Greetings to all those who celebrate Christmas. Zhipu and Minimax delighted fans with updated models such as GLM-4.7 and Minimax-M2.1 which show great results in early testing and bode very well for open source and open weight models.
Andrej Karpathy ended the week with a banger of a post on X talking about the challenges programmers face today in unlearning old habits as well as keeping up with a much more complex mental model around tooling.
Readers who think others in their family, friends and acquaintances who have a curiosity in knowing more about rapidly evolving AI tools/services/use cases would benefit from being subscribers of this weekly newsletter are encouraged to share this publication link to them and invite them to subscribe
The following messages were posted on the 'All Things AI' Telegram group from Sunday 21st Dec 2025 to Saturday 27th Dec 2025
Sam Rose from the nGrok team with one of the best explainers I've seen of how LLMs work overall
Prompt caching: 10x cheaper LLM tokens, but how? | ngrok blog
A far more detailed explanation of prompt caching than anyone asked for.
Courses on Skilljar which are a part of Anthropic Academy
Anthropic Courses
Learn to build with Claude AI through Anthropic's comprehensive courses and training programs.
1 | Zhipu AI Launches GLM-4.7, Top Open-Source Model for Coding and Reasoning The Beijing-based company's latest GLM series model shines on benchmarks, scoring 73.8% on SWE-bench Verified for coding, 95.7% on AIME 2025 math problems, and leading agent tasks like multi-step tool use at 87.4%. Available free on chat.z.ai or via API on OpenRouter, it quickly hit No. 1 on LMArena's open model leaderboard, beating rivals like Claude Sonnet 4.5. Zhipu AI's team celebrated the 'romance' of hard work behind it, with watchers calling the results insane as open models challenge closed giants. You’ve been invited to join the GLM Coding Plan! Enjoy full support for Claude Code, Cline, and 10+ top coding tools — starting at just $3/month. Subscribe now and grab the limited-time deal! |
2 | Kilo Code, the AI code startup from former GitLab cofounder Sid Sijbrandij is taking aim at Lovable with a new App Builder tool. Introducing App Builder: A Lovable Alternative in Kilo's Agentic Engineering PlatformBuild and preview live websites and applications with natural language  The New Vibe Coding Startup Taking Aim At Lovable (And Everyone Else)Kilo Code, cofounded by GitLab's former CEO Sid Sijbrandij, launched its App Builder on Monday after a six-week sprint, Upstarts Media exclusively reports.  |
3 | A look at Google's TPU evolution, its decision to sell chips to competitors after 12 years of internal use, the TPU vs. GPU architecture showdown, and more TPU ManiaIs GPU vs TPU the CISC vs RISC contest of the modern era?  |
4 | MiniMax Launches M2.1 Open-Source AI Model for Coding Excellence The Chinese company released model weights on Hugging Face today, enabling local runs and API access through platforms like Ollama and OpenRouter. Optimized as a 10B-activated mixture-of-experts, it shines in multi-language programming, app development for Android and iOS, and agent workflows, with self-reported scores like 74% on SWE-bench Verified beating Gemini 3 Pro and Claude Sonnet 4.5 in some tests. Developers quickly praised its speed for tasks like building 3D web apps and simulated macOS interfaces, though some note hiccups on niche challenges as independent evals roll in
MiniMax M2.1 is officially live
 Built for real-world coding and AI-native organizations — from vibe builds to serious workflows. A SOTA 10B-activated OSS coding & agent model, scoring 72.5% on SWE-multilingual and 88.6% on our newly open-sourced VIBE-bench, exceeding leading 
Proud of the team.
Beating Gemini 3 Pro and Claude Sonnet 4.5 on SWE-Multilingual and VIBE-Bench. The most powerful OSS model for the agentic era is here.
MiniMax M2.1 is officially live
Built for real-world coding and AI-native organizations — from vibe builds to serious workflows. A SOTA 10B-activated OSS coding & agent model, scoring 72.5% on SWE-multilingual and 88.6% on our newly open-sourced VIBE-bench, exceeding leading
.@MiniMax__AI M2.1 is now available in Cline.
200K context window. 128K max output. MoE architecture (10B active / 230B total params). Here's what's improved: 1. Better code quality - more readable, maintainable output 2. Smarter instruction following on complex multi-step 
Read more on minimax.io/news/minimax-m…
Open-source release in 2 days, stay tuned! |
5 | via Robby Yung
Hollywood is cooked.
A single ai artist just dropped a 7 minute film made 100% with ai tools and it went viral on Chinese social media in a week. The film explores the idea that our world might be a simulation. Built with Nano Banana, Veo3 and Runway, music by Suno. The era of 
|
6 | via Robby Yung
Here’s Claude absolutely RIPPING building a modern mansion draft with primitives—using UnrealMCP to control Unreal Engine by himself.
You prompt what you want, he does what you ask. He can be creative too. Tiny little mini tutorial thread  
so there’s a free thing called Unreal MCP that lets you prompt Claude to build stuff in Unreal Engine.
