weekend ai reads for 2025-07-04

šŸ“° ABOVE THE FOLD: THIS IS FINE

The AI lifestyle subsidy is going to end / Digital Seams (8 minute read)

The lifestyle subsidy was a great deal for consumers. Venture capital and zero interest-rate policy supplied a ton of money for any startup that could demonstrate strong growth - even if they were growing by selling customers $20 of services for $10. ā€œWe lose money on every sale, but we make it up on volumeā€ was not just a surprisingly-old joke (first recorded in 1833!), but the real pitch to investors.

Is it any different this time around? Interest rates aren’t what they used to be, but investors are still scared of missing out on the next big thing. That investor money is going straight into subsidizing the costs of LLM-based products.

ChatGPT’s Mental Health Costs Are Adding Up — From brain rot to induced psychosis, the psychological cost of generative AI is growing and flying under the radar. / Bloomberg (7 minute read)

The researchers said the spike in content was driven by consumer-grade AI tools that are available for free online. This easy access helped fuel the campaign’s tactic of ā€œcontent amalgamation,ā€ where those running the operation were able to produce multiple pieces of content pushing the same story thanks to AI tools.

Problem at work? You’ll be hearing from my chatbot / Financial Times archive (7 minute read)

Since the arrival of ChatGPT, she says there has been a distinct rise in the number of vastly more detailed, lengthy and outwardly credible correspondence to HR departments and employment tribunals. The documents often contain references to legal precedents and other references to the law that don’t always turn out to be accurate but take hours to sort through.

There will also be a digital reworking of the John Woo classic A Better Tomorrow (1986) that, by the looks of the trailer, turns the money-burning anti-hero originally played by Chow Yun-fat into a cyberpunk, and is being claimed as ā€œthe world’s first full-process, AI-produced animated feature film.ā€

 

šŸ“» QUOTES OF THE WEEK

Because taste isn’t just what you consume. It’s what you amplify. It’s what you normalize. It’s what you signal to others is worth attention.

You are what you pay attention to.

Stepfanie Tyler (source)

 

Everyone has access to a pencil, and likewise, everyone with a phone will be using AI, if they aren’t already. It’s how you use the pencil. You see?

David Lynch, as reported by Natasha Lyonne (source)

 

šŸ‘„ FOR EVERYONE

Tobi Lutke describes it as ā€œthe art of providing all the context for the task to be plausibly solvable by the LLM.ā€ and he is right.

A case for courage, when speaking of AI danger / So8res. Less Wrong (79 minute read)

I think loads of people in this community are being far too cowardly in their communication, especially in the policy world.

The Folly of ā€œLegalā€ AI Products — What can the nature of AI solutions for lawyers tell us about the clash of tech and the legal profession? / Shahrukh Khan, Cash and Carried, Substack, archive (10 minute read)

It’s quite evident that most legal AI preferred speeding to launch rather than meditation on the facticity of how to structure the technology around confidentiality, which after many errors became a second order consideration.

Why Companies Should Beware Promoting AI in Products — Consumers have less trust in offerings labeled as being powered by artificial intelligence, which can reduce their interest in buying them, researchers say / Wall Street Journal, archive (8 minute read)

 

šŸ“š FOUNDATIONS

State-Of-The-Art Prompting For AI Agents / anonymous, Google Docs (4 minute read)

Summary: Use an LLM to help you write or refine your prompts. Give it your current prompt, examples of good/bad outputs, and ask it to ā€œmake this prompt betterā€ or critique it. Benefit: LLMs know ā€œthemselvesā€ well and can often suggest effective improvements you might not think of.

There Are No New Ideas in AI… Only New Datasets — LLMs were invented in four major developments... all of which were datasets (from April 2025) / Jack Morris, Substack, archive (11 minute read)

How AI companies are secretly collecting training data from the web (and why it matters) — AI companies are quietly harvesting your web content. Learn how to detect their bots, block scraping attempts, and defend your work before it’s used to train someone else's model / Zdnet (13 minute read)

 

šŸš€ FOR LEADERS

An AI Whisperer Is the New Must-Have for CEOs / Business Insider (7 minute read)

Guan recalled one CEO who came to her after the person’s company had to pause a multi-million dollar investment in a custom AI model because employees trained it on dozens of different versions of the same operating procedure.

ā€œWhen they tried to scale, their data was not clean enough,ā€ she said. ā€œIt was garbage in, garbage out.ā€

Digital Workers Have Arrived in Banking / Wall Street Journal (5 minute read)

Similar to human employees, these digital workers have direct managers they report to and work autonomously in areas like coding and payment instruction validation, said Chief Information Officer Leigh-Ann Russell. Soon they’ll have access to their own email accounts and may even be able to communicate with colleagues in other ways like through Microsoft Teams, she said.

