AI Expressing Empathy, Freaky
You can read, take notes, and then pull it all together the old-fashioned way. Or, power your insight step with software.
The Teardown
Wednesday :: September 25th, 2024 :: Approx. 10 min read
👋 Hi, this is Chris with another issue of The Teardown. In every issue, I cover how we interact with technology that enables our day-to-day lives.
If you’d like to get emails like this in your inbox every week, subscribe below.
And, if you enjoy today’s thinking, let me know by tapping the Like (❤️) icon in email or on the web!
Where Is My Notebook?!
You’ve spent weeks combing through books, guides, papers.
You’ve talked with experts and friends.
And you’ve amassed, throughout those efforts, tons of research.
Now you must synthesize that research into something digestible. All of the physical and digital chicken scratch needs to be blended from chaos into learnings that stick to your neurons.
But, all the sudden, you don’t have your notes. They’re gone. They’re missing.
And now, you can’t synthesize in any short span of time. The notes you collected over time and without an assumption of loss might need to be reconstructed unless you unrealistically write every prior note verbatim somewhere new.
Let’s face it. You probably can’t.
So, what if you don’t have to start from scratch? Can you synthesize most things much faster?
Let’e explore that today.
Short answer: maybe.
No Moore’s Law In Synthesis
My last two posts explored a possible practical use of AI: therapy, or more broadly, self-help through coaching.
Talking about practical usage is something lots of you want to know more about. Some of you are using ChatGPT or other similar tools in your every day. But many are not at all or not in any useful way. The technology hasn’t proved it deserves a place in your active thinking.
Jim MacLeod is an author that writes about how AI is changing how we think about design. He posted on Threads about the ways people engage technology when making decisions:
The idea of "desktop decisions" has changed how I think about UX. Many users prefer the desktop experience for certain types of information and activities. Especially if it's something that requires their full attention.
I thought this phrasing was profound. And it reminded me that the research process that I illustrated at the beginning of this post is a desktop experience, for me anyways.
What that means is not precisely that you work or consume on a computer desktop. Instead, the process of ingesting, documenting, and synthesizing information requires a significant cognitive surface area.
And if you think about physical notebook, it is just that. An organized (being bound) but very large surface on which you collect ideas, notes, and other bits of information. A digital product like Notion is no different in basic form. You can collect, clip, type, and store an enormous amount of information.
At some point, you might want to halt your collection process and do something with the information. Maybe you’re on the verge of a decision. Or perhaps you need a refresh on what you wrote or collected at some point in the past.
There is no right way to perform that synthesis step. I think you can agree that the process is very much trial and error for you, though you might read a book or twenty about how to do it better.
Over time, this process of collecting, synthesizing, and summarizing builds on itself. You end up with a trove of notes about a topic or about thousands of topics. Prior summaries are also notes.
Perhaps they’re labeled, stored, and replicated in case of emergency. You trust your inner reference librarian to dig through that archive when you don’t remember something.
But it’s probably not efficient to do so.
Until, now?
Supercharged Research
Google recently released an experimental product called NotebookLM (“Notebook”) into the wild. It is ambitious (emphasis mine):
NotebookLM is your personalized AI research assistant powered by Google’s most capable model, Gemini 1.5 Pro
Collaborate with a virtual research assistant.
When you upload the documents that are central to your projects, NotebookLM instantly becomes an expert in the information that matters to you.
Go from information to insight in a snap.
The emphasis in the quotes illustrates two claims that I’ll file under marketing, for now. To better understand how expertized and instant Notebook could be, I turned it on me.
First, let me explain my intentions.
What I write here comes together through a collection of motions that probably aren’t unique to me. I write drafts that I love, hate, or sigh at. I sometimes start a post not to write in that moment but instead to later recall something I’m thinking about. There is a growing list of potential topics in a Notion page, some with bits of thoughts, some not. And, I’ll sometimes document in the moment with audio notes via Apple Notes, or Apple Voice memos, or more recently Superwhisper - a free tool that sits locally (i.e. on my phone) on top of OpenAI’s speech recognition model called Whisper.
I also comb through old drafts to recall themes or bits of writing that catalyze new pieces or drag me away from moments of writer’s block.
So, I have a good sense of what I’ve written this year, but not a perfect comprehensive word-for-word memory of posts, drafts, and every imaginable theme. You, as the reader, know less since you’re paying attention only some of the time and don’t see my drafts.
With that framing, I fed Notebook every published post from this (2024) year. That corpus contains well over ten thousand words. And, then I started poking and prodding.
I thought I might ask my new research assistant: what are the key themes that summarize the blogs content? Note: is this a blog? I have no idea what to call it.
My uploaded source posts are in the left bar and interactions with Notebook within the larger right pane:
What do you think? Did Notebook capture the essence of my writing? Of my soul?
