Between trip, end-of-year initiatives, the approaching holidays, and different hysteria, I haven’t provide you with an article this month. So right here’s a fast checklist of issues which have amazed me not too long ago.
Are we digital but?
I’m removed from the primary particular person to seek out NotebookLM superb, and I definitely gained’t be the final. I did a easy experiment: I pointed it at two of my latest posts, “Think Better” and “Henry Ford Does AI.” Both the abstract and steered questions NotebookLM supplied had been fairly good: They went past merely commenting on the 2 items and bought into the connection between the 2. But what blew me away was the podcast it generated: an eight-minute dialogue between two artificial individuals who sounded and engaged. (Here’s an outline of among the strategies Google places to make use of to make it occur.) Was it 100% right? No, however truthfully, if a human summarized my articles, I’d in all probability discover just a few issues to complain about.
Learn quicker. Dig deeper. See farther.
Being Google, after the preliminary expertise, the person interface was greater than a bit clunky. When I needed to return to the podcast just a few days later, I needed to play “guess what to click” method an excessive amount of. (Hint: Would you guess that you should click on on “Notebook Guide”? Why doesn’t the podcast participant seem by default?) But that’s actually a really minor drawback.
Models utilizing computer systems
Anthropic’s pc use API is now accessible in beta. Beta is correct—there’s clearly lots happening right here that’s harmful and simply abusable. But it’s additionally numerous enjoyable, and it factors towards a brand new course for AI improvement.
In essence (and I’ll have the essence mistaken), pc use lets you inform Claude the way to use a pc: browsers, editors, shells, something that may have a person interface on a display (and probably extra). Anthropic supplies a demo as a Docker container, so you’ll be able to run it safely. Once the container is working, you may give Claude an issue to resolve; it’s going to determine the way to resolve that drawback, and use the container’s digital Linux pc to do the work. For instance, you may ask it to fill out a spreadsheet with knowledge it collects from web sites. Claude will do all the press, copying, and pasting for you.
Is this revolutionary? My first response was “Big deal, I can upload a file to GPT and use it to browse the web for me.” In precept that’s true, though ChatGPT doesn’t enable net searching and file importing in the identical dialog. What’s actually new? Think in regards to the monstrous immediate you’d have to get GPT to learn a spreadsheet, discover out what knowledge was lacking, search for that knowledge on the internet, and generate a brand new up to date spreadsheet. It wouldn’t be easy. With pc use, most of that complexity disappears.
Does it actually disappear? We’ll discover out as we get additional in. We’re nonetheless on the stage the place hallucinations and misbehavior are cute relatively than crucial. It’s straightforward for Claude to be misled into deciphering one thing on a random web site as a immediate. It will probably be a discipline day for immediate injection assaults. And I can think about loads of enhancements. Computer use at the moment works by taking screenshots and transport them to Claude, which computes the place to click on. That appears extremely awkward, particularly on condition that many purposes have accessibility affordances which may make the screenshotting pointless.
For now, calm down and take a breath. Don’t use pc use for something critical but—it’s vital to heed Anthropic’s many warnings. But you must play with it and take into consideration what it means. An automated framework for testing net purposes, Selenium++? A instrument for negotiating with on-line distributors? We’re a lot nearer to an agent-filled world the place we ask a pc what to do and it does it for us.
Could this be the top of CRM?
Somewhat alongside the identical traces: Sam Lessin posted on Twitter (I gained’t name it X) a few very intelligent and helpful hack. He exported a few years of e-mail, used GPT to extract key elements, and uploaded it to NotebookLM (sure, once more), which permits him to ask questions on his conversations over the previous decade. Who did I speak to? Why? What are the subjects we talked about? That’s all helpful info.
Sam argues that that is the top of structured buyer relationship administration (CRM) software program. I gained’t supply an opinion for traders or founders, however his course of resonated with me instantly. I’ve labored with many authors and potential authors over the a long time, and my e-mail consists of conversations with hundreds of individuals. So after I need to ask a query like “I want to understand more about DDOS; who should I talk to?” my first step is to go to Gmail and begin looking. Email is my CRM system; I’ve by no means used a business CRM product.
Unfortunately and mockingly, Gmail’s capability to look is sort of poor. Using it for contact administration, although it may be made to work, isn’t nice. Can I simply ask NotebookLM? Absolutely.
Email-based CRM may even be startup concept, although it’s arduous to think about succeeding long-term. There wouldn’t be a lot of a “moat” to guard a startup towards bigger corporations—like Google itself. I can simply think about Google constructing this sort of AI-enabled search immediately into Gmail. They have already got all the info.
That’s it for this month. That wasn’t so dangerous—possibly I ought to do that extra usually.