Rami's Readings #135 - Mistral's Devstral 2 & OpenAI's GPT-5.2
The latest on AI, LLMs, Mistral's Devstral 2, OpenAI's GPT-5.2, Cohere's Rerank 4, VectorChord for PostgreSQL, Oracle Debt, Holiday Reading List, Disk Prices, and more.
Welcome to Rami’s Readings #135 - 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.
📈 Top Recent Editions According to Substack
Vacation Notice: Starting tomorrow until the middle of January, I’ll be on a holiday break. You might receive a few surprise newsletters during this time. Wishing you a splendid end to the year!
🤖 AI Reads
Introducing Devstral 2 & Mistral Vibe CLI
Notes: Excellent release from Mistral! Devstral Small 2 has already replaced GPT-OSS-20B for my fun projects. Per SWE-Bench Verified, Devstral Small 2 (24B) outperforms the much larger GPT-OSS-120B. It runs fast on my RTX 4090. Ollama model link.
Introducing GPT-5.2
Notes: Three things caught my eye in the latest GPT-5.2 release from OpenAI.
“New knowledge cutoff of August 2025”. The previous 5.x models knowledge cutoffs were September 2024.
Pricing is 🚀
The ARC-AGI improvement is wild!
ARC-AGI-1 (Verified), a benchmark designed to measure general reasoning ability, GPT‑5.2 Pro is the first model to cross the 90% threshold, improving from 87% by o3‑preview last year while reducing the cost of achieving that performance by roughly 390×.
Introducing Rerank 4: Cohere’s Most Powerful Reranker Yet
Notes: Outperforms the Qwen3-Reranker model family.
Towards a Science of Scaling Agent Systems
Notes: Some problems are best solved by teams, while others are not so much.
Centralized coordination improves performance by 80.9% on parallelizable tasks like financial reasoning, while decentralized coordination excels on dynamic web navigation (+9.2% vs. +0.2%). Yet for sequential reasoning tasks, all multi-agent variants degraded performance by 39-70%.
llama-farm / llamafarm: Deploy Any AI Model, Agent, Database, RAG, and Pipeline Locally or Remotely in Minutes
Notes: Fairly opinionated, the project uses Ollama & Chroma as default.
CocoIndex: Data Transformation for AI
Notes: Written in Rust. Example usage: Building a Knowledge Graph from meeting notes that automatically updates.
How We Made 100M Vector Indexing in 20 Minutes Possible on PostgreSQL
Notes: Engineering fundamentals win again. VectorChord is another PostgreSQL extension for vector search.
Strategic Decision Framework for Enterprise LLM Adoption
Notes: Interesting framework. It’s a fast read.
💼 Business Reads
Oracle Debt Trades Like Junk as Bond, CDS Spreads Flare
Notes: Yikes.
How the Dollar Store Industry Overcharges Cash Strapped Customers While Promising Low Prices
Notes: In my experience, it is not just dollar stores. We have developed a habit of taking pictures of prices and checking what appears on the register, even at fancy grocery stores. It is rare that there isn’t at least one pricing mistake.
The 82 Books That Top Business Leaders Couldn’t Put Down
Notes: I’m looking forward to the holiday break to read 1929.
AI Leaders Eye New Breakthrough to Build More Powerful Models
Notes: I previously shared research in this space in the newsletter. Check out MIT’s SEAL.
🔀 Other Reads
Disk Prices
Notes: Neat website that tracks disk prices across most form factors and technologies.
Signing off from Redmond, WA.