Rami's Readings #44
The latest on Coding, LLMs, Prompt Attacks, Nvidia, Rogue Amoeba, Edge AI, Deep Learning & Materials Science, Startups & Recessions, Obesity Pay Gap, Chrome & Bluetooth Exploits, and more.
Welcome to Rami’s Readings #44 - a weekly digest of interesting articles, papers, videos, and X (formerly known as Twitter) 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.
Vacation Notice: From Dec 16, 2023 to Jan 13, 2024, 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
A Coder Considers the Waning Days of the Craft
Hacking is forever.
MEDITRON-70B: Scaling Medical Pretraining for Large Language Models
Notes: New open-source LLM for Medical applications from EPFL!
Overall, MEDITRON achieves a 6% absolute performance gain over the best public baseline in its parameter class and 3% over the strongest baseline we finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of Med-PaLM-2.
Notus 7B: A New Open Source LLM, Fine-tuned Using Direct Preference Optimization and AI Feedback Techniques
Notes: Another new open-source model.
ChatGPT’s One-year Anniversary: Are Open-Source Large Language Models Catching up?
Notes: New paper from NTU in Singapore. I don’t need to read the paper to say yes unequivocally.
Google Researchers’ Attack Prompts ChatGPT to Reveal Its Training Data
Notes: This technique for extracting training data from many LLMs is intriguing and capturing much attention this week. Arguably much of the efficacy in Large Language Models (LLMs) stems from memorization; fully solving training data extraction is not reasonable. The paper on ArXiv.
How Jensen Huang’s Nvidia Is Powering the AI Revolution
Notes: Another great long read about Huang.
Audio Hijack for Mac Includes Speech to Text Transcription Powered by OpenAI’s Whisper
Notes: I am a fan of Rogue Amoeba. In my newsletter on Nvidia’s P/E ratio, I hypothesized companies leveraging open-source ML models will have their inference costs trend to zero.
There will be another category of companies that will distill or fine-tune existing machine learning models to suit their needs. This category will be larger than the previous one. Only some companies need to train their own models. This category’s training and inference capacity needs will be modest. They will also be the primary beneficiaries of open-source AI models and the optimizations making AI inference on edge devices a reality. Their costs will trend to zero. 🤔 Why buy entire data centers of expensive hardware when many open-source models with some fine-tuning suit your needs? Rent GPUs just enough to get by. Run inference on edge devices as much as possible. [From Rami’s Readings #23]
Rogue Amoeba is doing exactly thing… building incredibly valuable features (transcription) into their software with minimal ML training (not sure if they even fine-tuned Whisper at all) and zero inference costs (running locally on the fabulous Apple Silicon). #Edge-AI
It’s also completely private. When you use Transcribe, everything happens right on your Mac. That means your data is never sent to the cloud, nor shared with anyone else. [From Rogue Amoeba]
In a First, Beijing Court Grants Copyright to AI-Assisted Artwork
Notes: My friends in China, can you please share any additional local news coverage?
iA Writer can now track what you or ChatGPT wrote
Notes: Great UX approach for tracking authorship if you plan to use ChatGPT as a writing copilot.
Scaling Deep Learning for Materials Discovery
Notes: Fabulous!
💼 Business Reads
Great Recession Babies: How Are Startups Shaped by Macro Conditions at Birth?
Notes: A commonly heard adage from the startup community in Silicon Valley is that 'the best startups are founded in a recession.' This seems to be confirmed by this paper.
The Obesity Pay Gap Is Worse than Previously Thought
Notes: The Economist found men also face a large obesity-wage penalty. From personal experience (I actively lost 50 lbs of body fat in 4 years), I am not surprised. I vividly remember the difference in tone and interactions before and after.
The conclusion—that well-educated workers in particular are penalised for their weight—holds for both sexes. Moreover, the higher your level of education, the greater the penalty. We found that obese men with a bachelor’s degree earn 5% less than their thinner colleagues, while those with a graduate degree earn 14% less. Obese women, it is true, still have it worse: for them, the equivalent figures are 12% and 19%, respectively.
Bona Vacantia
Notes: What?!? Write wills people!
The king is profiting from the deaths of thousands of people in the north-west of England whose assets are secretly being used to upgrade a commercial property empire managed by his hereditary estate, the Guardian can reveal.
🛡️CyberSecurity Reads
PSA: Update Chrome Browser to Avoid an Exploit Already in the Wild
Notes: Update your iOS & Mac devices as well.
New Bluffs Attack Lets Attackers Hijack Bluetooth Connections
Notes: Distributing a fix for this will be fairly difficult.
Google Drive Users Complain of Missing Files, Months of Data Disappearing
Notes: More of a data/backup read. Also, why I still use a local NAS.
That is all for this week. Signing off from Redmond, WA.