Claude Gets a Blender Connector — Anthropic's First Move Into Creative Tooling

🧠 LAUNCH

Claude Gets a Blender Connector — Anthropic's First Move Into Creative Tooling

Claude now connects directly to Blender via a new MCP connector — debug scenes, build custom tools, and batch-apply changes across every object without leaving your DCC app. This isn't a toy integration: it's full agentic control over a professional 3D pipeline, and it signals Anthropic is done being "just a coding assistant." If you're a 3D artist or technical director, this collapses the feedback loop between intent and execution. (26,509 likes | 2,068 RTs) Read more →

Mistral Launches Workflows — An Enterprise Orchestration Layer With Durability Built In

Mistral ships Workflows in public preview — a production orchestration layer for enterprise AI with durability, observability, and fault tolerance baked in. This fills the gap that's killed most agent demos in production: what happens when a multi-step process crashes mid-run? Mistral's answer is checkpointed, resumable execution that survives failures. If you're building agents that need to run for hours, not seconds, this is the infrastructure layer you've been hand-rolling. (1,428 likes | 171 RTs) Read more →

Microsoft's VibeVoice Ships as the Open-Source Whisper Successor You Can Run Locally: Simon Willison's hands-on review confirms VibeVoice runs well on Apple Silicon — 60GB RAM, 4-bit quantized, with speaker diarization built in. MIT-licensed and production-ready, this is what Whisper should have become. (946 likes | 82 RTs) Read more →

Gemma 4 Drops an Apple Silicon-Optimized 31B Model: Google's Gemma 4 now has a 31B variant specifically tuned for local Mac inference. A frontier-tier open model you can run natively on your MacBook — pair it with last week's WebGPU demo and the local AI story just got a lot more practical. (880 likes | 66 RTs) Read more →

Sakana AI Opens Beta for Fugu Multi-Agent Orchestration: Sakana Fugu enters beta — a multi-agent orchestration system from the lab known for evolutionary model merging. Lighter than the full platform plays, with durable execution as a headline feature. A solid starting point for teams exploring multi-agent workflows without committing to a monolithic framework. (650 likes | 151 RTs) Read more →

📝 TECHNIQUE

Anthropic's Official Guide to Onboarding Claude Code Like a New Developer

Treat Claude Code like a new hire, not a magic autocomplete. Anthropic's official blog distills 17 years of development experience into a structured onboarding framework — CLAUDE.md files, persistent memory, and skill definitions dramatically improve output quality. The key insight: the more context you give Claude Code upfront, the less you correct it later. If you're using Claude Code without a CLAUDE.md, you're leaving performance on the table. Read more →

For a deeper dive on prompting strategies, see our guide: How to Effectively Prompt Claude Code.

🔧 TOOL

Claude Code CLI Ships 50+ Fixes in Four Releases — The Polish Sprint Continues: Claude Code just pushed 50+ stability and performance fixes across four rapid releases — faster resume, stable auth, and reliability improvements across the board. The pace of polish is accelerating, and it shows. Update now. (1,568 likes | 60 RTs) Read more →

Codex Can Now Auto-Migrate Your Repos to GPT-5.5: OpenAI's Codex adds automated model migration — point it at an existing repo and it upgrades your code to GPT-5.5. One agent command replaces hours of manual API version bumping. (616 likes | 35 RTs) Read more →

FigJam Gets MCP — Generate Architecture Diagrams Straight From Code: Figma's FigJam MCP integration now generates architecture diagrams and ERDs directly from your codebase. Agents can read and annotate your design boards, bridging the gap between code understanding and visual system design. (344 likes | 22 RTs) Read more →

OpenMed's Privacy Filter Catches 55+ PII Categories vs. OpenAI's 8: Same text, two filters, wildly different results. OpenMed detects medical record numbers, blood types, and dozens more sensitive categories that OpenAI's model misses entirely. If you're building healthcare AI pipelines, this is the de-identification layer HIPAA compliance actually requires. (876 likes | 75 RTs) Read more →

Transformers v5.7.0 Adds Day-1 Support for Poolside's Laguna Models: HuggingFace Transformers v5.7.0 ships with support for Laguna — Poolside's MoE model family — plus other new architectures. Day-1 support for a new model family confirms the library's role as the universal model interface. (2.7K downloads) Read more →

Ollama v0.22.0 Adds Nemotron 3 Omni and Laguna for Local Inference: Ollama v0.22.0 brings NVIDIA's Nemotron 3 Omni and Poolside's Laguna XS.2 to local inference. Two significant new model families you can run on your own hardware — the local ecosystem keeps expanding. Read more →

