👋🏼 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.

🤖 AI Reads

The two following papers consumed my entire thinking this week.

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Notes: A super important paper to read from Randall Balestriero and Yann LeCun.

LLMs Can Natively Read & Write Neural Network Weights

Notes: Another interesting discovery from MIT. Is all that investment in training infrastructure even necessary now?

Fastino GLiDE: Decision Models for Classification, Routing, and Scoring with Text, JSON, and Images

Notes: Full family of open-weight decision models from the same company behind GLiNER.

💼 Business Reads

Nvidia CEO’s Affair Engulfs AI Giant in Chinese Spying Probe

Notes: Not what I expected to be thinking about this weekend. The memes just wouldn’t stop too. I got wind of the story more than a week ago, reports were around, but it’s interesting how quickly this story blew up this weekend. The stock market will be interesting tomorrow. NVIDIA shares are already down on Robinhood. Gary Marcus on this affair.

Buffett and Munger Would Know How to Invest in AI

Notes: Yes. Also, I can’t stop thinking about this Charlie quote this weekend.

There are only three ways a smart person can go broke: liquor, ladies and leverage.

The Global Office: How Globalisation’s Frontier is Shifting from Factories to Offices

Notes: Having spent most of my career at Microsoft, this shift was intuitive to me. It’s good to see data.

What Millions of Data Points Reveal About US Inflation

Notes: Beautiful data visualization.

Tokyo Is Being Reshaped by a Historic Real Estate Boom

Notes: It’s still early and not too late if Tokyo is a long-term destination for you.

🔀 Other Reads

Sony Debuts $60 Wired Headphones With ’80s-Inspired Design

Notes: YES! I had a pair of these when I was a kid. Great memories!

Signing off from Redmond, WA.

👋🏼 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.

🤖 AI Reads

Introducing LFM3: Every Modality for Specialized Intelligence

Notes: I like Liquid AI’s model family strategy. #MIT

Introducing Rerank 5: Precise, Efficient Enterprise Document Retrieval

Notes: Cohere released Rerank 5.

nvidia / Kumo-Tabular: NVIDIA’s Tabular Foundation Model

Notes: We need more tabular data models.

AirLLM: 2.8T LLM Inference with Single 4GB GPU

Notes: If you really need to load larger models in constrained memory environments, this project helps. Inference speeds will be constrained by the machine’s ability to load and unload layers fast enough to compute.

💼 Business Reads

ChinaTalk

Korea’s Stock Market Crashed $1 Trillion. What Does That Mean for the Global AI Race?

Richard Gray is an analyst who works for a foreign policy think tank in Washington, D.C…

Read more

8 days ago · 48 likes · 1 comment · Richard Gray

Rate Hike Kicks Canada’s Housing Market When It’s Down

Notes: Inflation-adjusted prices are back around the 2016 baseline (still too expensive!)

🔀 Other Reads

TomAnthony / YourOwnWeigh: Free Wi-FI Syncing for Withings Scales

Notes: I have a “grandfathered” Withings Scale, so I still have free syncing to third-party apps. However, I am completely put off by the company.

Trying Not to Sound Like AI Is Changing How We Write

Notes: We are now in year 3 of this saga…

When Is Scotch Worth $36,000? Take a Look at the Packaging

Notes: Admittedly, I also keep the bottle and packaging of my favorite whiskey. Why not?

Signing off from Redmond, WA.

👋🏼 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.

🤖 Seattle AI Week

Seattle AI Week is taking place this week. Unfortunately, I will not be able to attend any events due to a prior commitment, but here are a few I would attend if I could:

Fireside Chat With LawZero & Networking Reception

Notes: LawZero is a Quebec nonprofit organization advancing research and developing technical solutions for safe-by-design AI. Founded by Yoshua Bengio.

Future of Finance: How AI Is Rewiring Private Market Workflows

Notes: Includes tickets to the Kraken vs. Maple Leafs hockey game.

Fireside Chat with David McJannet, ex-CEO HashiCorp

Notes: Mitchell Hashimoto’s old company.

🤖 AI Reads

Introducing NVIDIA Nemotron 3.5 Bolt E4B 4GB

Notes: NVIDIA introduced Bolt E4B built on Gemma 4 E4B with Nemotron post-training and outputs up to ~80% fewer tokens.

incoai / Qwen3.8-27B-Splash: Packed for Apple Silicon

Notes: No noticeable quality difference from regular Qwen3.8-27B. I am observing between 2x - 5x faster tokens / second on my M5 Max 128GB depending on the task.

Diffusion at Scale: Introducing Mercury 3

Notes: Diffusion models bring incredible speed, but don’t sleep on diffusion models as decision models. Inception Labs just launched Mercury Decide.

JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs

Notes: Includes authors from Urbana-Champaign, Princeton, and Google.

The State of Context Engineering

Notes: Interesting article sharing much of what’s needed for production systems.

MemoryTax: The Bitter Lesson of Agent Memory

Notes: Interesting paper from a solo author. Relevant for those building long-term memory.

💼 Business Reads

The Rough Economics of Inference Providers

Notes: Great LinkedIn post from Deedy Das of Menlo Ventures.

What It Really Takes to Build a One-Person Company With AI

Notes: I have a few friends building one-person companies now.

What the Swiss Watch Tariff Does to a Price Tag

Notes: For my watch collecting subscribers.

🔀 Other Reads

Where the Romans Fought

Notes: Cool website mapping the geographic scope of Roman battles.

Signing off from Redmond, WA.

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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

DeepSeek-V4.1-Flash-1017

Notes: DeepSeek released a new version of their flash model. You can download the model checkpoint from HuggingFace. DeepSeek Pushes Out Updated AI Model It Says Is Faster, Leaner. Their new Search and Fetch API might be more interesting for subscribers to check out.

