Antonym: The 150 Million Word Limit Edition
Take it apart. Understand it. Make a better one.
Dear Reader
Couple of things we learned in week one of 2025:
In March, you will be able to buy a desktop supercomputer capable of running AI models from NVIDIA, something no one pegged as a possibility in their 2025 predictions thereby immediately making them out of date.
NotebookLM Plus, the paid version of Google’s incredible note-managing, podcast-generating AI tool, will let you make on-the-fly personalised experts with 150 million words of data.
Conversations as a media
Three months ago, Google released an AI-powered research notes product, NotebookLM. It would have been an interesting but niche demo but for a feature called “Audio Summary” which created a realistic conversation in the style of a podcast about whatever notes and documents were stored in a notebook.
The effect was shocking for anyone hearing it for the first time. Literal gasps of astonishment.
But the Audio Summary feature was more than a party trick, it was very useful. Listening to an easy-to-understand conversation about whatever you are researching makes complex or specialist information easier to get a handle on. Even listening to AI-made conversations about topics you’re familiar with can give fresh perspectives as the system tries to explain your jargon to an imagined general audience.
Listening to the product team talking about NotebookLM revealed something fascinating. They didn’t rely on a technical breakthrough to create a feature that feels like an instant podcast. Instead, they carefully designed a process that guides the AI through the necessary steps to achieve the desired outcome.
The second big surprise of NotebookLM was: Anyone can do it.
How NotebookLM’s “audio summary” works
People who like understanding how computers and smartphones work love tear-down videos. When a new product comes out they will take it apart and explain what they can see inside – deducing the design decisions and engineering tricks that have been used to create it.
Let’s do the same thing for NotebookLM from a process perspective.
I made a video, but in case you’re not in the mood, here’s the diagram and prose version…
Your documents are read by the AI and turned into a script.
The script is not very good, but it is taken through a process of being criticised by an editor bot.
A new, better script is created. But if this were turned into voice it would sound like script-reading by slightly bored people.
The new script is passed to a bot that adds disfluencies (filler words, half finished sentences and repetition that feature in natural speech).
Then the script is converted by an AI voice synthesis software into speech with variances in emotion, pitch, cadence etc.
Ta-dah! The conversation is created and ready for use.
Advert from our sponsors
Brilliant Noise AI Power Hours
Accelerate your productivity with a practical 60-minute one-on-one coaching session designed to unlock the power of AI for your specific challenges. Whether you're new to AI or already an advocate, this tailored session will show you how to leverage everyday AI tools to work faster, smarter, and more effectively. Reply to this newsletter for info or book online.
More astonishment: NotebookLM Plus
We have been testing NotebookLM Plus and it is jaw-dropping all over again. Where you’re able to upload 30 documents to a Notebook in the standard product, you can upload up to 300 documents of up to 500,000 words (or 200MB) each, essentially creating a personal database that contains almost enough books to open a small bookshop. That’s per notebook, by the way. And you can talk to it about any of those books or about themes and facts from all of them.
150 million words is:
1,875 books.
Reading War & Peace 268 times.
Or the Harry Potter series 139 times.
4.3 years of reading 8 hours a day.
No. You shut up.
The size of that number gives an insight into the scale of Google’s datacentre network. Currently anyone can access the NotebookLM PLus for the cost of a taxi ride from Victoria Station to Oxford Circus. Or a pizza. Or a reasonably priced bottle of wine.
And then, there’s a new feature which is just as jaw-dropping as that first time you heard the Audio Summary feature. You can join the conversation.
The video’s about 3 minutes long, but if you’re not able or inclined to watch right now, this is what I show:
I show a notebook with some sources about trends that I used for the last edition of Antonym. I load and start the Audio Summary and then press “join now” and ask a question about video games, something not addressed specifically in the trends report. After a pause, the AI conversation continues extrapolating some of the trends and speculating about how they might affect video game play and design.
This feature is impressive. It is a little slow at the moment but it gives hints of how we will start to have conversations with data in the future. I will say that again, because I find it amazing: we will have conversations with data.
Fun stuff
Rude noises
That was way too serious. To shake it out, let’s take a look at a cool feature on the Elevenlabs website: AI generated sound effects. I can trust you all not to use this for silliness, can’t I? Go to the Elevenlabs website and choose the “Text to FX” button.
Can you poison an AI?
MIT Technology Review has details of a new tool called Nightshade that lets artists invisibly ‘poison’ the pixels in their artwork before uploading, potentially disrupting AI training data and damaging image-generating models. The aim is to protect artists’ rights, but there are concerns about how this could affect the effectiveness of AI models at large.
A supercomputer on every desk
For $3,000, Nvidia’s Project DIGITS offers a desktop machine boasting 128GB of RAM, 4TB of NVMe storage, and a Grace-Blackwell superchip promising up to a petaFLOP of AI performance. It’s like cramming a supercomputer into a paperback.
Ideal for taming models up to 200 billion parameters (compressed, naturally), DIGITS also lets you chain two systems together to play with Meta’s Llama 405B.
It’s a tool for AI developers who like their innovation spicy, their form factors tiny, and their wallets slightly lighter. Shipping begins in May.
Who will use it? Contacts and colleagues engaged in AI development were surprised by the announcement, but several saw immediate uses they could put it to, mostly when dealing with very large data-sets.
Recommendations
This week we are mostly watching The Traitors (UK series three), which continues to entertain and delight as human brains are fried trying to manage the dissonance between their almost irresistible urge to be part of a group and the point of the game, which is to lie your way to winning a modest cash prize. Essentially it is full-contact emotional mixed martial arts.
There’s powerful allegory amid the mayhem. In the absence of certainty, contestants try to work out which of them are “traitors” by consistently, constantly conjuring evidence (which is in no way actual evidence) and theories of which they are “hunnerd percent” convinced. Until they are quickly shown how wrong they are.
That’s all this week…
Thank you for reading – there are now 500 Antonym subscribers and I couldn’t be happier than when I’m writing for you.
Antony
NotebookLM Plus is great - the expansion of the number of sources per folder is amazing