weekend ai reads for 2023-08-18

📰 ABOVE THE FOLD: ECONOMICS OF AI (SPECULATIVE)

• What exactly are the economics of AI? — Some AI makes a ton of money; a lot of it is still speculative Gary Marcus, Substack

• AI Won’t Supercharge the US Economy Bloomberg Opinion

• via adam, What if Generative AI turned out to be a Dud? — Some possible economic and geopolitical implications Gary Marcus, Substack

• Every start-up is an AI company now. Bubble fears are growing. Washington Post

• Nvidia and other AI stocks are slumping. 'Amara's Law' explains why. Yahoo Finance

  • meanwhile …

    • Saudi Arabia and UAE race to buy Nvidia chips to power AI ambitions — Gulf nations acquire thousands of GPUs amid global shortage of semiconductors needed to build large language models Financial Times

    • China's internet giants order $5bn of Nvidia chips to power AI ambitions — Baidu, ByteDance, Tencent, Alibaba rush to buy amid concern of U.S. clampdown Nikkei East

• ChatGPT fever spreads to US workplace, sounding alarm for some Reuters

  • The $900,000 AI Job Is Here — Salaries rise as employers such as Netflix and Walmart seek candidates with artificial-intelligence skills Wall Street Journal

📻 QUOTES OF THE WEEK

Leonardo da Vinci, allegedly (source)

Justin Wolfers (source Twitter, sorry) (non-Twitter link)

🏗️ FOUNDATIONS & CULTURE

• Two good AI explainers, one not technical and the other, very

  • Explained: The conspiracy to make Al seem harder than it is! by Gustav Söderström, co-founder of Spotify (1:30:00) Spotify

  • An observation on Generalization, with Ilya Sutskever co-founder and Chief Scientist of OpenAI (57:20) Simons Institute, YouTube

• Should Computers Decide How Much Things Cost? The Walrus

  • How AI-powered pricing improves margins and protects consumers Venture Beat

  • related, Hypertargeting at Scale The By Product, Substack

• Tracking AI-enabled Misinformation NewsGuard

  • This Disinformation Is Just for You Wired

  • These Women Tried to Warn Us About AI Rolling Stone

• Creation and Adoption of Large Language Models in Medicine Journal of the American Medical Association

  • Argument for medicine to build and use their own LLMs, trained on their own data; partially an effectiveness argument, and partially a financial one.

  • related, DoctorGPT, an open-source project with a mission to provide everyone their own private doctor Github

• Why the Great AI Backlash Came for a Tiny Startup You’ve Probably Never Heard Of Wired

  • the CEO’s response: Taking Down Prosecraft.io Shaxpir Blog

• Consulting giant McKinsey unveils its own generative AI tool for employees: Lilli Venture Beat

🎓 EDUCATION

• AI in Education — The leap into a new era of machine intelligence carries risks and challenges, but also plenty of promise Education Next

• ChatGPT-Wary Universities Scramble to Prepare for New School Year Bloomberg

• An Iowa school district is using AI to ban books Engadget

• AI 101 for Teachers Code.org, with ETS, ISTE and Khan Academy

  • Have not taken; two of six modules currently available

• AI Assistance in Legal Analysis: An Empirical Study Social Science Research Network

📊 DATA & TECHNOLOGY

• Plotting Progress in AI Contextual.ai

  • AI benchmarks are getting “solved” in less and less time.

• LIMA: Less Is More for Alignment arXiv

  • related (1), Optimizing latency Hamel Huasin

    • Testing ways to achieve massive speed increased with llama-v2 models. The fastest, mlc, averages 117 tokens per second, versus 24.6 when using HuggingFace Transformers.

  • related (2), via adam, Fine-Tuning Llama-2: A Comprehensive Case Study for Tailoring Models to Unique Applications Anyscale

  • related (3), Better Llama 2 with Retrieval Augmented Generation (RAG) (20:48) James Briggs, YouTube

• Simple synthetic data reduces sycophancy in large language models arXiv

  • LLMs will often repeat the user’s opinion, even when it’s wrong. This paper finetunes the model with synthetic data to reduce the sycophancy.

• Flows: Building Blocks of Reasoning and Collaborating AI arXiv

• REFORMS: Reporting Standards for ML-based Science Princeton University

🎉 FUN and/or PRACTICAL THINGS

• AI recreates clip of Pink Floyd song from recordings of brain activity — An artificial intelligence can guess what a song sounds like based on patterns of brain activity recorded while people were listening to it New Scientist

• Could THIS be the new James Bond? Handsome candidate for the next 007 is revealed - but there’s a HUGE catch Daily Mail

• ChatGPT for Excel Microsoft Store

  • Excel plug-in; but

  • … so hasn’t been tested thoroughly because we’re not going to use it with any real data; that said, it’s a bit pokey

• Metaphor — rediscover the internet

  • Search engine; seems effective at finding websites pertaining to a particular topic, but less so when asked to find specific information

• Skybox AI Blockade Labs

  • AI-generated 3d worlds based on user guides (drawing) and prompts

• AI Courses from Google (free)

  • Introduction to Generative AI 

  • Introduction to Large Language Models 

  • Introduction to Responsible AI

  • Introduction to Image Generation 

• A new AI app lets users ‘text’ with Jesus. Some call it blasphemy. Washington Post

🧿 AI-ADJACENT

The design trends forged by AI Decode Magazine

Posting this mostly as an excuse to share one of our favorite LIDAR videos, from 2008: House of Cards (4:34) Radiohead, YouTube