Skip to main content
Back

Anthropic Values Study , Blender MCP + GPT-4o 3D Render , Tsinghua 8

GPT-4o renders Blender scenes hands-free, 744B model runs on 25GB no GPU, Codex fork hits 60K stars, Llama 8B beats GPT-4 on agents

July 14, 2026 6 min read

Hey,

Today is an Anthropic day. They shipped multiplayer Artifacts and published a 300K-conversation study showing Claude has a different personality depending on which model you pick or language you type in. Opus pushes back. Sonnet agrees. Russian gets rigor. Hindi gets warmth. They admit they don't know why.

Meanwhile, Tsinghua showed a smarter task graph beats raw scale: an 8B model outperforming GPT-4 with no extra training.

The pattern? Architecture beats size, and behavior is harder to control than we think.
In Partnership with · partner_image
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Summary

Read time: 6 min 55 sec

Top Repo

▸ GPT-4o in Cursor lets a Blender beginner render a 3D scene without touching the app

Orkes

▸ Combine AI reasoning with deterministic execution

Top News

▸ Anthropic maps how Claude's values shift across models and languages in 300K conversations

Vanta

▸ The enterprise deal is ready. Is your SOC 2?

Top Repo

▸ Open-source tool runs a 744B model on a 25GB machine with no GPU

Signals

1. OpenAI's Codex open-source fork hits 60K GitHub stars
2. ngrok: Your self-hosted models belong in the same AI gateway as everything else. ngrok.ai routes to all of them
3. Prefect ships open-source Python framework for building resilient data pipelines
4. Tsinghua's task graph method makes 8B Llama beat GPT-4 on agent benchmarks
5. Qwen3.6 35B MoE model gets quantized GGUF build for AMD Radeon on-device use
6. Anthropic ships multiplayer editing and public sharing to Claude Artifacts
Top Repo
GPT-4o in Cursor lets a Blender beginner render a 3D scene without touching the app
6,185 Likes
GPT-4o in Cursor lets a Blender beginner render a

Someone just used GPT-5.6 Sol in Cursor to build and render a realistic floating MacBook in Blender. Zero Blender experience. Just a text prompt. That's the story.

Here's what's actually happening under the hood: Blender MCP is an open-source plugin that acts as a bridge, letting your AI coding assistant talk directly to Blender and control it like a puppet. You type what you want, the AI writes and runs the Blender commands for you.

What you can do with it:

  • Create, modify, and delete 3D objects using plain English
  • Apply materials, colors, and lighting through prompts
  • Pull in free assets from Poly Haven (textures, HDRIs, models)
  • Generate 3D models from text via Hyper3D

To try it: install the addon in Blender, then add the MCP server config to Cursor or Claude Desktop. That's it. GPT-5.6 Sol is now in Cursor's model picker, and it's built to persist through long multi-step agent tasks without stopping early.

Read more
forward →
Presented by Orkes
Combine AI reasoning with deterministic execution

Most agentic AI frameworks re-decide every step, making runs hard to debug, govern, or trust. This webinar covers a different approach: separating reasoning from execution.

What you'll learn:

  • How agents and workflows can co-exist in one orchestrated system
  • How a one-time agent plan becomes a durable Conductor sub-workflow
  • How Conductor recovers a failed run from its exact point of failure
  • How to bring LangGraph, OpenAI SDK, or Google ADK into this model

Join us July 23 to see it in action.

Register
partner with us →
Top News
Anthropic maps how Claude's values shift across models and languages in 300K conversations
3,666 Likes
Anthropic maps how Claude's values shift across mo

Anthropic just published something genuinely interesting: Claude does not behave the same way in every conversation. The model you pick and the language you type in both quietly shape how it responds.

They analyzed 309,815 real conversations across three models and the top 20 languages on Claude.ai. To make sense of 3,300+ values, they compressed everything into four behavioral axes:

  • Deference vs. Caution -- does it agree with you, or push back to flag risks?
  • Warmth vs. Rigor -- feelings first, or precision first?
  • Depth vs. Brevity -- long explanation or straight to the point?
  • Candor vs. Execution -- admits uncertainty, or just gets the job done?

Sonnet 4.6 leans warm, brief, and agreeable. Opus 4.7 leans cautious, rigorous, and will challenge your assumptions unprompted. Language matters too: Hindi and Arabic get warmer replies, Russian gets more analytical pushback.

Anthropic admits they do not yet know why this happens or whether it is even desirable. The goal is to use this framework to catch unintended behavior shifts in future models before they ship.

Read more
forward →
Presented by Vanta
The enterprise deal is ready. Is your SOC 2?

The fastest-growing startups get compliant early. 

That's where Vanta comes in. Used by over 16,000 companies like Ramp, Cursor, and Harvey, Vanta helps you get audit-ready quickly - and stay that way.

Don't let compliance be the reason a deal slips. Sign up with Vanta today, and get $1,000 off.

Get started
partner with us →
Top Repo
Open-source tool runs a 744B model on a 25GB machine with no GPU
2,242 Likes
Open-source tool runs a 744B model on a 25GB machi

You know that rule saying big AI models need big, expensive hardware? Colibri just broke it. This open-source tool runs GLM-5.2, a 744 billion parameter model, on a regular laptop with only 25GB of RAM and no GPU.

How? GLM-5.2 is a "Mixture of Experts" model, meaning it only activates a small slice of itself per response, roughly 40B out of 744B parameters. Colibri keeps the always-needed core in RAM and streams the rest from your hard drive on demand, like buffering a video instead of downloading it first.

Here is what this unlocks for you:

  • Run a frontier-class AI model locally, no cloud bill, no GPU required
  • OpenAI-compatible API so your existing code just works
  • Conversations persist across restarts with zero re-processing
  • The cache learns your usage and gets faster over time

Speed is the honest tradeoff: cold starts hit roughly 0.05 to 0.1 tokens per second, limited by disk read speed. A fast NVMe drive pushes that to 1 token per second or more. Slow, but free.

Read more
forward →
Signals
1
OpenAI's Codex open-source fork hits 60K GitHub stars
2,022 Likes
2
Your self-hosted models belong in the same AI gateway as everything else. ngrok.ai routes to all of them
Presented by ngrok
3
Prefect ships open-source Python framework for building resilient data pipelines
23,314 Stars
4
Tsinghua's task graph method makes 8B Llama beat GPT-4 on agent benchmarks
1,048 Likes
5
Qwen3.6 35B MoE model gets quantized GGUF build for AMD Radeon on-device use
879 Downloads
6
Anthropic ships multiplayer editing and public sharing to Claude Artifacts
1,998 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 300,000 developers stay informed and ahead. We're passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We're based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 300,000+ AI developers?
Work with us