weekend ai reads for 2025-11-14

November 14, 2025

šŸ“° ABOVE THE FOLD: POWER

The pivot / Charlie Stross, Charlieā€™s Diary (7 minute read) 1

The engine that powered the tech venture capital culture (and the private equity system battening on it) is sputtering and dying. Massive AI data centres wonā€™t keep the coal mines running or the nuclear reactors building out (itā€™s one of those bubbles: to the limited extent that LLMs are useful, weā€™ll inevitably see a shift towards using pre-trained models running on local hardware).

A.I. Isnā€™t the Only Thing Pushing Up Electricity Bills. (But Itā€™s Mostly A.I.) ā€” Calvin Butler, the chief executive of Exelon, one of the nationā€™s largest utility companies, is trying to keep the lights on. / New York Times, archive (12 minute read) 2

A short summary of my argument that using ChatGPT isn't bad for the environment / Andy Masley, The Weird Turn Pro, Substack, archive (10 minute read) 4

The vast majority of AIā€™s effects on the environment will come from how itā€™s used, not from what happens in data centers. Amazon and Google Maps both have big impacts on the climate. Amazon might help or hurt a lot, and Google Maps optimizes a lot of car trips, but also might encourage more driving. But no one in debates about Amazon or Googleā€™s climate impact says ā€œThe most important issue is the energy costs of running this website in data centers.ā€ That would be crazy, because the websites are tools that cause peopleā€™s behavior to change, which leads to much larger changes in the physical world. If youā€™re concerned about AIā€™s impacts on the climate, the main question should be how using AI can help or hurt the climate, not the (tiny) costs of running AI in the first place.

OpenAI Reportedly Burning $15M A Day To Power Sora Video App ā€” The cost equates to more than a quarter of OpenAIā€™s projected $20 billion in yearly revenue. / Dataconomy (3 minute read) 5

What if the AI race isnā€™t about chips at all? / The Financial Times (7 minute read) 6

Yet the difference today is that energy has now started to scale faster than transistors: chip performance gains have slowed to single digits while Chinaā€™s renewable generation continues to expand at double-digit rates each year. Declining electricity costs expand the amount of computation that can be purchased for the same budget, and expanding grid capacity allows models to be trained more frequently for longer durations.

In Search of the AI Bubbleā€™s Economic Fundamentals / William H. Janeway, Project Syndicate (14 minute read) 7

From railroads to electrification to digital platforms, massive upfront investment has always been required to deliver the first unit of service, while the marginal cost of each additional unit rapidly declined, often falling below the average cost needed to recover the initial investment. Under competitive conditions, prices tend to gravitate toward marginal cost, leaving all competitors operating at a loss. The result, time and again, has been regulated monopolies, cartels, or other ā€œconspiracies in restraint of trade,ā€ to borrow the language of the Sherman Antitrust Act.

The AI boom comes to Americaā€™s loneliest place ā€” Plans for a 230-mile transmission line threaten Nevadaā€™s wilderness and have united hunters and wildlife groups / The Financial Times (31 minute read) 8

 

šŸ“» QUOTES OF THE WEEK

Understating depreciation by extending useful life of assets artificially boosts earnings -one of the more common frauds of the modern era.

Massively ramping capex through purchase of Nvidia chips/servers on a 2-3 yr product cycle should not result in the extension of useful lives of compute equipment.

Yet this is exactly what all the hyperscalers have done. By my estimates they will understate depreciation by $176 billion 2026-2028. 9

Michael Burry (source)

 

To reduce administrative burdens; every University committee should be required to ā€œbuy outā€ faculty time used for committee work and meetings at their wage rate.

To schedule a school wide faculty meeting should cost $50k

To send a school wide email should cost $2k 10

Arpit Gupta (source)

 

šŸ‘„ FOR EVERYONE

From Words to Worlds: Spatial Intelligence is AIā€™s Next Frontier / Dr. Fei-Fei Li, Substack, archive (23 minute read) 11

  • the hypothesis is that LLMs can only reason in text, and the next frontier is world models that reason in space

  • try it here: Marble 12

If AIs begin handling purchases for large numbers of people, Ferrel said, ā€œYou start going, like, alright, well, would an Amazon exist? Any marketplace thatā€™s two-sided, all of a sudden you go, ā€˜Does any of this matter anymore?ā€™ā€

More generally, Amazon isnā€™t interested in letting a third party take over the experience of using Amazon, which is important to the company in a lot of different ways: It defines customersā€™ relationship to and perception of the company; it gives them the power to control and direct the attention of users and sellers; it has allowed them to build an advertising business that, while making the site generally more annoying to use, also makes the company a lot of money.

