I love listening to In Our Time with Melvyn Bragg and guests (official site here). It’s the best radio.
There are almost 1,000 episodes (it has been broadcasting on Radio 4 since 1998) and when I want to learn about, like, Ancient Greek tragedies, or the evolution of teeth, this show is where I turn first. All the audio is online, which is amazing! Thank you BBC!
But actually trawling through the back catalogue is hard.
So I made a very unofficial website to find old episodes to listen to.
Braggoscope lets you explore the In Our Time archive. Check it out!
There are multiple ways to explore:
- Directory - browse episodes by topic, organised using standard Dewey decimal library codes
- From an episode page, e.g. Le Morte d’Arthur (Jan 10, 2013) you can pivot on guest (each discussion has three academics) and similar episodes which are surprisingly good.
Each episode links through to the BBC website so you can listen. The full show description and reading list are included too.
I guess I would call this pre-pre-alpha… there’s no real design yet, and there are surely some bugs with the data.
HOWEVER: I’m using it to discover new episodes already.
For example. I loved learning about the late Devonian extinction recently. Here’s the episode page. Now I can go down the reading list to find books, and find my way to similar episodes about the Permian-Triassic Boundary, the Cambrian Period, the fish-tetrapod transition and so on. Like I said, surprisingly good.
And browsing the Directory is super fun.
(Oh and hi to anyone from the Beeb who is reading this! I’ll take this project private if you need me to, or share the approach.)
HEY: you can stop reading here unless you want all the stuff about how it works and my opinions about AI.
For posterity… in the event that Braggoscope changes URL or disappears, I want to remember what it looked like in years to come. So here are a couple of screenshots: the Directory; an episode page.
I used GPT-3 for the heavy lifting and now I have oh so many opinions
I wrote up the tools I use in Braggoscope on the About page, but as a quick overview.
The process:
- Spider the official website and fetch all the HTML (a 1,000 pages or so, not too many). The show notes are what we’re interested in, but they’re not super well structured.
- We extract data like the episode synopsis, guests (names and affiliations), reading list (title, author, etc) in a machine-readable format. Imagine starting with prose, and ending up with tidy columns in an Excel spreadsheet: that’s what I mean by structured data.
- The data can then be enriched by classifying it and processing it to figure out similarities.
- Finally there is site build where the data is written out into HTML.
I’m a casual coder but the above is pretty straightforward.
Except the extract step. This is tedious. It’s a few days of writing fiddly code to catch all the different ways that guests might be listed, or how show notes might be written.
OR:
It occurred to me… why not just give this to OpenAI’s GPT-3?
So that’s what I did. It took 20 minutes to write the integration, then I left the code running overnight. It costs me pennies per inference so I’ve replaced a few days of boring graft with $30 on my credit card.
And this is interesting right?
I’ve been used to thinking about generative AI and LLMs (language models) as smart autocomplete.
But this is more like a universal coupling.
I set temperature=0 – this is a parameter that governs creativity, so by doing this I was asking GPT-3 to be pretty deterministic.
In the prompt, I specified that GPT-3 should return structured data as JSON (a data interchange format based on Javascript objects) and provided a type definition.
It doesn’t always return valid JSON. I have some wrapper code that fixes it up.
It was while I was getting structured data back for the synopsis that I thought: I wonder if I could get GPT-3 to classify this? How about using Dewey decimal classification…?
And sure enough, it works! It’s not perfect but it’s preeeeetty good.
(Now I read down the list of Dewey Decimals classes with some considerable side-eye. It has, uh, a particular perspective. And it turns out that Melvil Dewey was a seriously bigoted and unpleasant guy. But it’s a well-known hierarchy that is small enough to wrap your arms around, and it makes topics findable. So… I would love an alternative but that’s for another day.)
The “Similar episodes” list also uses OpenAI – each show synopsis is translated into an “embedding,” a ten thousand parameter vector representing its position in the “meaning space” of the language model. Then similar episodes are simply nearest neighbours (calculated with cosine similarity).
Again - this is surprisingly good! While I was developing Braggoscope I tried using tags too but honestly, for finding related shows, this embedding approach is way better.
This is pretty technical but you can explore the whole space yourself: here are all the episode embeddings in a single chart (hover over each dot for the title). This uses PCA (principal component analysis) on the embeddings, then the top two components (being the most significant vectors of variability) are the x and y axes. It’s code that OpenAI provides but will be pretty easy to customise - PCA is a ton easier than when I used it back in undergrad! - so I’m thinking about what to use this for.
I feel like this programmatic use of LLMs is where AI gets really interesting.
There’s the experience of it…
Using GitHub Copilot to write code (as previously discussed) and calling out to GPT-3 programmatically to dodge days of graft actually brought tears to my eyes. I’ve coded, mostly as a hobby, my whole life – it’s a big creative outlet alongside writing – it’s so rarely felt like this. It feels like flying.
But the actual literal engineering of it too…
Sure Google is all-in on AI in products, announcing chatbots to compete with ChatGPT, and synthesised text in the search engine. BUT.
Using GPT-3 as a function call.
Using GPT-3 as a universal coupling.
It brings a lot within reach.
I think the magnitude of this shift… I would say it’s on the order of the web from the mid 90s? There was a radical simplification and democratisation of software (architecture, development, deployment, use) that took decades to really unfold.
There is so much tooling to build around temperature=0 language model calls. There’s a startup or nine just in that.