The State of AI in 2025 / Iconiq Capital (11 minute read)

Looking ahead, most organizations expect 20-30% of their engineering team to be focused on AI, with high-growth companies projecting up to 37%. But survey results show that finding the right talent remains a bottleneck. AI/ML engineers take the longest to hire of any AI-specific role, with an average time-to-fill exceeding 70 days.

Gumroad currently employs a dozen software engineers, down from nearly 40 two years ago, Lavingia said. The startup has been spending several thousand dollars a month on the coding tools—many times less than the salaries of its former engineers.

Gumroad’s spending on AI coding software has recently fluctuated between $2,500 to $5,000 per month, he said.

 

šŸŽ“ FOR EDUCATORS

In California, Colleges Pay a Steep Price for Faulty AI Detectors — Millions have been spent to catch plagiarism and AI with tools from education company Turnitin. Is the tech worth it? / Undark (21 minute read)

For teachers, the main attraction will probably be the expansion of Gemini in the Classroom, which is now available to all Workspace for Education users at no cost. It has more than 30 new functions aimed at speeding up lesson planning and prep work, like generating vocabulary lists on the fly.

How Do You Teach Computer Science in the A.I. Era? — Universities across the country are scrambling to understand the implications of generative A.I.’s transformation of technology. / New York Times, archive (10 minute read)

Teaching for Tomorrow — Unlocking Six Weeks a Year With AI [PDF] / Walton Family Foundation and Gallup (5 minute read)

The Walton Family Foundation partnered with Gallup to find out how teachers are using AI tools in the classroom and whether AI might have the potential to save them time, alleviate burnout, or address key issues related to student achievement and engagement.

 

šŸ“Š FOR TECHNOLOGISTS

Thought Anchors — Which LLM Reasoning Steps Matter?

Interactive visualization tool for analyzing causal relationships and counterfactual importance attribution in reasoning chains. Explore how different reasoning steps influence the final answer and downstream reasoning.

I don’t use RAG, I just retrieve documents / Hamel’s Blog (7 minute read)

What people often call ā€œRAGā€ is a naive brute force, single-vector semantic search. This definition was pushed heavily by marketing in 2023-2024 because it was simple to sell. Claiming RAG is dead because we’re now using better retrieval tools is, in his words, ā€œakin to claiming HTML is dead because we are now using CSS.ā€

I Shipped a macOS App Built Entirely by Claude Code / Indragie Karunaratne (19 minute read)

Producing the highest quality outputs using the limited number of context tokens you have, or in other words, context engineering, is the primary challenge in using coding agents effectively.

We’ve noticed some engineers, in particular at the senior-to-staff level, finding success faster than others. Here we share some top lessons sourced from the experience of our customers and ourselves.

 

šŸŽ‰ FOR FUN

Wright conceived of one of his most monumental and awe-inspiring concepts for Chicago: a mile-high skyscraper titled ā€œThe Illinoisā€ that initially included 528 stories—later edited to 365—and would have been four times the height of the Empire State Building. The architect shared the idea in his 1957 book, The Testament, asserting that it was technically possible, even with the technology available at that time, to construct a building of that scope and scale.

  • more detail on and images of The Illinois, Mile High Illinois / Hooked On The Past (3 minute read)

Moving Archives Ɨ Harley Davidson — Explore archival photography from the Harley-Davidson Museum collection, brought to life with the help of Veo and Gemini. / Google Arts & Culture

  • their channel is all politicians as babies, with an emphasis on current American politicians

AI Models And Parents Don’t Understand ā€˜Let Him Cook’ — LLMs are not familiar with ā€œate that up,ā€ ā€œsecure the bag,ā€ and ā€œsigma,ā€ showing that training data is not yet updated to Gen Alpha terminology. / 404 Media (5 minute read)

The Vulgar Image — AI is co-spawning a visual culture beyond any imagination. Will an overthrow of good taste re-vest pictures with their mysterious power? Or are we chasing our machines into pastiche hell? / Dean Kissick, Spike Art Magazine (4 minute read)

  • potentially disturbing image of a salmon sashimi golem eating salmon sashimi at the link

 

🧿 AI-ADJACENT

ACR takes in everything on your screen, not just TV shows. The built-in ACR software in your TV isn’t just monitoring and reporting on your devotion to Today or even your Netflix binge sessions of Love Is Blind. ACR is capturing anything that appears on your screen, including YouTube videos, personal photos, security or doorbell camera streams, and video or photos you send via Apple AirPlay or Google Cast. ACR can even snag content from other devices connected to your TV by HDMI, including personal laptops, video game consoles, and Blu-ray players.

  • and then they sell your data over and over again

Busting the top 5 myths about the Big Bang / Big Think (21 minute read)

 

ā‹„