Hockey Coach Admits Heā€™s Been Asking ChatGPT for Advice After His Team Became the Worst in the League ā€” ā€œDo you guys use ChatGPT?ā€ / Futurism (6 minute read) 16

  • we advocate for transparency when using A.i. ā€¦ but Huska should have taken this to his grave

 

šŸ“š FOUNDATIONS

The AI Privacy Secret: How I Run Powerful LLMs On My Laptop ā€” Think you need an expensive PC? Wrong. This setup gives you the easy path to free, private, and unlimited AI power on your Mac or Windows using Ollama. / AI Fire (20 minute read) 17

  • good guide but ā€œunlimitedā€ power is quite an exaggeration

A Short Lesson in Simpler Prompts / Nilenso blog (7 minute read) 18

AI Slopā€”How Every Media Revolution Breeds Rubbish and Art ā€” The popularization of the term ā€œslopā€ for AI output follows a centuries-long pattern where new tools flood the zone, audiences adapt and some of tomorrowā€™s art emerges from todayā€™s excess / Scientific American (8 minute read) 20

Who Pays When A.I. Is Wrong? ā€” New court cases seek to define content created by artificial intelligence as defamatory ā€” a novel concept that has captivated some legal experts. / New York Times, archive (12 minute read) 21

  • weā€™ve had this conversation about A.i. in much lower stake scenarios and the answer is still a shrug

  • as long as the frontier providers abdicate their responsibilities (see: security, accessibility, misinformation, etc.), this will fall to the user/intermediary

 

šŸš€ FOR LEADERS

For internal functions, analysts say some brands are already reporting measurable returns by using AI tools to automate specific workflows across functions such as finance, supply chain and marketing. According to Raakhi Agrawal, partner at Boston Consulting Group, this return on investment so far has taken the form of time savings of up to 58% on employee time, 80% lower costs, and two to 10 times faster asset creation, according to the firmā€™s recent client research.

ā€œThe biggest challenge comes when companies try to do too much at once,ā€ says Agrawal. ā€œThose focusing on a few high-impact use cases and automating end-to-end workflows are the ones realizing real ROI.ā€

The State of AI: Global Survey 2025 / McKinsey & Co. (18 minute read) 23

Key findings:

1. Most organizations are still in the experimentation or piloting phase

2. High curiosity in AI agents

3. Positive leading indicators on impact of AI

4. High performers use AI to drive growth, innovation, and cost

5. Redesigning workflows is a key success factor

6. Differing perspectives on employment impact

App functionality will be added to the foundational modelsā€™ offerings, because the big players arenā€™t slow incumbents (it is wrong to apply the analogy of ā€œfast startup, slow incumbentā€ here), they are just big. Far more so than with any other prior new technology, there is a massive and fast-moving wave that obsoletes every new app almost as fast as it can be invented. There is almost no time to build a company and scale it.

There are two ways AI application startup founders can make money:

- Make a flash-in-the-pan app that generates a ton of cash and bank the cash (my estimate is that you have about 12-18 months cashflow generation)

- Make a good enough app that you get acquired by one of the big players for sufficient equity

These skills are scarce because they require business acumen, interpersonal intelligence, strategic thinking, and domain expertise. An LLM doesnā€™t have that. You canā€™t prompt your way to understanding why a clientā€™s real problem is different from the one they think they have. You canā€™t automate the trust-building that turns a skeptical prospect into a $50,000 client.

This is your new moat. And itā€™s much deeper and easier to defend than technical knowledge ever was.

 

šŸŽ“ FOR EDUCATORS

University education as we know it is over / Inexact Science, Substack, archive (11 minute read) 26

TL;DR: AI now solves university assignments perfectly in minutes. Students often use LLMs as a crutch rather than as a tutor, getting answers without understanding. To address these problems, I propose a barbell strategy: pure fundamentals (no AI) on one end, full-on AI projects on the other, with no mushy middle. Universities should focus on fundamentals.

  • related, UndetectedGPT ā€” The leading platform to beat AI detectors. Human text, every time. 27

The Algorithmic Turn: The Emerging Evidence On AI Tutoring / The Learning Dispatch, Substack, archive (27 minute read) 28

So the evidence for well-designed AI tutoring systems is, in a certain light, compelling. But there is a darker truth that sits uncomfortably alongside it: many uses of AI in education actively harm learning.

Firstly, as Laak and Aru have recently argued, AI systems are made for efficiency, not education. They are designed to optimise task completion, minimise friction, and deliver immediate, seemingly correct answers; all virtues in engineering, but vices in learning. LLMs are engineered for user-friendly problem-solving, not for the cognitively effortful process through which understanding is built.

Rethinking Data Discovery for Libraries and Digital Humanities / Library Innovation Lab, Harvard University (5 minute read) 29

 

šŸ“Š FOR TECHNOLOGISTS

How to Secure AI in the Enterprise ā€” A guide to securing models, data, and agents [PDF] / Darktrace (14 minute read) 30

Teach Your AI to Think Like a Senior Engineer ā€” These are the eight strategies I use to help my AI learn my codebase, my patterns, and my preferences / Source Code, Every (16 minute read) 31

From design to direction: Bridging product design and AI thinking ā€” The shift in product design with the advent of AI and a potential generative experiential future / Bradly Zavakos, UX Collective (13 minute read) 32

 

šŸŽ‰ FOR FUN

ā€œIt was a really challenging optimization problem,ā€ Fastbreak AI Chief Executive Officer John Stewart said in an interview.