I think it accomplished three things:
The content of the summary points is good. Notebook seems to understand the various themes that I weave in and out of my posts. I don’t see glaring inaccuracies.
Notebook is not evocative. These summaries look like they come from someone that needs some hobbies. They are bland. Google believes research super stars are kind of bland.
It referenced the source of its summarizations. This is important. I’ll come back to it.
Notebook’s summarizing doesn’t, at least in my view, need to be evocative. I’m ok with bland.
What I wanted was a digestible summary. I wanted to outsource the process of understanding everything I’ve written to more efficiently query that learning.
But Notebook can be evocative if you give it permission. How? Well, in a very human-like way: verbal conversation.
The Teardown … Podcast
As a disclaimer, I’ve not researched far and wide for other audio-based AI conversation products.
And, I might not bother at this point.
Podcasts are extremely popular these days. Everyone has one. Many famous podcasts have spawned all sorts of adjacencies, like Atomic Habit influencers, or AI that allows you to converse with the content in Andrew Huberman’s very long and often controversial speeches.
Some podcasts are personality compliments to the written words or professional happenings of their hosts. The sighing, stumbling, word choice, and sentence phrasing say communicates beyond the edited and structured words in blogs, or newsletters.
The opposite is also informative, as polish and speaking prowess adds to a host’s allure. Some writers sound like themselves on their podcasts, like what you might have imagined as the tone, pitch, and speed of their voice. Some are the opposite.
Most of you hadn’t heard me speak until I added audio overviews in two recent posts. My wife called them “cute and well delivered” so I guess we will stay married a bit longer.
I added those overviews for a few reasons:
To provide you with a concise review of a given post.
To experiment with how people engage with my posts.
To practice speaking about my writing without all the ums, ahs, sighs, and other things that flow so easily when lots of people speak. I am immeasurably impressed by people that have cut junk words and sounds from their vernacular.
So, you might have thought, will these overviews turn into a podcast? Do I get to hear your wonderful baritone tone, Chris?
We’ll see.
But, in the meantime, here’s The Teardown Podcast. I encourage you to listen to the entire twelve minutes if you have a little bit of time on your hands.
What The F—K Moment
Did you feel it too? I was mesmerized by what I heard and how I heard it.
Those podcast hosts oozed energy. They synthesized almost ten months of my writing in just twelve minutes. And, they seemed to really like each other, in that platonic bordering on flirting way.
And of course, they weren’t real. They were podcast hosts conceived, trained (on my writing), and voiced by Notebook.
Wow.
The Details Matter
In non-virtual life, podcast hosts often come under fire for saying things that might not be exactly right upon further examination. Joe Rogan is well known for it. Andrew Huberman is more recently a target. You have examples, too.
What you sometimes hear is apologies or restractions or restructured thoughts, especially if the hosts care about feedback and about the integrity of the information they verbalize. There are honest mistakes and outright lies and everything in between.
Notebook is supposed to be grounded in the source information you provide. But Notebook, like OpenAI 4o and o1, sits on top of a prediction-oriented LLM that may hallucinate.
So, what does that mean for Notebook’s podcasts?
Two things:
It might not understand the source material as intended during the ingestion/learning/something phase.
The virtual hosts might say something that isn’t accurate.
Only after I listened to the entire 12 minutes did I remember that I was still dealing with LLM technology. Notebook performed a Jedi Mind Trick on me.
Yet, I noticed a few problems:
The hosts open their show The Deep Dive (😂) by repeating my first and last name to the audience. Then, they call me Cocuzzo for the rest of the podcast. Like they’re my buddies. And let me tell you, they aren’t my buddies, guys. I an appreciate the colloquial handling of my otherwise great name, but don’t hear anything like it on other podcasts - run by humans. This issue isn’t one of information but of style.
Around 2:50 or so, the hosts claimed that I was trying to construct a “perfect Instagram-able” coffee-brewing process as part of a goal to build a community around coffee. Uh, no. In the source post, my overall point was simple: make and cherish time for things you enjoy, like coffee.
At 9:00 and later, the hosts talked about how I’m organizing golf outings for other fathers in the area. The summary point was not entirely untrue, as I played quite a bit of golf with friends this summer, but I was not posting flyers and organizing formal outings. So, I thought this summarization was a significant stretch. Pure lack of truth? No. But not right either.
I think the most important takeaway is this: NotebookLM will save you tons of synthesis time. And, it will help better validate ideas by maintaining a cognitive surface area that is very difficult to replicate. Unlike a human, NotebookLM doesn’t seem to forget anything it understands.
Every note, correction, and data point you feed it adds to a growing bank of query-ready summarizations that you don’t need to develop yourself.
But, like most everything else in AI right now, you will still need an intuitive of right, wrong, accurate, or inaccurate.
So, will I use NotebookLM? You bet. I won’t be regurgitating Notebook’s word for word here but do plan to use it to synthesize research for some upcoming posts.
Stay tuned.