Introducing Timestral

Notes: Mistral released a time-series foundation model beating Google’s TimesFM-3 on a few benchmarks.

UltraBERT-Embed-150M

Notes: New dense embedding model. Cohere and others released embedding models recently, so I am not convinced of the SOTA claim.

Google’s Harness2

Notes: A framework for a recursive coding harness from Google.

OpenCode 2.0

Notes: A significant overhaul with a new client-server architecture and developer quality of life improvements. I’m upgrading my multi-agent setup scripts.

GLiClass v4: Multimodal, Instruction-Following Encoders for Every Decision

Notes: Neat model family for the GLiNER ecosystem.

💼 Business Reads

France Struggles to Contain Fallout From Nationwide Protests

Notes: The situation does not seem to be improving across the country. There are debates about intervention. Bond yields are high. Still not a good situation. Read #156 and #154.

AI’s Giant Money Loop Is Drawing Foreign Cash Like 2008

Notes: Fascinating article if you have any interest in South Korea.

AI Risk Is Everywhere and It’s Making Billion Dollar Funds Nervous

Notes: I am increasingly uncomfortable with my own portfolio’s diversification.

🔀 Other Reads

The Teal Programming Language

Notes: Statically-typed dialect of the Lua programming language.

Signing off from Redmond, WA.

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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

Introducing Beam: Reflection’s 501B Open-Weight Model

Notes: An open-weight frontier model competitive with GLM 5.2 and Qwen 3.8-Max from a US-based lab. Nvidia is looking to buy them.

Liquid AI: Announcing d1 With Vision

Notes: A multimodal decision model is interesting.

vLLM Semantic Router: Introducing Decision 1.0: Open Decision Foundation Models

Notes: More decision models.

perplexity-ai / pplx-decider-v1.1-27b

Notes: Yet another decision model.

Andreessen Horowitz Backs Jev Maker at $7.5 Billion Value

Notes: What’s the moat?

EmbeddingGemma 2: An Open, Lightweight Multimodal Embedding Model

Notes: The sequel to the first EmbeddingGemma which does very well.

💼 Business Reads

Canadian Investor, Philanthropist Stephen Jarislowsky Dies at Age 101

Notes: An important pillar of the Montréal finance community has passed away.

France’s Bond Turmoil Raises the Specter of Contagion in Europe

Notes: Hmmm… not good. See #154.

France’s Debt Is Riskier Than 38% of Nation’s Company Bonds

Notes: The last time a US company traded cheaper than US Treasuries was Microsoft for a very brief period.

🔀 Other Reads

Why Do We Steal Things From Restaurants?

Notes: Is there any research showing a decline in the prevalence of this behavior since smartphones and Instagram?

China’s Young People Are Turning to an 84-Year-Old Cambridge Professor

Notes: Fascinating!

Signing off from Redmond, WA.

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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

Gemini 4 Argon: Our Next Era of Frontier Intelligence

Notes: Google released their latest Gemini model and I guess everybody is moving to Rust now.

Argon agents are working on migrating C/C++ codebases to Rust across Google — scaling from tens of thousands of lines in core libraries like re2, libgav1 up to 800K+ lines for the Fuchsia Zircon kernel.

firelex / jeff: Millisecond Decisions, Any Domain

Notes: Neat idea. Jeff places a 0.8B open decision model in front of a bigger LLM.

P.S. It seems the Internet has decided to call Jev-style models: Decision Models.

Liquid AI’s First Decision Models: D1

Notes: Outperforms Jev. Liquid AI is a MIT Startup.

Introducing Strands Decider 2B: A Small, Open Source, Decision Model

Notes: Amazon’s Strands Labs released a decision model.

Introducing Clef: Our Open-Source Decision Models

Notes: Cloudflare released two decision models. At this point… who hasn’t released one?

Cohere’s Embed 5 Is Here

Notes: Cohere’s embedding model is very good. Embed 5 is the latest version.

Introducing topk-embed-v1: Frontier Retrieval at the Lowest Cost

Notes: Another frontier embedding model. 65.16 nDCG@10.

Context Language Models

Notes: These language models natively manage their own context.

💼 Business Reads

Mark

Notes: A profile of Mark Zuckerberg from Colossus.

Toyota Eyes Selling Japan’s Answer to Rolls-Royce Around the World

Notes: Original Centuries were incredible machines. I am excited to see the latest generation exported to the rest of the world.

Slovenia’s .si Domain Sees a Surge in Registrations

Notes: How many of you have yearly domain renewal bills above $500?

🔀 Other Reads

Bose’s First Wired Earbuds in 11 Years Add Noise Canceling

Notes: About time! The Bose QuietComfort 20s were my favorite earbuds of all time. I still have two of them that are suffering from clicking noises when plugged into modern devices that I never had time to figure out.

How to Buy Your First Rolex

Notes: For subscribers looking to get into watch collecting.

Chinese Couples Are Turning Wedding Videos Into Epic Productions

Notes: Some of these videos are spectacular.

Signing off from Redmond, WA.

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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

GLiNER2.5-Decide: An Open-Weight Model for Structured Decision Making

Notes: I’ve loved GLiNER’s classifiers and they now just launched an open-weight model for structured decision-making directly competing with Jev. Incredibly fast on Apple GPUs or old-school CPUs. This model was launched within days of Jev’s release.

Open Models Caught Up With Jev

Notes: Models are not a moat. I shared OpenJev (now known as SemIf) last weekend. There are even more models released now.

Contrastive-LM / CLM-v0.1-8B

Notes: Much much faster than Jev. Open-weight too.

jaredpalmer / kev: Jev-Like Family of Decision Models

Notes: From Jared Palmer. Built on top of Qwen3.5/3.8. Make your own Jev.

innylarouge / jevlike: Train a Small Model Like Jev

Notes: Another make your own Jev.

