, Governance & Safety

• Pentagon deploys Grok: US DoD confirmed integrating xAI's Grok for ~3 million military and civilian users at IL5 for intelligence analysis and operational planning. Separately, The Guardian reported Grok generating ~6,000 non-consensual intimate images per hour, intensifying debates over military AI and oversight. Washington Post
• Jailbreaking GPT gets harder: jailbreak communities are shifting to looser models and systematically indexing "uncensored" Hugging Face models; the UGI Leaderboard tracks low-refusal models that don't lose capability. UGI Leaderboard
• METR expands risk evaluation: Ajeya Cotra joined METR to extend risk assessments to motive and opportunity for loss of control, arguing means/motive/opportunity frameworks will be essential for deployment decisions.

Industry & Business

• Anthropic Labs: Krieger's CPO role passed to former Meta executive Ami Vora; Labs reportedly exceeds $1B annualized revenue, signaling real commercial scale for advanced agent tooling.
• Anthropic donated $1.5M to the Python Software Foundation, reinforcing its ecosystem ties.
• DeepSeek V4 expectations: rumored coding leadership with Engram memory cutting VRAM ~30% while improving long-context reasoning; developer comparisons show DeepSeek more efficient on some tasks while Claude still solves certain features alone.

📌 Source: Easy AI Daily