Rami's Readings #145 - ✨ Gemma 4 & Claude Code
The latest on AI, LLMs, Gemma 4 Family of Models, Claude Code Leak, Liquid AI, TurboQuant, OpenShell, and more.
Welcome to Rami’s Readings #145 - a weekly digest of interesting articles, papers, videos, and X threads from my various sources across the Internet. Expect a list of reads covering AI, technology, business, culture, fashion, travel, and more. Learn about what I do at ramisayar.com/about.
👋🏼 Welcome New Subscribers
Hello! A hearty thank you for subscribing to Rami's Readings! There are quite a few new subscribers this week, thanks to recommendations from The AI Ethics Brief, Product Byte, and The VC Corner. I am thrilled to have you on board! In this newsletter, I curate the best papers, tweets, and articles I have read during the week focusing on LLMs, AI, economics, business, and technology news. You can learn more about me on my website.
Apologies for missing last week, the whole family caught a new variant of Covid which escaped our existing immunity. Felt surreal to lose my sense of taste again.
🤖 AI Reads
Biggest News of the Past Two Weeks: Claude Code Leaked
Notes: Original leak was through their npm registry, read this HN thread. The title link is to ccleaks which breaks down all the unreleased features. Read this analysis too along with the HN comments. Claude Code was rewritten in Rust using Codex in this open source project so as to be free from DMCA takedowns. Also, read this Reddit thread.
✨ Google Releases Gemma 4: Maximize Intelligence-Per-Parameter
Notes: 4 new models: E2B, E4B, 26B-A4B, 31B. I’m already running 26B-A4B and 31B locally. They are excellent! Gemma 4 31B might replace the Qwen 3.5 family of models in my local deployment (Qwen 3.5 is also new, see #142).
Liquid AI's LFM2.5-350M: No Size Left Behind
Notes: MIT’s Liquid AI is pushing impressive models at very small sizes.
thedotmack / claude-mem: Better Memory for Claude Code
Notes: Overkill for most projects unless it is >100K LOC and even then…
Nimbalyst: The Visual Workspace for Building with Codex and Claude Code
Notes: Neat project, but it’s hard to beat the customization you can build with any old-school terminal.
TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate
Notes: This paper supposedly lead to memory manufacturing stocks to decline. Yes, it’s a great development, but the stock market reaction was overblown.
Agents of Chaos
Notes: Paper covering extensive red-teaming of deployed agents.
PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost
Notes: Interesting paper for subscribers fine-tuning for agentic scenarios.
OpenClaw Arena
Notes: StepFun 3.5 Flash is number one! Congratulations to ex-STCA friends!
NVIDIA OpenShell
Notes: My local setup looks a lot like this new open source project from Nvidia, except without the K8s overkill, but with the same network activity filtering to locked down VMs. GitHub.
Lemonade: Refreshingly Fast LLMs on GPUs and NPUs
Notes: Supported by AMD. Another inference tool for running LLMs locally on AMD hardware.
Simon Willison is a subscriber favorite. Great interview with Lenny:
🔀 Other Reads
Supply Chain Attack on Axios Pulls Malicious Dependency from npm & LiteLLM Hack: Were You One of the 47,000?
Notes: The number of supply chain attacks is getting unwieldy…
Signing off from Redmond, WA.