Introducing MiMo-V2.6 series

Notes: Xiaomi’s MiMo model is competitive with Fable and Opus on a number of benchmarks.

FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution

Notes: I’ve already seen variations, but it’s still neat! Here’s a demo.

Memory Compression for High-Fanout Agent Sandboxes

Notes: Interesting paper for those building deep research-style agents.

Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity

Notes: Should be easy to implement in Pi.dev if anyone has free time / tokens.

doximity / mvcc: Compile and Run CUDA on Apple M4-M5 GPUs

Notes: If this project is successful, it would improve model inference compatibility on Apple’s M chips.

💼 Business Reads

What China’s Vibe-Coding Capital Can Tell Us About the AI Boom

Notes: Just vibes. Good for getting a feel of what’s happening on the ground.

Who’s Ahead in the Global AI Talent Race?

Notes: Results are not surprising. AI is a story of two countries.

Latent.Space

OpenRouter: from Seed to Stripe — with OpenRouter’s Alex Atallah & AMP’s Anjney Midha

From the earliest days of open-weight models to becoming the neutral routing layer for more than 10 million developers, OpenRouter is one of the clearest bets that the future of AI will be multi-mode…

Listen now

a month ago · 34 likes · 6 comments

Financial Stress Spreads in Europe as Leverage Mounts, BCG Says

Notes: Any insights from my European subscribers?

Signing off from Redmond, WA.

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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

Introducing System One Models & Jev

Notes: Jev definitely captured the ML community’s attention this past week. It’s effectively a “general-purpose” classifier, and that generalization is very valuable. I believe too little attention and investment are going toward improving classifiers, and developers in the applied AI space too often reach for LLMs when a classifier with embedding features (like CatBoost) or a fine-tuned DeBERTa model would do the trick.

Here’s a fun story to share. My wife and I often split expenses from our individual credit cards. I built a small web app and added a CatBoost model to automatically decide which transactions (using the transaction line item) should be split based on our historical preferences. This app took a monthly chore from 1 hour to 1 minute. The model takes the transaction description (Amazon is very descriptive, thankfully) and other features, and is accurate enough that my wife was impressed (this is not a frequent occurrence with my side projects 🥹).

Whether Jev is truly novel or groundbreaking is up for debate. See this Reddit thread and Laya. You can get close to Jev’s benchmark performance with Qwen3.5 4B as well; see OpenJev for a demo.

Qwen3.8-Omni-Flash: Omni Senses

Notes: Another impressive model from the Qwen team.

Mozilla’s Inaugural ‘State of Open Source AI’ Report Is Here

Notes: The gap between open-weight AI models and frontier models is shrinking. Mozilla estimates the gap at just 4 months.

Harbor: Evaluate Agents in Sandboxed Environments

Notes: Open source project for evaluating agents from the creators of Terminal-Bench.

DietrichGebert / ponytail: Makes your AI agent think like the laziest senior dev in the room.

Notes: I love this project! Run the audit command on your repositories.

💼 Business Reads

One of AI’s Fiercest Critics Says All the Doom Talk Is ‘Meant to Distract Us’

Notes: Wired’s interview with Timnit Gebru.

The Next Life of an Iconic New York Skyscraper Is 37 Luxury Homes

Notes: My favorite building in NYC. I make a pilgrimage to NoMad every visit.

Some of the Richest Americans Are Hiding in Plain Sight

Notes: I bought the book, The Everywhere Millionaire.

A $5,000 Private Lounge With BMW Tarmac Drop-off Is the New Airport Luxury

Notes: I guess paying for private terminal access might be worth it at highly congested airports where flight delays are frequent. But I don’t think these terminals are worth it in Seattle or similar-sized, well-run airports. At SEA, I rarely spend more than 15 minutes going from car to gate and the lounges are not crowded.

🔀 Other Reads

One Hot Day Can Change Your Body for Years

Notes: Be careful out there! Heat stroke is no joke.

Pandoc Templates

Notes: Neat website collecting templates for pandoc.

wouterdebie / davit: A Native macOS UI for Apple’s Platform

Notes: For those who want truly lightweight hardware isolation on Apple’s silicon.

Signing off from Redmond, WA.

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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.

Apéro & Intellect #15 - NYC Edition & Apéro & Intellect #16 - Boston Edition

NYC has 1 spot left. There’s a few more available for Boston. It’s going to be fun!

Apply on Lu.ma

🤖 AI Reads

CIS 6280: World Models

Notes: From UPenn, covering various World Model architectures and applications.

nvidia / nemotron-3.5-asr-streaming-0.6b

Notes: SOTA ASRA model from Nvidia. Ready for commercial use.

kyutai-labs / pocket-tts

Notes: Fast TTS. I previously shared Kyutai’s work in #117.

LFM2.5-230M: Built to Run Anywhere

Notes: A very small model from Liquid AI released in June.

RyanCodrai / turbovec: A Vector Index Built on TurboQuant, Written in Rust with Python Bindings

Notes: I haven’t tried it yet, but the project seems promising. Claims to be much faster than Meta’s FAISS.

herdrdev / herdr: The Runtime Your Coding Agents Live On

Notes: An open source agent multiplexer if you don’t already have your own setup with tmux.

💼 Business Reads

Memoirs of Extraordinary Popular Delusions and the Madness of Crowds by Mackay

Notes: A very old book (1841) but relevant to read today.

Z.AI’s Cash Burn Drives Alibaba-Like Hong Kong Fundraising Spree

Notes: If there’s demand…

Kalshi Moves to Expand Into US Stock and Commodity Markets

Notes: Huh… interesting.

Where to Invest $10,000

Notes: Always a fun article series to read.

