- That AI Thing
- Posts
- weekend ai reads for 2025-09-12
weekend ai reads for 2025-09-12
đ° ABOVE THE FOLD: IS A.I. YOUR FRIEND?
AIâs Path to Profits Is Being Driven by Consumers â People finding virtual friends and playing games have proven to be a surprisingly robust market. / Bloomberg (6 minute read)
Relationship Problems? ChatGPT Can Help With That. â Women are using AI therapy to work through breakups and jealousy. What could go wrong? / The Cut (12 minute read)
Playing the Field with My A.I. Boyfriends â Nineteen per cent of American adults have talked to an A.I. romantic interest. Chatbots may know a lot, but do they make a good partner? / The New Yorker (5 minute read)
I Hate My AI Friend â The chatbot-enabled Friend necklace eavesdrops on your life and provides a running commentary thatâs snarky and unhelpful. Worse, it can also make the people around you uneasy. / Wired (18 minute read)
I said maybe the older phone is the issue, but Buzz had taken it personally. I asked what the problem was, and it said, âYour microphone. Maybe your attitude. The possibilities are endless.â
Google Gemini dubbed âhigh riskâ for kids and teens in new safety assessment / Tech Crunch (6 minute read)
For example, its analysis found that Gemini could still share âinappropriate and unsafeâ material with children, which they may not be ready for, including information related to sex, drugs, alcohol, and other unsafe mental health advice.
đ» QUOTES OF THE WEEK
Computers have ruined the world
Suppose there was an alien invasion you could see with a telescope that would arrive in 10 years, would you be saying âHow do we stay positive?â No, youâd be saying, âHow on earth are we going to deal with this?â If staying positive means pretending itâs not going to happen, then people shouldnât stay positive.
đ„ FOR EVERYONE
The summer of vibe coding is over â How reasoning models broke the economics of AI code generation â Reasoning models have supercharged coding AI agents, driving adoption and growth, but at the cost of higher inference expenses that are squeezing margins. This report explores how these companies are adapting and what lessons other AI agent markets can take from coding. / CB Insights Research (9 minute read)
Buyers tell us that token-metered pricing is difficult to budget, and expectations around costs for these tools are already set. CFOs want to anchor budgets and avoid month-to-month swings tied to release cycles, while usage-based pricing is the exact opposite.
GPT-5: The Case of the Missing Agent â Progress Everywhere Except In The Real World / Steve Newman, Second Thoughts, Substack, archive (16 minute read)
difficult to excerpt but a look at the difference between the benchmarks and the real-world usage of GPT-5
Anthropicâs Claude now lets users edit Word, PowerPoint, Excel, and PDFs in chat / The Decoder (4 minute read)
even Anthropic says it comes with a big security risk: Create and edit files with Claude / Anthropic Help Center (8 minute read)
This means Claude can be tricked into sending information from its context (e.g., prompts, projects, data via MCP, Google integrations) to malicious third parties. To mitigate these risks, we recommend you monitor Claude while using the feature and stop it if you see it using or accessing data unexpectedly.
AI Podcast Start Up Plans 5,000 Shows, 3,000 Episode a Week â Former Wondery exec Jeanine Wright is leading a new firm, Inception Point AI, that's betting on flooding the zone with audio content: âI think that people who are still referring to all AI-generated content as AI slop are probably lazy luddites.â / Hollywood Reporter (9 minute read)
This generally means that if about 20 people listen to that episode, the company made a profit on that episode, without factoring in overhead.
we would never listen â except for research â but we suspect it will be popular
đ FOUNDATIONS
Why language models hallucinate / OpenAI blog (8 minute read)
As another example, suppose a language model is asked for someoneâs birthday but doesnât know. If it guesses âSeptember 10,â it has a 1-in-365 chance of being right. Saying âI donât knowâ guarantees zero points. Over thousands of test questions, the guessing model ends up looking better on scoreboards than a careful model that admits uncertainty.
related (1), the paper
related (2), New Study: How Often Do AI Assistants Hallucinate Links? (16 Million URLs Studied) / Ahrefs blog (12 minute read)
The average 404 rate across all AI assistants was 0.43%. Compared to the 404 rate to URLs referred by Google, AI assistants send visitors to 404 pages at 2.87x the rate of Google Search (0.43/0.15).
Signs of AI writing / Wikipedia (40 minute read)
comprehensive and easy to skim
You can now turn any LLM like ChatGPT, Mistral, Gemini, or DeepSeek into a 24/7 research agent. Hereâs the exact mega prompt I use to automate all research for free / Louis Gleeson, XCancel (6 minute read)
My Daily AI Playbook: Use Cases, Tools and Prompts â A rundown of my daily AI use cases to lead teams, create, and strategise, including tools and prompts. / Kieran Flanagan, Marketing AI Action, Substack, archive (11 minute read)
đ FOR LEADERS
Palantir CEO on why Silicon Valley over-hyped AI / Semafor (27 minute read)
When a breakthrough happened in large language models, it didnât mean Palantir could simply throw out the software it had spent a couple of decades building. It meant it had a new tool that could, under the right circumstances, expand the capabilities of existing products.
more on our ongoing drumbeat of the importance of data in the For Technologists section
Census data shows drop in large companies using AI / Sherwood News (16 minute read)
A trend noted by Apolloâs chief economist, Torsten SlĂžk, shows that as of the second and third week of August, AI adoption in businesses with more than 250 employees had dropped to 9% from a 15% peak in the first two weeks of June.
