weekend ai reads for 2025-01-17

January 17, 2025

đź“° ABOVE THE FOLD: WHAT IS GOING ON?

via mike, Brainwash An Executive Today! / Ludicity (24 minute read)

And conversely, if you do not deploy your system to a working state but can someone make your VP or CEO happy, then you did ship. This sounds even stranger. How can you ship if the code doesn't work?

It’s called lying, and it’ll solve all your problems!

  • and

Why would non-technicians be so focused on a database of all things, a concept so dull that it is Effective Communication 101 to try and avoid using the term in front of a lay audience? It’s because if you buy Snowflake then you’re allowed to get onto stages at large venues and talk about how revolutionary Snowflake was for your business, which on the surface looks like a brag about Snowflake, but is actually a brag about the great decisions you’ve been making and the wealth you can deploy if someone becomes your friend. And the audience is full of people that are now thinking “If I buy Snowflake, I can be on that stage, and everyone will finally recognize my brilliance”.

  • the closing line gave this section its title, but it’s much funnier in context

CES was a giant exercise in AI gaslighting — There was plenty of shouting about AI at the tech world’s mega trade show, but not a lot of substance. / Fast Company (8 minute read)

AI Mistakes Are Way Weirder Than Human Mistakes / IEEE Spectrum (14 minute read)

AI errors come at seemingly random times, without any clustering around particular topics. LLM mistakes tend to be more evenly distributed through the knowledge space. A model might be equally likely to make a mistake on a calculus question as it is to propose that cabbages eat goats.

 

đź“» QUOTE OF THE WEEK

You can be efficient or effective. When it comes to innovation, choose effective.

Christina Wodtke (source)

 

đź‘Ą FOR EVERYONE

How AI could change EV charging / The Verge (7 minute read)

“There’s a big role for AI to play at the grid edge,” says Siobhan Powell, a postdoctoral researcher at ETH Zürich who was not involved in the study. “It didn’t used to be the case, right? There wasn’t a lot interesting going on and now that we have a chance to do control, there’s more opportunity and more value in knowing what’s going on.”

Copilot Chat’s features live in the Microsoft 365 Copilot app, a rebranding of the Microsoft 365 app.

  • branding and marketing are hard, huh?

Singapore uses AI to improve eldercare with early detection tools — The country has a dire shortage of nurses, so to fill the manpower gap, it’s using AI for preventive care. / Rest of World (7 minute read)

OpenAI’s ChatGPT adds scheduled tasks feature in beta — The AI chatbot can now set reminders and perform recurring actions. / The Verge (5 minute read)

  • using an LLM for this seems like overkill, but we’re still chuffed about it

 

đź“š FOUNDATIONS

The Rise of the Generalist — As AI handles specialized tasks with increasing sophistication, the ability to connect dots across domains has never been more valuable. / Figma Blog (8 minute read)

  • it feels like every new technology spurs a plethora or these articles

  • they’re not wrong, but maybe stop hyping specialists so much so these aren’t as necessary

Why AI language models choke on too much text / Ars Technica (17 minute read)

The key innovation behind transformer-based LLMs is attention, a mathematical operation that allows a model to “think about” previous tokens. Before an LLM generates a new token, it performs an attention operation that compares the latest token to every previous token. This means that conventional LLMs get less and less efficient as the context grows.

 

🚀 FOR LEADERS

The CEO’s guide to Generative AI , Second edition — What you need to know and do to win with transformative technology [PDF] / IBM Institute for Business Value (25 minute read)

  • useful, practical suggestion

  • “data” seems important

What Companies Succeeding with AI Do Differently / Harvard Business Review (13 minute read)

  • link to an archive because their website is infuriating

AI-generated phishing emails are getting very good at targeting executives — Hyper-personalized emails use "an immense amount" of scraped data. / Ars Technica (4 minute read)

Crossing the generative Al tipping point — From quick wins to sustained growth [PDF] / Google Cloud (9 minute read, plus videos)

 

🎓 FOR EDUCATORS

A Tutorial on Teaching Data Analytics with Generative AI [PDF] / Robert L. Bray, Northwestern University (24 minute read)

However, making my class “10000% ChatGPT enabled” was risky because it wasn’t clear whether a series of open ChatGPT quizzes could yield a meaningful grade distribution. Fortunately, this risk didn’t manifest, as students missed questions at nearly the same rate with AI as without AI. Indeed, a slight increase in quiz-question difficulty was sufficient to secure a healthy bell curve.

Kevin Byran on AI for Economics Education / Markus’ Academy, YouTube (71 minute video)

On Thursday, December 12, Kevin Bryan joined Markus’ Academy for a conversation on "AI for Economics Education." Kevin Bryan is an Associate Professor at the University of Toronto.

  • starts at “Pros and cons of AI in education”, you can stop about halfway

 

đź“Š FOR TECHNOLOGISTS

Build Everything with AI Agents: Here's How / David Ondrej, YouTube (40 minute video)

Sky-T1: Train your own O1 preview model within $450 / Nova Sky, University of California Berkeley, GitHub (6 minute read)

We introduce Sky-T1-32B-Preview, our reasoning model that performs on par with o1-preview on popular reasoning and coding benchmarks. Remarkably, Sky-T1-32B-Preview was trained for less than $450, demonstrating that it is possible to replicate high-level reasoning capabilities affordably and efficiently. All code is open-source.

  • we haven’t tried this but the hype is interesting

Here’s how Apple’s AI model tries to keep your data private — Apple Intelligence uses models built to work on your device for privacy purposes. Can it still keep privacy in check when going to the cloud? / The Verge (8 minute read)

 

🎉 FOR FUN

  • that headline is being generous about how terrible they look

DeepSeek - AI Assistant / Deep Seek, Apple App Store (3 minute read)

  • Google Play

  • one of the better free OpenAI/Anthropic alternatives

  • do read the privacy policy before jumping in; then get a VPN and/or a burner

How to gauge community sentiment with NotebookLM — Turn messy community discussions into clear insights with this technique. / AI Supremacy, Substack (sorry) (12 minute read)

 

🧿 AI-ADJACENT

  • some of the photos are reminiscent of early ai-cruft

 

â‹„