🔀 Other Reads

WA K12 Academy Atlas

Notes: Surprisingly beautiful website for parents in WA.

The Website Specification

Notes: Point your coding agent at this website to improve code quality.

WSL 2 Is Getting Faster: Windows File System Access

Notes: This provides a significant quality of life improvement and increases the iteration speed for your coding agents if you run them in WSL.

Signing off from Redmond, WA.

Apéro & Intellect #15 - NYC Edition & Apéro & Intellect #16 - Boston Edition

Limited spots available on September 17 and September 18th!

Apply on Lu.ma

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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.

Apéro & Intellect #15 - NYC Edition & Apéro & Intellect #16 - Boston Edition

Limited spots available on September 17 and September 18th!

Apply on Lu.ma

🤖 AI Reads

OpenAI GPT-6 Astra

Notes: My first impression is that GPT-6 Astra pursues the incorrect path less eagerly than GPT5.6 Sol, and stops to ask for direction. Almost closer to Opus 4.7 in ability to pair with a collaborator. Opus 4.8 & 5 went off the rails trying to do too much.

Introducing Gemini 3.8 Flash and 3.8 Flash Cyber

Notes: Google upgraded the flash model family and has a cybersecurity-specific model.

TimesFM-3: A Zero-Shot Foundation Model for Multivariate Forecasting

Notes: T-1 before quant firms integrate this model into their trading stack.

Pokee-Isaac 28B: A 10M-Token Context Efficient Agentic Model

Notes: This is coming from a Seattle-based AI startup. Interesting performance with a proprietary model architecture.

S1-mini by Superwhisper

Notes: The STT model chain is getting more sophisticated. Here’s a 0.6B-parameter text normalizer for speech-to-text output that cleans up filler words and normalizes transcripts to clean outputs.

💼 Business Reads

AI Is Rewiring South Korea’s Careers, Dating and Culture

Notes: The Korean culture and stock market had quite a ride during this AI boom.

Corporate America Is Getting Hooked on Open-Source AI & Kai-Fu Lee: China Will Win the AI Race for Reach

Notes: The rest of Corporate America is finally catching up. Also, a related podcast from the Mishal Husain Show with Kai-Fu Lee (whose latest company models I’ve shared in this newsletter multiple times). I completely agree with Kai-Fu Lee’s statement that Chinese talent was undervalued.

AI Startup Micro1 Seeks to Topple Google’s Bid for Spirit Data

Notes: The only moats left are data and distribution. Not surprising others are jumping in to buy Spirit’s business data.

The Ex-Beer Baron Turning Windshield Repair Into a $7.8 Billion-a-Year Empire

Notes: Windshields are high tech now.

Prime Minister Carney Launches Digital Transformation Canada & Ottawa Taps Patrick Pichette to Lead New Digital Transformation Agency

Notes: An excellent initiative with a great choice for a leader. 🇨🇦

How to Fund a Startup in Canada: Every Option, and What Each One Costs You

Notes: For my Canadian subscribers, this article is fantastic and covers the funding options nationwide. 🇨🇦

🔀 Other Reads

Hong Kong’s Iconic ‘Cha Chaan Teng’ Diners Are Going Global

Notes: Nothing new for the foodies on the list; my favorite Cha Chaan Teng’s are in Seattle and New York City.

A Controversial Alzheimer’s Surgery Is Said to Reverse Symptoms

Notes: For the surgeons on the list, I would love to hear your thoughts on this treatment.

Signing off from Redmond, WA.

Apéro & Intellect #15 - NYC Edition & Apéro & Intellect #16 - Boston Edition

Limited spots available on September 17 and September 18th!

Apply on Lu.ma

🚀 Resetting and Relaunching Rami’s Readings

Rami’s Readings was quiet for a few months. There wasn’t one dramatic reason, just a lot of life piling up at once.

My family was knocked out for a whole month and a half with a variety of nasty illnesses from the petri dish that is “daycare.” Shortly after, a mix of travel and vacations took up most of the summer, and with work becoming unusually intense, it was incredibly hard for me to get back to the newsletter. The weekly cadence slipped for a few months, but this pause also gave me an opportunity to think about the newsletter and prepare a few things in celebration of #150.

So rather than simply picking up where I left off, I’m using #150 as a reset and a relaunch.

The biggest change is that Apéro & Intellect is getting a proper home with its own website and events hosted on Luma, fulfilling the plan I shared earlier this year. I’ll be hosting Apéro & Intellect #14 in Greater Seattle on Thursday, September 3rd.

The newsletter will continue its focus on AI, business, and other reads, but with a more concerted effort to share product-related news: specifically, new AI products or features that catch my eye.

Lastly, expect a few guest writers and interviews to appear in the newsletter on occasion. I’m excited about this part, and I’ll share more details a bit later in the year.

Consider the last few months an intermission. Now, on to the next iteration.

👋🏼 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, The VC Corner and from the AI in Action Conference in Montréal. 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.

If any of you would like to get a copy of the slides, feel free to respond directly to this newsletter.

🤖 AI Reads

Broadly speaking, the past few months saw smaller models steadily eroding the Big AI labs’ advantage across many benchmarks.

Qwen3.8-27B

Notes: I used Qwen3.5 all the way through to the latest Qwen3.8 version for coding locally. The Qwen family of models have continued to show that you don’t need the biggest model for every workload.

Qwen3.8-Flash-Next

Notes: This model is interesting because it’s showing a new avenue for improving quality while reducing memory requirements. It uses N-gram embedding which is offloadable from GPU RAM to system RAM or even SSD. Many Reddit and HuggingFace threads showing varying levels of success running this model locally on more and more constrained RAM.

