I’d heard that the NotebookLM feature that creates podcasts was cool, but it’s wild! I’ve used it several times now. Here, you can listen to its summary of my chip tasting shorts. These are 60 second videos where my daughters and I try different chips - like foie gras, holiday stuffing, or just BBQ. You just give it some stuff and it comes up with a male host and female host doing a two people talking podcast. Somehow it figures out what to talk about and adds commentary to it.
By default, it comes out as slightly cliché, sometimes getting to levels of like Christopher Guest movie but without the jokes. But, other times it’s pretty amazing. Sure, it’s “slop,” but to gets close to human-made slop. The summarizing it does it good too.
I’ve done many more, including one five years of our State of Kubernetes survey (pretty all over the map) and Jeffrey Dahmer (my wife listens to a lot of true crime podcasts, so I wanted to get her take on the output).
Using a tip from “Optimal-Fix1216,” I figured out how to put “producer notes” into the podcasts. These are simulated notes that a producer has given the hosts. It allows for some guiding of the content: putting in pauses for ad breaks, being a little more sarcastic and telling jokes, and maybe making the episodes longer, like 20 minutes instead of 8. Here’s a generic one, and here’s one for doing a Software Defined Talk episode. You add these notes as sources in the NotebookLM notebook and - maybe? - it uses them to guide the two. Sometimes it obviously work: they’ll do an ad break. And maybe instructing them to go longer works too. Here’s a 32 minute one going over my book of my The Register columns, Digital WTF.
I’m trying to get them to do a scat musical intro/jingle thing, but no luck yet. However, they’ve started burping in some episodes…?
It’s fun to make and listen to these. Here’s some theory on uses:
Summaries, quick learning. One of the first uses I remember hearing about is making a personal podcast. For example, get those ponderous academic PDFs and have it make a discussion of each.
Modeling analysis and thinking. Related, I’m thinking it might be good supplemental tutoring for my kid’s school. My son is starting to get into “write an essay about this book” classes. Maybe these two could have some discussions about the books. Sure, this could be cheating if he just converted what they said into an essay, but more of what I’m interested in is modeling how to think about that kind of literature analysis. Also, my kids might be more aural learners than readers - or they could be both! Multi-channel learning.
Story pitches and outlines. Another type of modeling could be writing corporate copy, blog posts, podcasts. For example, you could take the keynotes from a big conference, throw in blog posts, even put in analyst and blogger commentary. I did this for our big conference, Explore. The Audio Overview did a good job of picking out themes and coming up with with something to say about them. Sure, it was basing that on the content I fed it. But, have you ever sat in front a pile of content and tried to figure out what to write about it? The generated episodes are a good start. And, you could just convert them back to single author prose - I haven’t tried that yet.
Instant reactions. Let’s say you’re at a big tech conference and just watched their keynote. If you had a recording of it, you could give it to Audio Overview, wait ten or 15 minutes, and have an instant, on the floor take. If you pre-loaded the notebook with a bunch of context and background material, it might make it better. Based on the experiments I’ve been doing, these would be really helpful. Would you publish them? I mean…yeah?
Simulating press, analyst, and influencer interviews. When you’re doing a big announcement, you do a lot of press/analyst/influencer interviews. What are they going to ask? What would you answers be? What are those people going to say on their weekly podcast or round-table video wrap-up? You can get the start of an idea from the Audio Overview. Of course it’s not going to be perfect or reliably predictive. That’s not the point: it’s to get started.
Brainstorming. Because it’s a podcast, the two hosts are sort of figuring out what they think on the fly. That’s the way a lot of podcasts are - especially the ones I do! In that way, if you gave them some rough ideas, you could simulate a brainstorming session: getting some people in a room, starting with basic concepts, and working through what to think of them. An important aspect here would be that you can re-generate the episodes over and over. So, you could have three, five, etc. brain-storming sessions. Here, you could chain things together soon. Have it generate five sessions, convert them back to text, and then have another AI session pick the best and worst (and neglected) ideas from those sessions.