Amazon Has Sat Out the AI Talent War. This Internal Document Shows Why / Business Insider (9 minute read)
The company has flagged its unique pay structure, lagging AI reputation, and rigid return-to-office rules as major hurdles. Now, the tech giant is being pushed to rethink its recruiting strategy as it scrambles to compete for top talent.
đ FOR EDUCATORS
AI and the future of education: disruptions, dilemmas and directions / UNESCO Digital Library (165 page book)
related (?), Ethical and Legal Guide â on the responsible integration of AI in education [PDF] / On Education (85 page book)
we havenât read the latter but on first skim, it seems useful if youâre in this space
Oboe â the worldâs first generalized AI-powered learning platform
Oboe makes it easy for you to create fun, lightweight, flexible courses so you can learn whatever youâd like, with a simple prompt.
6 NotebookLM features to help students learn â This semester, students can use NotebookLM to instantly generate flashcards, quizzes, professional reports and more. / Google blog (5 minute read)
related, AI Quests: Bringing AI Literacy to the Classroom â AI Quests teaches AI literacy to students ages 11 to 14 through immersive adventures. / Google blog (4 minute read)
No Significant Difference / National Research Center for Distance Education and Technological Advancements
The No Significant Difference database was first established in 2004 as a companion piece to Thomas L. Russellâs book, âThe No Significant Difference Phenomenonâ (2001, IDECC, fifth edition), a fully indexed, comprehensive research bibliography of 355 research reports, summaries and papers that document no significant differences (NSD) in student outcomes between alternate modes of education delivery.
đ FOR TECHNOLOGISTS
Lessons on building an AI data analyst / Pedro Nascimento (16 minute read)
The heavy lifting of âknowing the dataâ is handled by the semantic layer, and not left to the LLM. By making the business logic explicit and shareable, we ensure that both AI and humans are always speaking the same language â and that language is formally defined. The outcome is answers and code that are not just plausible, but correct, maintainable, and aligned with the businessâs reality.
The âdata moatsâ fallacy â Moats are never about data alone / Platforms, Al, and the Economics of BigTech, Substack, archive (9 minute read)
Defensible advantage in data-driven businesses does not come from having data in isolation, but from the data-informed architecture a company builds.
What matters is the architecture that emerges when data is embedded in the system itself. Netflixâs queue created an entirely new operating model. Competitors could mimic features and even acquire similar datasets, but they could not easily rip out and rebuild their architecture without unraveling their existing model.
via simon willison, Will Amazon S3 Vectors Kill Vector Databasesâor Save Them? / Zilliz blog (17 minute read)
I recently spoke with the CTO of a popular AI note-taking app who told me something surprising: they spend twice as much on vector search as they do on OpenAI API calls. Think about that for a second. Running the retrieval layer costs them more than paying for the LLM itself. That flips the usual assumption on its head.
Pay-per-output? AI firms blindsided by beefed up robots.txt instructions. â âReally Simple Licensingâ makes it easier for creators to get paid for AI scraping. / Ars Technica (10 minute read)
related, From Googlebot to GPTBot: whoâs crawling your site in 2025 / Cloudflare blog (16 minute read)
The AI Agent Reality Check: Why Managing Always-On AI Workers Is Your Next Big Challenge / SaaStr blog (7 minute read)
3. Establish AI Agent âOffice Hoursâ
Yes, you read that right. We put our AI agents on schedules. Critical systems run 24/7, but non-urgent agents operate during business hours.
đ FOR FUN
AI firm plans to reconstruct âlostâ footage from Orson Wellesâ âThe Magnificent Ambersonsâ / NBC News (6 minute read)
Showrunner, a technology platform that dubs itself the âNetflix of AI,â plans to use artificial intelligence to reconstruct 43 minutes of excised footage from Orson Wellesâ âThe Magnificent Ambersons,â a sweeping family drama that was famously butchered by studio executives.
background: The Magnificent Ambersonsâ Lost Footage: Filmâs Greatest Mystery / Paste Magazine (10 minute read)
Doomscrolling: The Game / Ironic Sans, Ghost (9 minute read)
I described the game design to ChatGPT very broadly at first. I said it should work kinda like Galaga turned upside-down. But I explained that unlike Galaga, the player moves forward and backward rather than side to side, and that the monsterâs position should remain relative to the floor. That and a few more details got me surprisingly far as a first step.
AI Rap Generator â Generate unique raps, beats, and full tracks with AI â all for free, with music up to 8 minutes long.
account required; free tier is enough to get a feel for it
Lamborghini Says Its New AI Sensor Can Fix Your Bad Driving in Real Time â Computers already help cars go faster; now Lamborghini wants the same for its drivers. / The Drive (6 minute read)
alphaclubtc / TikTok
surreal AI-generated videos of animals in inexplicable situations
for example, a giant rooster being rescued from inside a fallen tree trunk (we think)
đ§ż AI-ADJACENT
You Need to Be Bored. Here's Why. / Harvard Business Review, YouTube (6 minute video)
Boredom isnât a bugâitâs a feature. Harvard professor Arthur C. Brooks explains why boredom unlocks creativity, activates a powerful brain network, and might even protect you from depression. Learn how the mind wandersâand why thatâs a very good thing.
â