🔬 RESEARCH

GPT-5.4 Pro Helps Crack a 60-Year-Old Erdős Conjecture

A 60-year-old Erdős problem has been solved with help from GPT-5.4 Pro — not by brute-forcing a proof search, but by contributing genuine reasoning steps that human mathematicians built on. This is a concrete, peer-verifiable milestone for AI-assisted research, and it changes the conversation about what "understanding" means when a model contributes to novel mathematics. OpenAI researchers discuss what shifts as AI gets genuinely good at math. (1,614 likes | 146 RTs) Read more →

GPT Co-Creator Alec Radford Releases Talkie — A 13B Model Trained on Pre-1931 English: Talkie is a 13B model trained exclusively on 260B tokens of historical text — a fascinating experiment from Alec Radford in what LLMs learn when the training data predates the modern internet. Simon Willison's notes are worth the read. (223 likes | 24 RTs) Read more →

NVIDIA's Nemotron 3 Nano Omni Handles Documents, Audio, and Video in One Model: Nemotron 3 Nano Omni processes documents, audio, and video in a single multimodal model with long-context support — purpose-built for agent workflows that need to reason across diverse input types without switching models. Read more →

💡 INSIGHT

OpenAI Models Land on Amazon Bedrock — The Multi-Cloud Era Begins: Following the Microsoft exclusivity dissolution, OpenAI models are coming to Amazon Bedrock. Altman and Garman's Stratechery interview makes the strategy explicit — OpenAI is now multi-cloud, and enterprise procurement just got a lot more interesting. If you're on AWS, evaluate OpenAI-on-Bedrock before your next model decision. (160 likes | 58 RTs) Read more →

OpenAI Bets That $8 ChatGPT Will Push Subscribers to 122 Million: OpenAI projects its ad-supported $8 ChatGPT tier will drive subscribers to 122M this year — while cannibalizing $20 Pro users. A massive bet that volume at lower price points beats premium retention, and it could reshape consumer AI economics for everyone. Read more →

Who Owns the Code Claude Code Wrote? The IP Question Gets Real: A trending Hacker News legal analysis digs into IP ownership of AI-generated code — specifically Claude Code's output. As agentic coding scales from side projects to enterprise codebases, the copyright question becomes unavoidable. Review this with your legal team before it reviews you. (231 likes | 275 RTs) Read more →

🏗️ BUILD

VibeVoice Repo Hits GitHub Trending — MIT-Licensed Production Voice Pipeline: Microsoft's VibeVoice is trending on GitHub and Hacker News — an MIT-licensed speech-to-text pipeline with speaker diarization, positioned as the open-source successor to Whisper for production voice workflows. Star the repo and test on your audio data. (311 likes | 166 RTs) Read more →

🎓 MODEL LITERACY

Durable Execution: Mistral's Workflows and Sakana's Fugu both launched this week with "durability" as a headline feature — but what does that actually mean? Durable execution guarantees that a long-running agent process can survive crashes, restarts, and network failures by checkpointing its state at every step. If a workflow dies mid-run, it resumes from the last checkpoint instead of starting over. Think of it like a save game for your AI agent — without it, a 4-hour research task that fails at hour 3 means losing everything. This is the missing reliability layer that separates demo agents from production ones, and it's why both Mistral and Sakana are leading with it.

⚡ QUICK LINKS

Anthropic Opens Sydney Office: Theo Hourmouzis named GM of Australia & New Zealand as APAC expansion continues. Link
Claude Code v2.1.122: Adds Bedrock service tier selection and PR URL resume in /resume. Link
Xiaomi's MiMo-V2.5: The MIT-licensed agent model optimized for 1,000+ tool calls gains real adoption on HuggingFace. (127 likes | 2.7K downloads) Link
Pika Launches Persistent AI Creative Agents: Not one-shot generation — ongoing creative collaborators with memory and style consistency. (317 likes | 54 RTs) Link
Musk Takes the Stand in OpenAI Trial: Argues the case sets precedent for charitable trust enforcement — the outcome could reshape AI company structures. Link

🎯 PICK OF THE DAY

A 60-year-old Erdős conjecture falling to AI-human collaboration isn't just a math milestone — it's a category shift. Most AI achievements in formal reasoning have been either brute-force search (AlphaProof) or pattern-matching on known techniques. This is different: GPT-5.4 Pro contributed novel reasoning steps that human mathematicians validated and built upon. It's the first clean proof that frontier models can participate in open research problems rather than regurgitate training data dressed up as insight. The uncomfortable question the math community now faces isn't whether AI can do math — it's what "understanding" means when a machine helps prove a theorem it can't explain in natural language. For the broader AI industry, this is the kind of result that moves the conversation from "useful assistant" to "research collaborator," and that distinction has massive implications for how we fund, credit, and structure scientific discovery. Read more →

Until next time ✌️