Anyhow! I recall seeing people automate some of this with the APIs for Gemini. I of course just use the web interface. There’s some whacky world where you have an AI pick out interesting news stories in the morning for you and then generate one of these episodes, puts it in your podcast feed, and then you can listen to it in the morning.
“Surprisingly interesting podcast from a text file of pure blank spaces”
“But it’s not about catching criminals. It’s about catching vibes.” Bop Spotter
“(I think I keep Instagram more intimate than the average person. My gauge for if I follow someone is ‘If I saw them at the grocery store, would I stop and talk to them for awhile?’)” Riley Walz
“Often programmers use “die” to mean “decline in relative marketshare”, not absolute marketshare.” Here.
“The candidate or the couch.”
In enterprise vendor marketing, it’s easy to fall into the trap of only talking about the benefit and outcomes of a product. Often, you also need to sell the idea of the product itself, and even the educate people on how to use it. Kubernetes wouldn’t have been successful if that community only talked about the benefits; you had to see all the demos for years of how it worked, and also be convinced that this is how infrastructure should be managed and how applications should be packaged, architected, and configured to run on Kubernetes.
“Whatever it is you want to do, somebody has already figured it out for you.” Here.
“The ribeye served with a smile over clean linen is fine, but it’s got nothing on tacos unceremoniously dropped on a plastic table you can afford to share with someone you love.” Here.
“their big ‘this is what we are doing’ conference…” // Great phrase for vendor conferences.
RTO or GFTO - RTO is like a dress code.
Survey Says: Tech Spending Is Up, But AI Rollouts Slower Than Expected - Initial AI hype is finally going back to reality. // " The AI wave is still building, but the future has been slower than anticipated. Today only 5.5% of identified AI use cases are in production, a sobering reality check on respondents' Q1’24 projection that 52% of identified use cases would be in production over the next 24 months."
CEO Kurian: ‘When I Started, Most People Told Me We Didn’t Have a Chance’ - “We’re now the fourth largest enterprise software company.”
The Top 1 Percent Paid a Lot of Taxes in 2021 - ”In 2021 the top 1 percent of taxpayers in the United States paid 36 percent of all federal taxes (they have 21.1 percent of income). This figure had been below 20 percent until the mid-1990s, and as recently as 2019 it was just 24.7 percent (they had 15.9 percent of the income that year).”
gabrielchua/open-notebooklm - Prompt for generating a podcast: “You are a world-class podcast producer tasked with transforming the provided input text into an engaging and informative podcast script. The input may be unstructured or messy, sourced from PDFs or web pages. Your goal is to extract the most interesting and insightful content for a compelling podcast discussion.”
The top five AI lessons learned from IBM’s presence at SaaStr 2024 - ”Build solutions that solve a specific enterprise pain point” // as we said in the PaaS days: “find the app.”
Enterprises funnel IT spend into AI and data, Accenture says - ‘Enterprises are shifting IT spending to investments that support data modernization and AI technologies rather than increasing overall budgets. - “What we are seeing is that as they’re saving money, they want to invest it in things like AI and data,” Sweet said Thursday. “That’s really the dynamic that’s going on: save to invest.”’
Platforms Engineering - Not a fan: “Platform engineering isn’t gluing a bunch of off-the-shelf tools together. More templates, variables, and YAML isn’t going to make it easier for application developers to deploy faster, reduce cognitive load, and still innovate (a topic for another time). And the CFO isn’t going to fund multiple platform teams–at least not part of the same organization.”
McDonald’s touchscreen kiosks were feared as job killers. Instead, something surprising happened - If you can come up with other ways staff can spend time, automation could be good for workers and customers. For example, come up with more complex products that require more human touch, speeding up delivery, or just busing tables. But, yeah, if you’re not going to change your business, then automation probably means firing people.
Why So Few Matt Levines? - There’s not enough news with the same types of events that happen over and over with clear, public records.
Some advice and good practices when integrating an LLM in your application - Yup!
I mentioned re-starting the Software Defined Interviews podcast. We just recorded our first episode yesterday, and it’ll be out tomorrow. Make sure you subscribe to it to start listening. I’ll of course put a link in here next episode.