The Vibes Have Come for Search

Hello, and happy Sunday! Dan Shipper’s piece about the return of generalists got us chatting in the Every Discord about generalists versus specialists. Many of us consider ourselves unabashed generalists. “My go-to line growing up (that my parents despised) was, ‘Specialization is for insects,’” entrepreneur in residence Brandon Gell told us. Read on for more about that piece along with everything else we published last week, and our take on the latest tech news.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

---

Knowledge base

"Why Generalists Own the Future" by Dan Shipper/Chain of Thought: Forget what you've heard about AI making generalists obsolete. Dan argues that in a world where LLMs can match specialist knowledge, it's the adaptable, curious minds that will thrive. Generalists armed with AI tools can navigate "wicked"—or complex and ambiguous—environments where experts (and machines) struggle. Read this if you want to understand why being a jack-of-all-trades might just be your superpower in the age of AI.

"Vibes-based Search" by Evan Armstrong/Napkin Math: The power of LLMs to transform search engines has been dramatically underestimated. AI can now intuit what you mean, search any dataset, and co-create answers with you. It represents a genuine threat to Google and a wholesale change to the way the internet works. Read this if you want to see how AI is evolving beyond finding information to understanding and creating it—and why that's a massive opportunity for startups.

🎧 "How to Supercharge Your Writing With AI Tools" by Dan Shipper/Chain of Thought: We publish lots of work around here—AI is how we do it with collapsing from exhaustion. Dan and Evan spill the beans on their AI-powered writing processes, from using ChatGPT to discover your literary taste to generating metaphors with Claude. Watch this if you want to learn how to use AI to improve your writing, or if you just want to see two word nerds geek out over their favorite tools. 🎧 Watch on X or YouTube, or listen on Spotify or Apple Podcasts, and access the full transcript. ✎ Sign up for our new course, How to Write With AI, to learn how to do it yourself.

🔏 "The Key to Great AI Prompting? Show, Don’t Tell" by Michael Taylor/Also True for Humans: Few-shot learning—providing AI with a handful of task examples—is the secret sauce to getting better results from ChatGPT and other language models. It's like teaching a child to apply sunscreen: Show, don't just tell. Read this if you want to level up your AI prompts and get more reliable, creative outputs. 🔏 Paid subscribers get access to code samples for testing few-shot prompts.

---

Fine tuning

Chatbots are getting bigger than you think. A New York Times report pegged ChatGPT at having 200 million weekly active users—twice as much as nine months ago. Meanwhile, The Information reported that Meta AI’s chatbot has 40 million daily active users, with 400 million logging in per month—twice what it was last November. That there are multiple new consumer products scaling to hundreds of millions of users at once flies in the face of consumer tech’s slump.

10 employees just raised $1 billion. It helps when world-famous AI researcher Ilya Sutskever is one of those ten. His new company, which he formed after leaving OpenAI over the failed coup against Sam Altman, is trying to build what it calls safe superintelligence—which is also the name of the company. These rounds for big AI research projects are becoming difficult to get done (and the returns are uncertain), but I love the ambition. Venture capitalists underwriting science risk is always a good thing.

A Korean deep-fake porn crisis. AI is a wonderful invention, but image models make it too easy for non-consensual adult imagery to be generated. Using the Telegram messaging app, group chats with thousands of individuals from more than 500 schools and universities sent images in which AI stripped women’s clothing. One Telegram group had 2,000 members dedicated to sharing images of under-age girls just from one school. Talk with your children about if/when they post their faces online. This stuff is too easy to do. 

Chip up or shut up. There are three buckets of inputs that, when multiplied, determine the success of an AI model. Algorithmic architecture times data times compute infrastructure equals the strength of an LLM. For example, while almost all models have ended up with GPT4-level capabilities, this clustering of performance is less an issue with transformer algorithms running out of juice and more a case where the major providers were using a similar size of compute . While public attention is on the algorithms, most of the capital and focus behind the scenes is on chips. In that regard, this was a big week: OpenAI is looking to spend "tens of billions" in the U.S. to build out compute infrastructure, and Elon Musk made a 100,000 GPU cluster, the biggest in the world, in under four months for X.AI. The bottleneck holding back the next generation of models is being removed. GPT-5-class models seem likely to blow your socks off.—Evan Armstrong 

---

Data mining 

AI adoption on the march. Enterprise AI adoption rates continue to increase across every sector:

Source: JPMorgan/Michael Cembalest.

Adoption is highest in software and customer service, but most other sectors have increased their adoption by at least 10 percentage points since last year. It’s no surprise that coding is leading the charge—that’s also where the lion’s share of recent funding has gone. As Dan wrote this week, LLMs excel in “kind” environments, like a codebase, which has a relatively structured dataset with clearly defined relationships (because otherwise the code wouldn’t work). That said, adoption rates are still just barely scraping 50 percent at the higher end  we’re still in the very early innings of this technology, and there is enormous territory left to conquer. You know what else has a relatively structured dataset and clear ontology? It’s that sector furthest behind the curve and all the way to the right: law.—Moses Sternstein  

---

Keyword extraction

Michael Taylor, who wrote about the power of examples in AI prompting, shared one good watch:

🖇 Coding demonstration with Cursor AI by Ricky Robinett: "If you've ever described yourself as non-technical, watching this 8-year old code a fully functioning app with Cursor AI may make you reconsider. I use Cursor every day, and while it's not as easy as she makes it look, the barrier to entry to building your own stuff just dropped by 10x."

---

Alignment 

Viral wisdom. Last week I was scrolling on X and saw five words that grabbed me by the scruff of the neck: “Absolutely unmixed attention is prayer." This quote, shared by an anonymous account, introduced me to the world of 20th-century French philosopher Simone Weil. Her work is a revelation. I instantly bought her first and last notebooks (because the best work is always in notebooks), and her emphasis on the value of paying attention feels radical. She challenges our modern-day conceptions of productivity and presence that’s making me rethink how I spend my precious time on this planet. To that anonymous poster: thank you. X’s serendipity led me to Weil, but I think Weil is leading me to myself.—Ashwin Sharma

Sentiment analysis

“Totally agree. I also think that having a specialist area (one you've dived deep into in the past) helps you know what questions to ask about the new thing you want to learn about. To be able to do both might be where the superpower is at.”—A reader in response to Dan’s piece about generalists

“One thing we are doing, and seeing other startups do as well, is using third-party firms on short-term hire—i.e., hire a marketing firm for a few months every year vs. a full-time team, or [take] a project-based approach. [It] lets us stay fluid on messaging yet keep the fixed costs low.”—A reader in response to Evan’s piece about gritty startups

Want to chat? DM Dan or Evan on X.

---

Hallucination

What if Tesla made a Cybercamper?

That’s all for this week! Be sure to follow Every on X at @every and on LinkedIn.