I would like to see frameworks and programming languages that have first class support for this as a pattern.
Anyway!
Hey, some trivia: I was involved in setting up the In Our Time podcast, way back in 2004. It was the first podcast by the BBC, and the BBC was the first national broadcaster to do any podcasting at all. I hand wrote the first XML files that were uploaded to the servers! Still a fan.
I love listening to In Our Time with Melvyn Bragg and guests (official site here). It’s the best radio.
There are almost 1,000 episodes (it has been broadcasting on Radio 4 since 1998) and when I want to learn about, like, Ancient Greek tragedies, or the evolution of teeth, this show is where I turn first. All the audio is online, which is amazing! Thank you BBC!
But actually trawling through the back catalogue is hard.
So I made a very unofficial website to find old episodes to listen to.
Braggoscope lets you explore the In Our Time archive. Check it out!
There are multiple ways to explore:
Each episode links through to the BBC website so you can listen. The full show description and reading list are included too.
I guess I would call this pre-pre-alpha… there’s no real design yet, and there are surely some bugs with the data.
HOWEVER: I’m using it to discover new episodes already.
For example. I loved learning about the late Devonian extinction recently. Here’s the episode page. Now I can go down the reading list to find books, and find my way to similar episodes about the Permian-Triassic Boundary, the Cambrian Period, the fish-tetrapod transition and so on. Like I said, surprisingly good.
And browsing the Directory is super fun.
(Oh and hi to anyone from the Beeb who is reading this! I’ll take this project private if you need me to, or share the approach.)
HEY: you can stop reading here unless you want all the stuff about how it works and my opinions about AI.
For posterity… in the event that Braggoscope changes URL or disappears, I want to remember what it looked like in years to come. So here are a couple of screenshots: the Directory; an episode page.
I used GPT-3 for the heavy lifting and now I have oh so many opinions
I wrote up the tools I use in Braggoscope on the About page, but as a quick overview.
The process:
I’m a casual coder but the above is pretty straightforward.
Except the extract step. This is tedious. It’s a few days of writing fiddly code to catch all the different ways that guests might be listed, or how show notes might be written.
OR:
It occurred to me… why not just give this to OpenAI’s GPT-3?
So that’s what I did. It took 20 minutes to write the integration, then I left the code running overnight. It costs me pennies per inference so I’ve replaced a few days of boring graft with $30 on my credit card.
And this is interesting right?
I’ve been used to thinking about generative AI and LLMs (language models) as smart autocomplete.
But this is more like a universal coupling.
I set temperature=0 – this is a parameter that governs creativity, so by doing this I was asking GPT-3 to be pretty deterministic.
In the prompt, I specified that GPT-3 should return structured data as JSON (a data interchange format based on Javascript objects) and provided a type definition.
It doesn’t always return valid JSON. I have some wrapper code that fixes it up.
It was while I was getting structured data back for the synopsis that I thought: I wonder if I could get GPT-3 to classify this? How about using Dewey decimal classification…?
And sure enough, it works! It’s not perfect but it’s preeeeetty good.
(Now I read down the list of Dewey Decimals classes with some considerable side-eye. It has, uh, a particular perspective. And it turns out that Melvil Dewey was a seriously bigoted and unpleasant guy. But it’s a well-known hierarchy that is small enough to wrap your arms around, and it makes topics findable. So… I would love an alternative but that’s for another day.)
The “Similar episodes” list also uses OpenAI – each show synopsis is translated into an “embedding,” a ten thousand parameter vector representing its position in the “meaning space” of the language model. Then similar episodes are simply nearest neighbours (calculated with cosine similarity).
Again - this is surprisingly good! While I was developing Braggoscope I tried using tags too but honestly, for finding related shows, this embedding approach is way better.
This is pretty technical but you can explore the whole space yourself: here are all the episode embeddings in a single chart (hover over each dot for the title). This uses PCA (principal component analysis) on the embeddings, then the top two components (being the most significant vectors of variability) are the x and y axes. It’s code that OpenAI provides but will be pretty easy to customise - PCA is a ton easier than when I used it back in undergrad! - so I’m thinking about what to use this for.
I feel like this programmatic use of LLMs is where AI gets really interesting.
There’s the experience of it…
Using GitHub Copilot to write code (as previously discussed) and calling out to GPT-3 programmatically to dodge days of graft actually brought tears to my eyes. I’ve coded, mostly as a hobby, my whole life – it’s a big creative outlet alongside writing – it’s so rarely felt like this. It feels like flying.
But the actual literal engineering of it too…
Sure Google is all-in on AI in products, announcing chatbots to compete with ChatGPT, and synthesised text in the search engine. BUT.
Using GPT-3 as a function call.
Using GPT-3 as a universal coupling.
It brings a lot within reach.
I think the magnitude of this shift… I would say it’s on the order of the web from the mid 90s? There was a radical simplification and democratisation of software (architecture, development, deployment, use) that took decades to really unfold.
There is so much tooling to build around temperature=0 language model calls. There’s a startup or nine just in that.
I would like to see frameworks and programming languages that have first class support for this as a pattern.
Anyway!
Hey, some trivia: I was involved in setting up the In Our Time podcast, way back in 2004. It was the first podcast by the BBC, and the BBC was the first national broadcaster to do any podcasting at all. I hand wrote the first XML files that were uploaded to the servers! Still a fan.