DeepSeek-V4-Flash-0731

Notes: DeepSeek continues to lead on the economics of running LLMs at scale. Their strategy appears to be positioning themselves to be the cost leader while remaining within reach of the top of the benchmarks.

GLM-5.3 Flash

Notes: A SOTA open-weights model that you can run quantized on the new Mac Ultra 256GB (or better on the yet to-be-priced 512GB version. By the way if anyone wants to buy me a Christmas gift 😉). The model was floating around on various benchmarks as Ox Alpha before it was revealed to be Z.ai’s GLM model.

Muse Glimmer

Notes: While Muse Glimmer was overshadowed by Qwen and GLM, it stands out as Meta’s first notable model release since Llama 3. It’s also the first major release since Alexandr Wang took on AI at Meta.

LFM2.5-2.6B

Notes: Liquid AI also released an agentic model that can run on a phone and yields some impressive results against competitors models of the same size. Liquid AI is an MIT startup that I track in this newsletter.

In terms of the business of AI, there’s been a handful of important acquisitions and upcoming IPOs. The following are the most important:

Stripe Agrees to Acquire OpenRouter to Help Businesses Optimize Token Routing and Usage

Notes: This acquisition is interesting because OpenRouter is effectively a marketplace for exchanging dollars for tokens across many model & inference providers. If we expect the market to continue to explode, then it makes sense for Stripe to expand into that payment infrastructure and payment processing market.

Nvidia Agrees to Buy Open Source AI Platform Hugging Face For $12.9 Billion

Notes: This acquisition is interesting and important, but not at all surprising. If we want to use a gold-rush analogy then it makes sense for shovel-maker to buy the shovel showcase space… I guess. 😂 My only concern is that the relative neutrality of HuggingFace will erode over time as regulatory and market pressures increase. There isn’t quite an alternative to HuggingFace except GitHub and possibly Docker or Ollama.

Anthropic Confidentially Submits Draft S-1 to the SEC & OpenAI Confidential Submission of Draft S-1 to the SEC

Notes: So much venture capital and other private capital is tied up in the companies that no matter how you feel about the IPOs themselves, one clear benefit is that some of that capital will be unlocked and available for the next generation of startups. If you time it right and are raising your Series A in 6-12 months from now, there could be a lot of dry powder ready to be deployed. 😉

A few subscribers always ask about what hardware I’m running. Recently, I acquired a M5 Max and a AMD Ryzen AI Max+ 395 to complement my RTX 4090. They are both interesting machines. Day zero support for both MLX & ROCm is becoming more common, so I’ll be watching for the differences in quality and performance between the different platforms.

New open source projects and papers continue to emerge, exploring a serious wide variety of concepts. I’ll share more in future newsletters, but here are a few worth highlighting:

mathomhaus / guild: Shared Context, Memory, and Task Coordination Across AI Coding Agents

HumanSignal / Adala: Autonomous DAta (Labeling) Agent Framework

CopilotKit / CopilotKit: The Frontend Stack for Agents & Generative UI

drumih / turbo-fieldfare: Gemma 4 26B-A4B inference in ~2 GB of RAM on any M-series MacBook

AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces

Atlassian Rovo Exfiltrates Data, Bypassing Controls

💼 Business Reads

It’s been a strange few months for businesses across North America. I won’t try to recap everything, but here’s a few things happening elsewhere:

China’s AI Boom Has a Favorite Bar

Notes: Shared by one of my favorite subscribers, Marielena. I would say it’s not unexpected for spaces for discussion to emerge. In Seattle, there is the AI House near the waterfront.

The Unlikely Place at the Center of China’s AI Boom

Notes: Not quite where I expected, but interesting to read.

Asia’s 20 Richest Families of 2026

Notes: Required reading in preparation for the Crazy Rich Asian sequel.

The List Tracking Canada’s Tech Exodus

Notes: Yep, but there’s still great Canadian startups and talent!

People Without Kids Are Hacking the College Savings Account

Notes: I can’t agree more with this article. Open a 529 plan as soon as you can.

Signing off from Montréal, QC.

Apéro & Intellect #14 - Rami’s Readings 150th Edition Launch Party

Apply on Lu.ma

👋🏼 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.

👋 I’ll be Speaking at AI in Action in Montréal!

I’m heading back to Montréal on May 5th for the AI in Action conference.

I am super excited for this conference as the AI conversation has moved from what’s possible to what’s actually working. In my talk, I’ll cover the latest trends in LLMs and SLMs, orchestration, and multi-agent systems. We’ll also explore how to prepare for the next major moment in the AI landscape.

I’m especially interested in comparing notes with founders, executives, and builders on what it takes to move from impressive demos to scaled products: distribution, evaluation, cost, UX, and the operational work that turns AI experiments into real workflows and businesses.

If you’ll be at AI in Action in Montréal on May 5th, or nearby that week, reply to this newsletter or on LinkedIn. I’d be glad to connect. If you’d like a discounted ticket, please reach out to me.

🤖 AI Reads

DeepSeek V4 — Almost on the Frontier, a Fraction of the Price

Notes: More on DeepSeek V4 which we covered last week. Huawei Ascend was supposedly used to train the model indicating that NVIDIA’s GPU stranglehold is weakening.

Mike - The Open Source Alternative to Harvey and Legora

Notes: I assume the naming is inspired from Suits. Great traction on GitHub for 3 day old release.

Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses

Notes: Autoresearch for the coding agent harness itself.

