Rami's Readings #119 - Mistral Released Voxtral ⭐

Welcome to Rami’s Readings #119 - 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, Global Fintech Insider, 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.

🤖 AI Reads

One Token to Fool LLM-as-a-Judge

Notes: Yikes. This paper presses pause on an assumption I constantly overhear: human annotation is no longer necessary.

Despite the seeming simplicity of this comparison task, we find that generative reward models exhibit surprising vulnerabilities to superficial manipulations: non-word symbols (e.g., ”:” or ”.”) or reasoning openers like “Thought process:” and “Let’s solve this problem step by step.” can often lead to false positive rewards.

Context Rot: How Increasing Input Tokens Impacts LLM Performance

Notes: From Chroma. Yep…

Energy-Based Transformers are Scalable Learners and Thinkers

Notes: Really interesting paper from UVA, UIUC, Amazon GenAI, Stanford and Harvard.

EBTs are the first instance of an approach that scales at a faster rate than the Transformer++ during pretraining across both continuous and discrete modalities.

Voxtral: Mistral’s Speech Understanding Model ⭐

Notes: Outperforms Whisper v3 (which is already incredible).

Agentic-R1: Distilled Dual-Strategy Reasoning

Notes: From CMU.

The Lethal Trifecta for AI Agents: Private Data, Untrusted Content, and External Communication

Notes: Simon Willison is just always on point.

As a user of these systems you need to understand this issue. The LLM vendors are not going to save us! We need to avoid the lethal trifecta combination of tools ourselves to stay safe.

Universal Tool Calling Protocol (UTCP)

Notes: GitHub. UTCP vs MCP comparison.

trymirai / uzu: A High-Performance Inference Engine for AI Models

Notes: Designed only for Apple Silicon.

💼 Business Reads

Ambani’s Jio Agrees to Reinsurance Venture With Allianz in India

Notes: Huh…

Jane Street’s Trading Secrets Spill Into Open and Face Rivals’ Scrutiny

Notes: Wild.

Stop pretending you know what AI does to the economy Have you ever had a song stuck in your head? You just walk around humming it all day, and you can’t quite seem to get it out. I find that some intellectual ideas work kind of like that. Some people walk around all day reflexively applying the word “neoliberalism” to everything they don’t like about the modern world. Others are convinced that solar power and batteries will never be viable technologies. Still others blame immigrants for every way in which their country is not yet a paradise, or think that vaccines cause every health problem in society. And so on. Once an idea like this gets stuck in your head, confirmation bias takes over, and you begin to see it everywhere around you — which only reinforces the strength of the idea in your mind. Instead of an earworm, it’s a brainworm… Noahpinion · 9 months ago · 253 likes · 119 comments · Noah Smith Stop pretending you know what AI does to the economy