Discover more from Rami’s Readings

I curate the best papers, tweets, and articles I have read during the week and send them to you in a weekly digest. Less widely shared articles, more reads missed among the noise. Learn about Abloh's 3% Rule for Creativity, GenAI, Culture, Tech, etc.

Rami's Readings #19

The latest on AI anxiety and predictions, LLMs in the medical field, Passkeys, world trade data, WEF’s Future of Jobs report, and more.

Rami Sayar

May 7, 2023

Welcome to Rami’s Readings #19 - a weekly digest of interesting articles, videos, Twitter threads from my various sources across the Internet. Expect a list of reads covering technology, business, culture, fashion, travel and more. Learn about what I do at ramisayar.com/about.

Are you enjoying this newsletter?

If yes, I have a favor to ask you. If you know someone else who would benefit, please forward or share it with them!

Share Rami’s Readings

If you received this newsletter from a friend, subscribe here to get the next one.

Many things happened this week; First Republic got acquired by JPMorgan Chase, King Charles was crowned, and an article about moats around AI/LLMs was leaked (no comment, see disclosure in #7). The following are other notable reads from the week, and do not miss the Future of Jobs Report:

🤖 AI Reads

AI Experts Aren’t Always Right About AI

Notes: Tyler Cowen spot on again. If you had to read one article this week, this is it.

I do not hear much engagement with the economic arguments that, while labor market transitions are costly, freeing up labor has been one of the major modes of material progress throughout history. The US economy has a remarkable degree of automation already, not just from AI, and currently stands at full employment. If need be, the government could extend social protections to workers in transition rather than halt labor-saving innovations.
…
So I will declare mine: decentralized social systems are fairly robust; the world has survived some major technological upheavals in the past; national rivalries will always be with us (thus the need to outrace China); and intellectuals can too easily talk themselves into pending doom. All of this leads me to the belief that the best way to create safety is by building and addressing problems along the way, sometimes even in a hurried fashion, rather than by having abstract discussions on the internet. So I am relatively sympathetic to AI progress. I am skeptical of arguments that, if applied consistently, also would have hobbled the development of the printing press or electricity.

Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes

Notes: Distilling is the key to Edge AI. Not the first paper showing great results. See #18, #16, #15.

ChatGPT3 in Sub-Saharan Africa: A Threat or an Opportunity?

Notes: Short under 5 pages paper.

PMC-LLaMA: Further Finetuning LLaMA on Medical Papers

Notes: LLaMA continues to be consequential overall. In this case, fine-tuning on medical data showed an improvement on medical tasks.

Evaluating ChatGPT's Information Extraction Capabilities: An Assessment of Performance, Explainability, Calibration, and Faithfulness

🛡️CyberSecurity Reads

Google Is Rolling Out Passkeys, the Password-Killing Tech to All Accounts

Notes: Great news!

Using the iPhone Recovery Key to Lock Owners Out of Their iPhones

Notes: This hack is mitigated by turning on Face ID, but it is still disappointing there isn’t an additional check on the device before setting a recovery key. I suspect Apple will fix this soon given the media coverage.

Hackers Hit Dallas City Servers, Limiting Some Police and IT Systems

Notes: Sigh.😮‍💨

Poisoning Language Models During Instruction Tuning

Notes: Supply Chain Cybersecurity is critical.

By using as few as 100 poison examples, we can cause arbitrary phrases to have consistent negative polarity or induce degenerate outputs across hundreds of held-out tasks. Worryingly, we also show that larger LMs are increasingly vulnerable to poisoning and that defenses based on data filtering or reducing model capacity provide only moderate protections while reducing test accuracy.

I wrote a disclosure for this newsletter in #7. Please consider reading it.