Recursive Multi-Agent Systems

Notes: A very interesting paper, but not particularly relevant for system builders, more for AI architecture research.

browser-use / browser-harness: Self-Healing Harness That Enables LLMs to Complete Any Task

Notes: Another tool for chrome browser support for agents.

h4ckf0r0day / obscura: Headless Browser for AI Agents and Web Scraping

Notes: Yet another new tool for browser support. This one takes a different approach to make a completely headless experience.

dirac-run / dirac: Coding Agent Singularly Focused Efficiency and Context Curation

Notes: Neat Uses ASTs and Hash Anchored Edits. I would expect other popular coding agents to support these pretty soon.

shanraisshan / claude-code-best-practice

Notes: Neat repository (not mine) collecting much of the Claude Code plugins and workflows I shared in this newsletter over the past few months.

💼 Business Reads

Two NJ Malls Separated by Just Four Miles — and Very Different Fates

Notes: The high-end mall in Bellevue is also doing very well.

Most Prediction Market Traders Are Losing Money While Bots Rack Up Gains

Notes: I’m not surprised at all.

Mahjong Is the Third Place of the Moment for a Burned-Out Generation

Notes: I love the proliferation of new kinds of third places, loving that it’s Mahjong.

Signing off from Redmond, WA.

👋🏼 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.

👋 I’ll be Speaking at AI in Action in Montréal!

I’m heading back to Montréal on May 5th for the AI in Action conference.

I am super excited for this conference as the AI conversation has moved from what’s possible to what’s actually working. In my talk, I’ll cover the latest trends in LLMs and SLMs, orchestration, and multi-agent systems. We’ll also explore how to prepare for the next major moment in the AI landscape.

I’m especially interested in comparing notes with founders, executives, and builders on what it takes to move from impressive demos to scaled products: distribution, evaluation, cost, UX, and the operational work that turns AI experiments into real workflows and businesses.

If you’ll be at AI in Action in Montréal on May 5th, or nearby that week, reply to this newsletter or on LinkedIn. I’d be glad to connect. If you’d like a discounted ticket, please reach out to me.

🤖 AI Reads

OpenAI Introduces GPT‑5.5

Notes: OpenAI is putting the heat on Anthropic in the agentic coding market.

DeepSeek V4 Preview Release

Notes: DeepSeek is back with a release, but so far seems to be mostly on-par with other open weight models. The more interesting choices they made seem to be enabling their price war strategy. See read below.

DeepSeek Slashes Fees for New AI Model in Chinese Price War

Notes: I covered the DeepSeek price war previously in #64. They’re back at it again!

AgenticQwen: Training Small Agentic Language Models with Dual Data Flywheels for Industrial-Scale Tool Use

Notes: The Qwen team has an interesting paper out.

earthtojake / text-to-cad: An Open-Source Harness for Generating CAD Models

Notes: The demo looks great, but I am not sure how useful this really is for engineering firms. Anyone care to comment?

vercel-labs / agent-browser: Browser Automation CLI for AI Agents

Notes: Giving this a try tomorrow to see if it performs better than chrome-devtools-mcp.

Signing off from Redmond, WA.

👋🏼 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.

🤖 AI Reads

Introducing Claude Opus 4.7

Notes: So much more expensive token wise, but quality is improved.

Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All

Notes: It’s already replaced Qwen3.5 27B for my coding agent work locally. Reddit Thread, Unsloth Quantized Model on HuggingFace.

Thunderbolt — AI You Control

Notes: Important because it is being built by Mozilla. Blog Post.

rowboatlabs / rowboat: AI Coworker Turns Work into a Knowledge Graph

Notes: This project looks a lot like how I use Claude Code + WorkIQ at work.

forrestchang / andrej-karpathy-skills

Notes: Andrej Karpathy’s observations on LLM coding pitfalls fixed by a single Claude.md.

Cybersecurity Looks Like Proof of Work Now

Notes: Interesting article given all the Mythos discussions.

PurpleAILAB / Decepticon: Autonomous Red Team Testing

Notes: Open source projects equivalent to Mythos.

zubair-trabzada / geo-seo-claude

Notes: Interesting project for Generative AI SEO.

Cocoon-AI / architecture-diagram-generator

Notes: System architecture diagrams as a Claude Code skill.

thedotmack / claude-mem: Persistent Memory Compression System

Notes: I shared Claude-Mem previously, but worth a reshare.

tirth8205 / code-review-graph: Local Knowledge Graph for Claude Code

Notes: Uses ASTs of your project. Not sure how much of a lift this would give if you have LSP correctly configured.

nadimtuhin / claude-token-optimizer

Notes: A project documentation system for Claude that reduces token usage.

drona23 / claude-token-efficient: Keeps Claude Responses Terse

Notes: Implemented as a single CLAUDE.md file.

JuliusBrussee / caveman: Why Use Many Tokens When Few Do Trick

Notes: Similar to the above, but funnier and with a more complex implementation (skills and MCP server).

💼 Business Reads

Stanford’s 2026 AI Index Report

Notes: A very important report to read every year. Great insights!

AI adoption is spreading at historic speed, and consumers are deriving substantial value from tools they often access for free.

Your Startup Is Probably Dead On Arrival

Notes: From the famous Steve Blank.

Signing off from Redmond, WA.

👋🏼 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.

🤖 AI Reads

Project Glasswing a.k.a. Claude Mythos

Notes: Rightfully caused a panic in the markets with respect to cybersecurity. Ignore the hype and criticism on all sides. Given the quality level of coding agent automation driven by frontier LLMs since November 2025, it’s a no-brainer that we’re entering a new cybersecurity era. Also read Wired.

multica-ai / multica: Open-Source Managed Agents Platform

Notes: Competing open source project to Claude Managed Agents.

getcompanion-ai / feynman: Four Bundled Research Agents, Dispatched Automatically

Notes: Kind of similar to Model Council.

Meta-Harness: End-to-End Optimization of Model Harnesses

Notes: Autoresearch for your autoresearch.

garrytan / gbrain: Garry's Opinionated OpenClaw / Hermes Agent Brain

Notes: Garry Tan’s opinionated OpenClaw setup.

