Welcome back, Insiders! 

My grandfather was the editor of a local Rhode Island newspaper, The Woonsocket Call. He was long retired when the internet started feasting on local print papers, but watching the transition from a distance filled him with a mix of horror, fascination, and confusion.

He had plenty of logistical questions like how do you pay for an online paper and how do you share stories if you don’t have a physical paper to flip through and pass around? (Newspaper sections scattered my grandparents' house like toys in a daycare center.) But my grandfather also had plenty of philosophical questions like, could anyone publish anything on the internet? What would that mean for journalists? For a well-informed society? 

He was equally awed and confused when I’d type a question into Google and get an answer in an instant. “Who is answering?” 

My grandfather was on my mind this week as internet search stepped into the AI spotlight. 

 Google launched AI search summaries. While I was delighted by how quickly I could help my son remember the Pythagorean Theorem and the significance of the Battle of Antietam, other users found it less helpful. Like the person who was told to eat rocks and another who was encouraged to use glue on pizza (helps the cheese stick, duh). I couldn't help but hear my grandfather, "Who is answering?" ⛰️🍕⁉️

The answer, it seems, is no one (not even the Great and Powerful Altman) really knows. 🤷

But thanks to a team of researchers at Anthropic, this week the answer is a little less mysterious. While the summary of the research into how Large Language Models learn is fascinating, the study also helps to serve as a stabilizing bookend to the wacky (and potentially dangerous) hallucinations of AI search. The more we learn about how these LLMs work, the more we can learn how to limit their misuse, threats, and just plain wrong information.  🙌

If you’re thinking, what if my kid is home alone and wants some pizza, Googles how to make it, and we end up in the ER for glue ingestion, you’re not alone you’re probably assuring yourself that even a 12-year-old would know a hallucination when they read it. 🤞

But it’s even more reassuring to know that the Anthropic team was able to manipulate the output of certain queries showcasing how this research “could allow A.I. companies to control their models more effectively.”

As more and more slop is likely being used to train these LLMs, knowing how to make them more accurate is critical.

📁 (File under New AI Term You Should Know: slop - “what you get when you shove artificial intelligence-generated material up on the web for anyone to view.”)

Coupled with the other big story – Scarlett Johansson’s voice – the question of the week seems to be, is AI moving too fast?

👉 (ICYMI: After hearing OpenAI’s voice assistant’s eerie resemblance to her own, ScarJo wrote in a highly publicized statement, that after being approached by Sam Altman to voice the OpenAI assistant, she declined. OpenAI asked her to reconsider. Before she could decline, the voice assistant was released to the public.) 

AI is moving fast. It’s probably not a coincidence that OpenAI launched its new voice capabilities during the same week Google launched AI search summaries. It’s not surprising that while Google and Open AI race to dominate search, smaller players like Perplexity are rethinking search entirely in an attempt to compete. 🏅

AI is moving fast but it’s getting bipartisan support for greater government regulation. Creatives are pushing back. Even AI engineers are questioning the AI "rat race." 🚫

The lure of billion-dollar valuations is tempting. Flashy AI tools with only the promise of returns and no real use cases are being rewarded for potential. But that’s simply not practical. Why build a tool that has no real value? 

Is AI moving too fast? 

It’s a valid question but one alwaysAI is not focused on. Instead, we’re focused on:

• Does AI provide business value?
• Is it scalable?
• Are there significant use cases that could solve persistent problems?

In the case of Vision AI, with real-time visual data, the answer is yes. Wherever you could use an extra set of eyes to observe, record, or alert, Vision AI is useful.  

Even in a time before computers and the internet, my grandfather would often caution, “Anyone can tell you anything.” I don’t know exactly what he’d have to say about the rise of AI but I know for sure he’d still remind me to “verify everything you see and hear.” I’d be happy to tell him that Vision AI is a verification. It simply shows you what’s happening in your business, even if no one is watching. 👀