weekend ai reads for 2023-08-11

📰 ABOVE THE FOLD: CRIME

This was inevitable, but criminals have created their own ChatGPT clones, and are tailoring them for sophisticated attacks. For example: scammers used AI to fake my daughter’s kidnap.

AI is also being used outside of LLMs to make life more dangerous. New acoustic attack steals data from keystrokes with 95% accuracy. This reminded us of the research from a 2020 where vibrations from a light bulb can be used to “hear” inside a room.

And it’s not just crime — it’s also bad behavior from the models and users. New York Times reports that AI-generated guidebooks are a new way to scam travelers; low-quality guidebooks, complete with hallucinations are and will cause problems for readers. AI is acting ‘pro-anorexia’ and tech companies aren’t stopping it. Supermarket AI meal planner app suggests recipe that would create chlorine gas. Innocent pregnant woman jailed amid faulty facial recognition trend.

And just wow: An Asian Woman Asked AI to Improve Her Headshot and It Turned Her White. The CEO’s response wasn’t amazing:

📻 QUOTE OF THE WEEK

Claire Silver (source)

🏗️ FOUNDATIONS & CULTURE

• If you haven’t kept up on a weekly basis on the LLM-world, this talk is worth your time: Catching up on the weird world of LLMs Simon Willison

• Poll Shows Overwhelming Concern About Risks From AI as New Institute Launches to Understand Public Opinion and Advocate for Responsible AI Policies Artificial Intelligence Policy Institute

  • The AI Crackdown Is Coming — Five ways for Washington to hold Silicon Valley accountable The Atlantic

  • OpenAI will crush Sarah Silverman — But Congress could change that Jovono, Medium

• Agent Bench — Evaluating LLMs as Agents

  • An interesting open source benchmark for LLMs

  • GPT-4 performs better than all others; no open source performs better than any API-based

• AI language models are rife with different political biases — New research explains you’ll get more right- or left-wing answers, depending on which AI model you ask. MIT Technology Review

• ChatGPT has higher emotional awareness than humans, study finds The Decoder

• Our Natural Human Defenses Against A.I. Culture Culture

🎓 EDUCATION

• Sal Khan (30:29) (no direct link to the episode) Unconfuse Me with Bill Gates

  • transcript [PDF]

• via bob, Shockwaves & Innovations: How Nations Worldwide Are Dealing with AI in Education The 74 Million

• ASU unveils first-of-its-kind organization to advance AI across the enterprise Arizona State University

  • related: Students can use AI on applications, Arizona State law school says Reuters

• Craft AI Literacy Resources Stanford University

📊 DATA & TECHNOLOGY

Operationalizing AI within organizations and software is becoming important, especially for under-resourced organizations. Two articles and a paper that helped to shed more light on this:

• Good governance essential for enterprises deploying AI MIT Technology Review

  • with Stephanie Zhang, Head of ModelOps, AI and ML lifecycle management and governance at JPMorgan Chase

• Patterns for Building LLM-based Systems & Products Eugene Yan

• Operationalizing Machine Learning: An Interview Study arXiv

• Agent Bench — Evaluating LLMs as Agents

  • GPT-4 performs than all others, no open source performs better than any API-based

  • related: ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs arXiv

• Leveraging LLMs on your domain-specific knowledge base — With RAG to Riches: wielding the power of LLMs using Retrieval-Augmented Generation to talk to your data ML6 Team

• OWASP Top 10 for Large Language Model Applications Open Worldwide Application Security Project

  • not mentioned in the headline but these are the top 10 vulnerabilities OWASP identified

• FacTool: Factuality Detection in Generative AI EthanC111, Github

  • Paper arXiv

🎉 FUN and/or PRACTICAL THINGS

• Heuristica AI-powered knowledge exploration

  • tool that helps make mind-maps (?) of any topic; explore the topics through various lenses; probably does much more

• Introducing PlayHT2.0: The state-of-the-art Generative Voice AI Model for Conversational Speech PlayHT

  • examples at the link; a little uncanny valley

• Machine Learning for Business Decision Optimization Weights & Biases

  • free course

• Casehopper Write Immigration Documents with AI

• Recast AI-summarized audio of articles intended to reduce reading time; the summaries seem adequate, but also seems to be a lot of repetition which could be an artifact of the writing itself

• The TextFX Project Google

  • introduced by Lupe Fiasco (MIT Visiting Scholar at Google!), AI-tool that seems to be designed to help songwriters

🧿 AI-ADJACENT

• The Edited Latecomer’s Guide to Crypto Molly White

  • Crypto skeptics annotate a 2022 article from New York Times on crypto