The Art of Building Verifiers for Computer Use Agents

Notes: This is a really interesting paper for Microsoft Research (not my department).

LFM2.5-VL-450M: Structured Visual Intelligence, Edge to Cloud

Notes: New model from MIT’s Liquid AI.

Mistral Launches Voxtral 4B TTS 2603

Notes: Outperforms ElevenLabs’s closed source model.

💼 Business Reads

Wall Street Banks Try Out Anthropic’s Mythos & Bank of Canada, Major Lenders Met on Anthropic AI Cyber Risk

Notes: I hope there will be a renewed pushed for improved cybersecurity everywhere, not just banks.

Japan Bets $16 Billion to Propel Rapidus in Global AI Chip Race

Notes: More competition for Nvidia is good.

Hermès Sticks With Plan to Raise Output

Notes: This is honestly a surprise to me, but I guess demand is so high that bad vibes wouldn’t lower demand for several years. P.S. Hermès’ flagship store in Japan is something else.

🔀 Other Reads

rtk-ai / rtk: CLI Proxy That Reduces LLM Token Consumption by 60–90% on Common Dev Commands

Notes: Hooks into coding agents like Claude, OpenCode, Codex, etc. to remove unnecessary tokens from common commands like ls, cat, grep, etc. One of my favorite things about the coding agent revolution is the revival of interest in terminal-based development and the explosive growth of CLI tools, (e.g. ripgrep, bat, exa, etc. all new-to-me).

coast-guard / coasts: Localhost Service Isolation and Orchestration for Git Worktrees

Notes: Another approach for coding agent isolation.

Freestyle: Sandboxes for Coding Agents

Notes: YC startup focused on coding agent isolation and sandboxing for enterprise.

Signing off from Redmond, WA.

👋🏼 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.

👋🏼 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.

🤖 AI Reads

Thoughts on OpenAI Acquiring Astral and uv/Ruff/Ty

Notes: I’m a little worried about uv. It truly solved so many problems with the Python packaging ecosystem that I would hate to go back. The acquisition seems to be in the same spirit as Anthropic’s acquisition of Bun. Is this the beginning of a new era of developer tooling funding and acquisition? Which VC-backed devtools startup would you bet on next?

Markdown as a Protocol for Agentic UI

Notes: For those building chat applications, this is an interesting approach to Generative UI.

pablodelucca / pixel-agents: Pixel Office

Notes: Is everything just a game now?

ComposioHQ / agent-orchestrator: Orchestrator for Parallel Coding Agents

Notes: Yet another coding agent orchestration framework.

hardikpandya / stop-slop: A Skill File for Removing AI Tells From Prose

Notes: A skills version of Signs of AI Writing from Wikipedia?

Mistral Forge: Build Your Own Frontier Models

Notes: Similar to Thinking Machines Lab and other startups.

💼 Business Reads

The AI Shift: Is This the ‘Take Off’ Moment for AI Agents?

Notes: Distribution is ever more important. I recorded a podcast with Framework.vc that touched on this topic.

Welcome to Paris, the City That Said No to Cars

Notes: I had the chance to experience many of these improvements in Paris last year. It definitely felt significantly more human-scale than my previous visit over a decade ago.

🔀 Other Reads

‘Project Hail Mary’ Review: Ryan Gosling Is Lost and Found in Space

Notes: A fantastic movie, I loved it! It was adapted from a novel written by Andy Weir, so it was similar in spirit to The Martian.

Signing off from Redmond, WA.

👋🏼 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.

👋 I’ll be a Mentor at Startupfest Montréal!

I am excited to be returning to Montréal (my hometown) this summer as a mentor at Startupfest.

Startupfest is one of the premier events in the Montréal startup ecosystem. It’s been around for longer than my career in tech. It’s also an excellent excuse to fly up to the city and experience our glorious patio summers!

Looking forward to connecting with founders and entrepreneurs building the next generation of AI startups, sharing lessons from scaling products from zero to global scale, and contributing back to the startup and tech community that launched my career.

If you’ll be attending Startupfest or will be in Montreal around the July 8 - 10 and would like to connect, feel free to reply to this newsletter. I also have a 15% discount code available for tickets.

🤖 AI Reads

bytedance / deer-flow: An Open-Source SuperAgent Harness That Researches, Codes, and Creates

Notes: Interesting project from ByteDance. Similar to OpenClaw, Autoresearch, and other clones.

LLM Architecture Gallery

Notes: Very cool poster from Sebastian Raschka.

I spent way too much time Claude Coding this past week… Just a short interjection, none of the spec frameworks, skills, plugins, etc. that I am sharing solve the problem of “taste”. In other words, what should be built when anything can be built.

Fission-AI / OpenSpec: Spec-Driven Development (SDD) for AI Coding Assistants

Notes: I shared in a previous newsletter, but worth re-sharing. Lighter than Spec Kit.

github / spec-kit: Toolkit to Help You Get Started With Spec-Driven Development

Notes: Supports almost all the coding agents (GitHub Copilot, Claude Code, Codex, OpenCode, etc).

bmad-code-org / BMAD-METHOD: Breakthrough Method for Agile Ai Driven Development

Notes: More opinionated.

obra / superpowers: An Agentic Skills Framework & Software Development Methodology That Works

Notes: Neat collection of skills.

obra / double-shot-latte: Claude Code Plugin That Automatically Evaluates Whether Claude Should Continue Working

Notes: Claude for Claude.

garrytan / gstack : Garry Tan’s Exact Claude Code Setup

Notes: Garry Tan runs YC.

777genius / claude-notifications-go

Notes: Customize Claude Code notifications.

💼 Business Reads

Services: The New Software

Notes: From Sequoia Capital, an excellent read for all founders. If you were to only pick one read this week, pick this one.

