🔥 Big Story
NVIDIA-Groq $20 Billion Deal: The Inference Wars Have Begun
This week, NVIDIA announced a landmark $20 billion licensing deal with Groq — the company's largest acquisition agreement ever. The deal includes hiring Groq CEO Jonathan Ross and key team members, bringing Groq's revolutionary LPU (Language Processing Unit) technology into NVIDIA's AI inference stack.
Why it matters: While NVIDIA has dominated AI training, this move signals their aggressive push into the rapidly growing AI inference market. Groq's technology has demonstrated 10x faster inference speeds compared to traditional GPU solutions. NVIDIA just acquired the technology to dominate both sides of the AI compute equation.
⚡ Quick Updates
Anthropic & OpenAI Double Holiday Usage Limits
Both companies gave premium subscribers a holiday gift: 2x message limits through December 31st. OpenAI even rolled out a festive "Santa Codex" model variant.
Read more →
Z.ai Releases GLM-4.7 Open-Source Coding Model
Chinese AI company Z.ai dropped GLM-4.7, achieving 73.8% on SWE-Bench Verified and 87.4% on τ²-Bench — the highest scores for any open-source model. Supports Claude Code, Cline, and Roo Code with 200K context.
Read more →
LG Unveils CLOiD Humanoid Robot
LG announced CLOiD at their CES 2026 preview, featuring two articulated arms with 5-finger dexterity powered by "Affectionate Intelligence" AI. Part of LG's ambitious "Zero Labor Home" vision.
Read more →
China Proposes AI Disclosure Regulations
New draft rules would require AI services to inform users they're interacting with AI at login and every 2 hours. Transparency in AI interactions is becoming a global regulatory priority.
Read more →
Claude Code Gets Major Updates (v2.0.70-2.0.74)
Anthropic shipped Native LSP support, Claude in Chrome (Beta), enhanced terminal UI, and improved clipboard navigation. Developer experience just got significantly better.
📄 Top Research Papers
1. Optimizing Decoding Paths in Masked Diffusion Models
Introduces "Denoising Entropy" — a novel metric that optimizes how Masked Diffusion Models generate sequences. By measuring uncertainty at each denoising step, the approach enables smarter token scheduling decisions, improving generation quality on reasoning and code benchmarks without additional training.
Possible Impact: Diffusion models are emerging as alternatives to autoregressive LLMs. This research makes them more practical for complex reasoning tasks.
2. Parallel Token Prediction for Language Models
PTP (Parallel Token Prediction) enables LLMs to generate multiple tokens simultaneously instead of one at a time. On Spec-Bench with Vicuna-7B, the approach achieved 4+ tokens per inference step, dramatically cutting latency.
Possible Impact: Sequential generation is a fundamental bottleneck in LLM inference. This could lead to 3-5x faster production LLMs.
3. Your Reasoning Benchmark May Not Test Reasoning
A critical analysis challenging popular benchmarks including ARC-AGI. The researchers found that 80% of VLM failures stem from perception errors, not reasoning deficits. When given correctly perceived inputs, models performed significantly better.
Possible Impact: We may be significantly underestimating AI reasoning capabilities. This could reshape how we design and interpret AI benchmarks.
📦 Top GitHub Repos
💻 smith
Blazing-fast AI coding agent for terminal, written in Rust. Supports Claude and OpenAI.
💻 awesome-agent-skills
Comprehensive collection of modular capabilities for building production-ready AI agents.
💻 agent-key
Secure credential management for AI agent workflows. Compatible with CrewAI, LangGraph, and Agno.
💻 Email-Ai-Agent
Privacy-first email agent using local LLMs. No cloud, no data sharing.
💻 mcp-server-bluesky-py
MCP server enabling AI agents to interact with Bluesky social network.
🛠️ Top AI Products
DiffSense |👍 406 upvotes
Local AI git commit generator using Apple's AFM 3B model. Zero latency, complete privacy.
NBot |👍 404 upvotes
AI curator that reads the entire internet and surfaces the 1% that actually matters. Kill the noise.
Alpie Core |👍 181 upvotes
32B reasoning model running entirely at 4-bit precision. Frontier performance, fraction of the compute. Open source.
🐦 Top Tweets
@karpathy
"I've never felt this much behind as a programmer. The profession is being dramatically refactored…"
@ycombinator
"A hacker who has learned what to make, and not just how to make, is extraordinarily powerful."
@claudeai
"Starting at midnight PT tonight, all Pro and Max plans have 2x their usual usage limits through New Year's Eve."
@github
"Sick of being on the hunt for bugs in your code? 👀"
Until next newsletter, Stay Curious and Keep building!!
Brain Pulse
