weekend ai reads for 2025-08-15

šŸ“° ABOVE THE FOLD: HEALTH CARE

Illinois bans therapeutic use of AI without clinician input — AI will be banned for therapeutic decision-making, but will still be allowed for administrative and supplementary support. / Healthcare Finance News (4 minute read)

The Future Of Health Data In The Age Of AI — Sensitive data about your health used to be relatively safe and anonymous. AI is making it much easier for that information to be used against you in opaque ways. / Noema Magazine (19 minute read)

Artificial intelligence helped health professionals to better detect pre-cancerous growths in the colon, but when the assistance was removed, their ability to find tumors dropped.

Where Are All the AI Drugs? — In an industry where 90 percent of drug candidates fail before reaching the market, a handful of startups are betting everything on AI to beat the odds. / Wired, archive (25 minute read)

Why AI Chatbots Can Be Risky for Autistic People / Wall Street Journal, archive (7 minute read)

 

šŸ“» QUOTES OF THE WEEK

I am clearly not capable of solving this problem. The code is cursed, the test is cursed, and I am a fool. I have made so many mistakes that I can no longer be trusted.

Google’s Gemini (source)

 

If you terrorists don’t kill us then the politicians will

Jeremy Messersmith (source)

 

For investors, treat ā€œAIā€ the way we treated ā€œmobileā€ circa 2011: infrastructure bets can still clear the hurdle rate if you triple the time discount, but the application layer is a graveyard of demos wearing revenue costumes.

Adam Butler (source)

 

šŸ‘„ FOR EVERYONE

  • clear reporting on how technology is being used to kill in one part of the world, and how easy it would be for military states to do this globally

  • you choose the companies and policies you support with your content, time, and attention

  • their approval is in writing, along with acceptable levels of violence that minors would be exposed to; they drew the line at ā€œgoreā€ if that makes it any better

  • related, Why A.I. Should Make Parents Rethink Posting Photos of Their Children Online — Artificial intelligence apps generating fake nudes, amid other privacy concerns, make ā€œsharentingā€ far riskier than it was just a few years ago. / New York Times, archive (11 minute read)

The result? Each one reflected the same assumptions - those of English-speaking, Western European societies. None aligned with how people in Africa, Latin America, or the Middle East actually build trust, show respect, or resolve conflicts.

The age of industrialized imagination / Free Think (13 minute read)

AI will lower the cost of getting ideas out of the heads of wannabe film and gamemakers the same way e-publishing lowered the cost of getting written stories out of the heads of wannabe novelists. Again, when the price of failure drops, risk-taking and new ideas take off.

SEO Is Dead. Say Hello to GEO. — Search-engine optimization now feels dated. Generative-engine optimization is all about trying to trick AI chatbots. / New York Magazine (17 minute read)

A close examination of two pictures of the coffee table appears to show differences in the damage, leading the woman to believe they were digitally manipulated or generated by AI. She says the host was retaliating because she ended her tenancy early.

  • spoiler: Airbnb were initially completely useless in helping resolve this matter

 

šŸ“š FOUNDATIONS

GPT-5 Prompting Guide: Best Practices for General Usage / Designs, Magic Path (4 minute read)

GPT-5 represents a significant shift in how language models process and respond to prompts. Unlike previous models, GPT-5 is particularly sensitive to instruction structure, style, and tone.

Claude Code Is All You Need / Gareth Dwyer (26 minute read)

Give it a lot of input. The more input you give it, the better its output is. It’s a magic tool, but you’d still better be damn good at communicating, either by typing thousands of words into text files or the interactive window, or using TTS (I haven’t tried this because I hate the sound of my own voice, but others have reported great results).

How I Code with AI on a budget/free / Tom, wuu73 (7 minute read)

  • it involves having 14+ tabs open to free LLMs, and then some

 

šŸš€ FOR LEADERS

LLMs as models for analogical reasoning / Science Direct (90 minute read)

We assess the performance of both human participants and LLMs on tasks focusing on reasoning from semantic structure and semantic content, introducing variations that test the robustness of their analogical inferences. Advanced LLMs match human performance across several conditions, though humans and LLMs respond differently to certain task variations and semantic distractors.

  • a deep dive into the limits of LLMs as a ā€œbag of heuristicsā€; helps to build one’s mental model on what LLMs can and can’t do

Generative Al and the evolving role of marketing — a CMO’s Playbook [PDF] / Cap Gemini (7 minute read)

There's a calculation that every AI executive should know by heart, but most have never done: an on-premises GPU server costs roughly the same as six to nine months of renting equivalent cloud capacity.

when you’re spending time with an ai—whether coding, writing, or thinking—you always max out on quality. nobody opens claude and thinks, ā€œyou know what? let me use the [worse] version to save my boss some money.ā€

 

šŸŽ“ FOR EDUCATORS

I didn’t realize how irreplaceable I’d believed myself, how like a John Henry of the networked Humanities, until my students shared their findings. Yes, the majority preferred my feedback—it was noted that the AI models demonstrated an unhelpful fixation on ā€œimproving transitions,ā€ whatever that means—but even my strongest advocates noted that their AI tutors often gave advice similar to mine, and faster.

South Korea pulls plug on AI textbooks / The Korea Herald (5 minute read)

The amendment narrows the legal definition of textbooks to printed books and e-books, excluding ā€œlearning support software using intelligent information technology." This reclassifies AI-powered textbooks as just another type of educational material, not official textbooks.

Guiding Principles for Artificial Intelligence in History Education [PDF] / American Historical Associaction (6 minute read)

Given the speed at which technologies are changing, and the many local considerations to be taken into account, the AHA will not attempt to provide comprehensive or concrete directives for all instances of AI use in the classroom. Instead, we offer a set of guiding principles that have emerged from ongoing conversations within the committee, and input from AHA members via a survey and conference sessions.

 

šŸ“Š FOR TECHNOLOGISTS

Many vibe coding platforms store this data, even temporarily, often without the builder fully understanding the data flow. Some platforms retain data for debugging, analytics, or other operational purposes

  • our rule of thumb: don’t vibe code anything that involves real user data, or is public-facing

Exfiltrating data is possible because ChatGPT supports markdown code in responses, and so can display images from remote URLs. And when an image is requested from a remote server, local data can be passed as parameters in the URL, leading to data exfiltration.

 

šŸŽ‰ FOR FUN

MoovAi — AI Movie Recommendations & Film Discovery

Describe what you want to watch in natural language and get AI-powered movie recommendations. Create moodboards, explore the community, and discover your next favorite film.

  • the free tier is three recommendations per day

Can we just have one day when no one mentions AI? — The gap between hype and reality is frustratingly wide and growing / Financial Times, archive (7 minute read)

San Francisco’s Billboards Aren’t For You / Bay Area Current (8 minute read)

And the message is this: if you do not understand these billboards, this city is not for you. You don’t know what an API is? You’re clearly a Luddite, stuck in the analog past. Your economic value is minimal. Why are you still here? This place is for building the future, where everything is shiny and frictionless and all human labor is replaced by AI, which really means a towering stack of B2B software products, each with its own freeway billboard.

 

🧿 AI-ADJACENT

Executive Communication w/ Harrison Metal / Heavybit (44 minute read)

It doesn’t sound as warm and human as it might if Barbara were here. She’d say, ā€œGo ahead and make it feel the way you want it to feel.ā€ She'd say, ā€œPeople have been decorating stuff for 10,000 years. Go ahead and decorate your writing. But don’t lose the skeleton, don’t lose the underlying architecture of SCQA because it’s a gift to the reader.ā€

  • the Minto Pyramid Principle is something we think about almost daily

 

ā‹„