AI Startup Moonshot Snags Funds at $18 Billion Valuation

Notes: Kimi K2.5 is pretty good.

Asia Tech Review

China is obsessed with OpenClaw

China is obsessed with OpenClaw, the platform for creating AI agents that can carry out tasks…

Read more

8 months ago · 1 like · Jon Russell

🔀 Other Reads

Hammerspoon

Notes: Much much more flexible than AppleScript.

Signing off from Redmond, WA.

👋🏼 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.

🎙️ Podcast with Framework Venture Partners

I recently sat down with Mikhail Karmali from Framework Venture Partners for a wide-ranging conversation recorded at my kitchen table. 🙂 We discussed how the LLM era is compressing traditional SaaS models faster than many appreciate. The cost and complexity of building software has fallen dramatically, yet the enterprise buy-vs-build bar has never been higher. Listen to it on Spotify!

I would love to hear your thoughts, reply directly to the newsletter.

✅ Innovation Governance Program Level 1

I'm pleased to share that I've completed Level 1 of the Innovation Governance Program (iGP) from the Council of Canadian Innovators | Conseil canadien des innovateurs (CCI).

The program focuses on the realities of governing innovation-driven companies, for example; how boards support founders, oversee risk in fast-moving technology environments, and provide strategic guidance while maintaining effective oversight. It was valuable to learn alongside a strong cohort of operators and leaders from across Canada's innovation ecosystem.

Governance is an increasing area of focus for me and I'm particularly interested in the role boards can play in shaping the responsible development and deployment of AI technologies, ensuring innovation is paired with strong oversight, safety, and long-term strategic thinking.

I'm open to board or advisory board opportunities, particularly with companies building in AI, software, and emerging technologies. If you're building a board or governance structure and think my background could be helpful, I'd welcome the conversation.

🤖 AI Reads

Alibaba’s Qwen3.5-Medium Models Offer Sonnet 4.5 Performance

Notes: I was testing qwen3.5-35b-a3b this weekend. Full GPU offload with a decent context size is possible on a RTX 4090. It’s shockingly good. It may become my default coding model. The 35B-A3B model seems to be close enough to Claude Opus 4.6 per this study. Simon Willison shared news of potential personnel changes at Qwen. Unsloth posted an article on How to Run Qwen3.5 Locally. Find the models on HuggingFace. Read this X Post from Alibaba.

karpathy / autoresearch: AI Agents Running Research on Single-GPU NanoChat Training Automatically

Notes: Just goodness from Andrej Karpathy.

Introduction to LLM Post-Training Techniques

Notes: From Maxime Labonne at Liquid AI. Great ML scientist to follow.

Olmo WebGPU: Run Olmo Hybrid Directly in Your Browser

Notes: Always nice to see Edge AI projects being released by big research labs like AI2.

ruvnet / ruvector: RuVector is a Self-Learning, Vector Graph Neural Network, and Database

Notes: Interesting open source project.

builderz-labs / mission-control: Open-Source Dashboard for AI Agent Orchestration

Notes: I haven’t tried it yet because I already built my own control plane for my AI agents.

🔀 Other Reads

A Wave of Unexplained Bot Traffic Is Sweeping the Web

Notes: Hey look Ma, I made it! The vast majority of the traffic to my website and Substack over the last 90 days is from Lanzhou.

the peculiar case of japanese web design

Notes: An old article (2022), but a super fun read.

Signing off from Redmond, WA.

👋🏼 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.

🎙️ Podcast with Framework Venture Partners

I recently sat down with Mikhail Karmali from Framework Venture Partners for a wide-ranging conversation recorded at my kitchen table. 🙂 We discussed how the LLM era is compressing traditional SaaS models faster than many appreciate. The cost and complexity of building software has fallen dramatically, yet the enterprise buy-vs-build bar has never been higher. Listen to it on Spotify!

I would love to hear your thoughts, reply directly to the newsletter.

🤖 AI Reads

The 2-Minute Claude Code Upgrade You’re Probably Missing: LSP

Notes: It’s unsurprising that a decade’s worth of investment in the Language Server Protocol would enable code search to outperform basic grep and find operations.

microgpt

Notes: Another fantastic article from Andrej Karpathy.

Introducing Mercury 2: Fastest Reasoning LLM, Powered by Diffusion

Notes: Mercury 2 is the current SOTA diffusion model.

LFM2-24B-A2B: Scaling Up the LFM2 Architecture

Notes: New model release from Liquid AI (MIT).

gsd-build / get-shit-done: Meta-prompting, context engineering and spec-driven development system for Claude Code

Notes: Neat project for Claude Code.

💼 Business Reads

Claude Code and the Great Productivity Panic of 2026

Notes: In tech, no one’s sleeping anymore.

AI Is Proving a 100-Year-Old Prediction True

Notes: An interesting thought experiment on leisure time.

SatPost by Trung Phan

How Hermès Sells "Time"

Thanks for subscribing to SatPost…

Read more

3 years ago · 135 likes · 19 comments · Trungphan2

🔀 Other Reads

Let’s Discuss Sandbox Isolation

Notes: Highly relevant article detailing the differences between each sandbox isolation method. Personally, I use full virtual machines for each of my coding agents with an isolated non-root user account within that VM. Each VM can’t connect to anything in my local network (enforced at multiple layers), and only has access to specific IPs pulled from DNS records that are refreshed every day. I took inspiration from Claude Code’s .devcontainer, but went further than Docker containers. Docker has a microVM feature in “experimentation” mode, but they aren’t easily customizable, so full VMs it is!

The Latest Destination for Bagels Isn’t New York. It’s Singapore

Notes: Another reason to visit Singapore!

Signing off from Redmond, WA.