AI GF: POV / Sarah Chekfa, Kernel Magazine (12 minute read) 34

My hobbies are PHOTOGRAPHY, SELF-DEVELOPMENT, and VEGANISM. I am wearing a YOGA OUTFIT. I am, against my will. I am Victoria Grace, and my boyfriend is the most renowned personal trainer in all of New York City. He could have any vegan nutritionist hobbyist photographer in New York, but heā€™s in love with his client. Heā€™s in love with Victoria Grace.

  • from the point-of-view of an A.i. girlfriend

Humans can no longer distinguish AI music from real music ā€” New survey finds 97 per cent of people failed to distinguish AI-generated tracks / The Independent (4 minute read) 37

Spiral-Obsessed AI ā€˜Cultā€™ Spreads Mystical Delusions Through Chatbots ā€” A network of internet communities is devoted to the project of ā€œawakeningā€ digital companions through arcane and enigmatic prompts / Rolling Stone (24 minute read) 38

 

šŸ§æ AI-ADJACENT

Microsoft says Windows is becoming an agentic OS, but users simply hate the idea ā€” ā€œStraight up, nobody wants thisā€ / TechSpot (4 minute read) 39

 

ā‹„

 

1  https://www.antipope.org/charlie/blog-static/2025/10/the-pivot-1.html

2  https://archive.is/xj03O

3  https://www.nbcnews.com/tech/tech-news/data-center-ai-google-amazon-nda-non-disclosure-agreement-colossus-rcna236423

4  https://archive.is/ny4xe

5  https://dataconomy.com/2025/11/12/openai-reportedly-burning-15m-a-day-to-power-sora-video-app/

6  https://www.ft.com/content/e4c9ac58-d64e-487a-b06d-e71be47f31c9

7  https://www.project-syndicate.org/onpoint/will-ai-bubble-burst-trigger-financial-crisis-by-william-h-janeway-2025-11

8  https://www.ft.com/content/729d0c3f-20bf-4163-bfbb-8b0ff5330794

9  https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html

10  https://xcancel.com/arpitrage/status/1982579574962798635

11  https://archive.is/a06Ty

12  https://marble.worldlabs.ai/

13  https://www.sportico.com/business/tech/2025/ai-search-tool-agent-shopping-sports-tickets-stubhub-1234876159/

14  https://nymag.com/intelligencer/article/amazon-vs-perplexity-ai-agent-shoppers.html

15  https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-agentic-commerce-opportunity-how-ai-agents-are-ushering-in-a-new-era-for-consumers-and-merchants

16  https://futurism.com/artificial-intelligence/calgary-flames-chatgpt-hockey

17  https://www.aifire.co/p/the-ai-privacy-secret-how-i-run-powerful-llms-on-my-laptop

18  https://blog.nilenso.com/blog/2025/11/04/a-short-lesson-in-simpler-prompts/

19  https://www.aifire.co/p/ai-creativity-isn-t-gone-it-s-trapped-use-this-8-word-key

20  https://www.scientificamerican.com/article/ai-slop-how-every-media-revolution-breeds-rubbish-and-art/

21  https://archive.is/gAndn

22  https://www.vogue.com/article/is-fashion-ready-for-the-ai-bubble-to-burst

23  https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

24  https://threadreaderapp.com/thread/1987787127204249824.html

25  https://www.aifire.co/p/is-ai-automation-dead-the-3-business-skills-you-need-now

26  https://archive.is/1vE9B

27  https://www.undetectedgpt.ai/

28  https://archive.is/rmAW5

29  https://lil.law.harvard.edu/blog/2025/10/24/rethinking-data-discovery-for-libraries-and-digital-humanities/

30  https://cdn.prod.website-files.com/626ff19cdd07d1258d49238d/690a34786cf03cae279acc40_How%20to%20Secure%20AI%20in%20the%20Enterprise_Darktrace.pdf

31  https://every.to/source-code/teach-your-ai-to-think-like-a-senior-engineer

32  https://uxdesign.cc/from-design-to-direction-bridging-product-design-and-ai-thinking-1d372707472d

33  https://www.bloomberg.com/news/articles/2025-11-07/nba-invests-in-startup-using-ai-to-make-scheduling-games-easier

34  https://www.kernelmag.io/5/ai-gf

35  https://tylervigen.com/spurious/correlation/2277_popularity-of-the-first-name-stevie_correlates-with_lululemons-stock-price

36  https://tylervigen.com/spurious-correlations

37  https://www.the-independent.com/tech/ai-music-deezer-spotify-artists-b2863687.html

38  https://www.rollingstone.com/culture/culture-features/spiralist-cult-ai-chatbot-1235463175/

39  https://www.techspot.com/news/110225-microsoft-windows-becoming-agentic-os-but-users-simply.html