“make a Victorian manor. here’s a reference pic. use the assets in this folder.” And he just…does posting a vid on it Monday, but if you’re bored tonight, 10/10 magic |
A curated list of awesome Claude Skills, resources, and tools for customizing Claude AI workflows
Usable across different surfaces within Claude as well as in Droid from FactoryAI
GitHub - ComposioHQ/awesome-claude-skills: A curated list of awesome Claude Skills, resources, and tools for customizing Claude AI workflows
A curated list of awesome Claude Skills, resources, and tools for customizing Claude AI workflows - ComposioHQ/awesome-claude-skills
1 | Nvidia Licenses Groq's AI Inference Tech in $20 Billion Deal, Hires Key Leaders The non-exclusive licensing agreement lets Nvidia use Groq's high-speed Language Processing Units for AI inference, while Groq continues independently with its GroqCloud service. Founder Jonathan Ross, ex-Google TPU designer, and President Sunny Madra join Nvidia, with Simon Edwards as Groq's new CEO. CNBC reported a $20 billion cash deal for assets from the $6.9 billion-valued startup, confirmed by outlets like Reuters as sidestepping antitrust issues amid rising demand for efficient inference chips.
Today Groq entered into a non-exclusive licensing agreement with Nvidia for Groq’s inference technology. Along with other members of the Groq team, I’ll be joining Nvidia to help integrate the licensed technology. GroqCloud will continue to operate without interruption.
Learn 
groq.com
Groq and Nvidia Enter Non-Exclusive Inference Technology Licensing Agreement to Accelerate AI...The Groq LPU delivers inference with the speed and cost developers need.
Why @nvidia acquihiring @Groq for $20B is so interesting:
It looks like a direct (and fast!) reaction by Nvidia to Google successfully using its own TPU chips for training AI models and inference (generating AI content) Because until recently, they used mostly Nvidia's GPUs 
Asked @pieterbot on AOL why @nvidia acquired @GroqInc
And they apparently did not, they just did a licensing deal for $20B Possibly to avoid antitrust laws (so Nvidia doesn't get a monopoly) "In a Christmas Eve deal, Nvidia has entered into a nonexclusive agreement to license 
Nvidia wishes you merry christmas by acquiring Groq for $20B its largest acquisition ever and here’s why this move will define the next decade of AI infrastructure
personally what fascinates me is how Jensen eliminated his greatest threat on christmas eve  understand the  |
2 | OpenAI quietly released a selection of “Prompts Packs” for different roles. You can find them in the OpenAI academy. These packs include more than 300 carefully curated prompts, covering over a dozen departments including IT, HR, Sales, Product, and Customer Success, with around 25 prompts tailored to each role. Think of it as getting thousands of hours of prompt engineering, completely free |
1 | --- Vercel just published a retrospective on their internal text-to-SQL agent. They realized they were solving problems the model could handle on its own. So they deleted 80% of the code. They stripped the agent down from a complex suite of tools to a single capability. They gave it a file system with basic Unix commands like grep and ls and stopped trying to think for the model. The results were immediate: → Success rate on evals went from 80% to 100% → It became 3.5x faster → It used 37% fewer tokens We removed 80% of our agent's tools - VercelWe spent months building a sophisticated text-to-sql agent, but as it turns out, sometimes simpler is better. Giving it the ability to execute arbitrary bash commands outperformed everything we built. We call this a file system agent.  |
2 | Two of the most influential people in 'vibe coding' talking about the challenges programmers face today in unlearning old habits as well as keeping up with a much more complex mental model around tooling.
I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become
I feel this way most weeks tbh. Sometimes I start approaching a problem manually, and have to remind myself “claude can probably do this”. Recently we were debugging a memory leak in Claude Code, and I started approaching it the old fashioned way: connecting a profiler, using the
Andrej Karpathy literally built the neural networks running inside coding assistants.
He taught the world deep learning at Stanford. He ran AI at Tesla. If he feels “dramatically behind” as a programmer… that tells you everything about where we are. The confession here is
I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between. I have a sense that I could be 10X more powerful if I just properly string together what has become
I think the rise of cheap coding plans from Zhipu/Minimax and their integration into the usual suspects of agentic coding harness's such as Cline, Droid, Claude Code, Kilo Code, Cursor, Windsurf etc is going to bring forth a new generation of programmers assuming that educational institutions adjust their teaching quickly |
Below is my personal website which aggregates links to many of my socials as well as content and community that I curate. Feel free to share this link to others who you think may find this content/community useful to them
The cover image of this newsletter via generated via the Seedream 4.5 model within the Freepik tool via the following prompt
A vibrant, high-detail image of a futuristic mechanical blues band, featuring humanoid robots with polished chrome surfaces and intricate gears, passionately playing classic blues instruments like electric guitars, harmonicas, and a drum set. The scene is set in a dimly lit, smoky blues club with warm amber and deep blue lighting casting dramatic shadows across the metallic figures. Neon signs glow faintly in the background, and the atmosphere feels soulful and nostalgic despite the robotic performers. The text 'Happy All Things AI New Year' appears in bold, glowing neon script above the band, adding a festive touch to the moody, immersive setting.
Wishing everyone Happy New Year 2026 in advance
Regards, Yusuf