Between vacations, year-end projects, upcoming holidays, and other hysteria, there hasn’t been an article coming out this month. Here’s a quick list of things that have surprised me recently:
Are we still virtual?
I’m not the first person to think NotebookLM is great, and I certainly won’t be the last. I did a simple experiment. I pointed out two recent posts: “Think Better” and “Henry Ford Does AI.” Both the summary and suggested questions provided by NotebookLM were pretty good. It goes beyond simply explaining the two parts and goes deep into the relationship between the two parts. But what surprised me was the podcast it generated. It was an 8 minute discussion between two synthetic people who sounded interested and engaged. (Here’s a description of some of the techniques Google uses to make this happen.) Was it 100% accurate? No, but honestly, if a human were to summarize my article, they would probably find a few things to complain about.
Learn faster. Take a deeper dive. Look further.
Being Google, the user interface was a bit clunky after my initial experience. When I wanted to listen to the podcast again a few days later, I had to play the “guess what to click” game one too many times. (Hint: Think you need to click on “Laptop Guide”? Why doesn’t the podcast player show up by default?) But this is actually a very minor issue.
model using computer
Anthropic’s computer-enabled API is now in beta. Beta is right. There is a lot going on here that is dangerous and easily exploitable. But it’s also a lot of fun and opens up a new direction for AI development.
In essence (and I may be wrong about the essence), using a computer teaches Claude how to use a computer: a browser, an editor, a shell, anything that can have a user interface on the screen (and possibly more). Anthropic provides the demo as a Docker container so you can run it safely. Once your container is running, you can give Claude a problem to solve. Figure out how to solve that problem and use virtual Linux machines in containers to do it. For example, we may ask you to create a spreadsheet with data collected from your website. Claude does all the clicking, copying, and pasting for you.
Is this revolutionary? My first reaction was, “This is great. I can upload files to GPT and use them to browse the web.” In principle, yes. However, ChatGPT does not allow web browsing and file uploading in the same conversation. What’s really new? Think about the enormous prompt you’ll have to have GPT read your spreadsheet, determine what data is missing, find that data on the web, and create a new, updated spreadsheet. It won’t be simple. Using a computer takes away most of this complexity.
Does it really go away? If you go further, you will find out. We’re still at that stage where hallucinations and misbehavior are more cute than critical. It’s easy to mislead Claude into interpreting something on a random website as a prompt. This will be a field day for rapid injection attacks. And I can imagine many improvements. Using the computer currently involves taking a screenshot and passing it to Claude, who calculates where to click. This seems very awkward, especially considering that many applications have accessibility affordances that can make screenshots unnecessary.
For now, just relax and breathe. Don’t use your computer for anything serious yet. It’s important to heed Anthropic’s many warnings. But you have to play with it and think about what it means. Selenium++, an automated framework for web application testing? Is it a tool for negotiating with online suppliers? We are much closer to a world full of agents who ask computers what to do and they do it for us.
Is this the end of CRM?
Somewhat along the same lines: Sam Lessin Posted On Twitter for a very clever and useful hack (I won’t call you X). He exported years of emails, extracted the key parts using GPT, and then uploaded them to NotebookLM (yes, I say that again). This allows us to ask questions about conversations over the past 10 years. Who did I talk to? why? What topics did we talk about? That’s all useful information.
Sam argues that this is the end of structured customer relationship management (CRM) software. I would never give an opinion to an investor or founder, but his process immediately resonated with me. I’ve worked with many writers and potential writers for decades, and my emails contain conversations with thousands of people. So, “I want to know more about DDOS. “Who should I talk to?” The first step is to go to Gmail and start searching. Email is my CRM system. I have never used a commercial CRM product.
Unfortunately and ironically, Gmail’s search functionality is quite poor. It might work if you use it for contact management, but it’s not enjoyable. Can I ask NotebookLM? entirely.
Email-based CRM may be a good startup idea, but it’s hard to imagine it being successful in the long term. There won’t be many ‘moats’ to protect your startup from a large company like Google. It’s easy to imagine Google building this kind of AI-powered search feature directly into Gmail. They already have all the data.
That’s it for this month. It wasn’t that bad. Maybe I should do this more often.
