a list for 'against comparison'

Explore

a list for 'against comparison'

Using Coda AI to classify research

Last edited 274 days ago by Gareth Simpson

⁠

Coda's AI block classifying text from news articles using 'inference'.

Coda's AI block classifying text from news articles using 'inference'.

⁠

TLDR - Coda is an exceptional tool for classifying research.

PR & comms, investagive journalism, sports leagues, editorial diaries, local media

Focus on a realively narrow subject & topic to start >700 texts of >12,000k each.

Don't automate scrapping the web to start, have a human editor.

⁠

Preparing for an interview

I was interviewing for a role at SongTradr, a company that has grown over a decade through funding rounds and acquiring businesses at a global level. I was aware that Songtradr’s values would be new to a lot of the teams I could be working with they would have established cultures formed under different ownership.

I take an interest in corporate values - the poetry of them - collaboration is an art, not a science - I have worked in creative organisations that practice what they preach here - legacies of leaders like Charles Allen and Michael Grade both now appointed Lords - there more ‘community’ organised and that excel in cross-functionality.

I was unsuccessful as a candidate - however, if life gives you lemons, than make lemonade and I can’t really document this use-case for

Coda⁠

’s AI without revealing that it was Songtradr and that I didn’t get the job.

So, how would I go about researching SongTradr and trying to put data-points on something pretty anecdotally like an example of company values? Build a database of course! 📚

⁠

Searching Google News for results

I started by building a pipeline using

Pipedream⁠

, a low-code automation tool similar to Make or Zapier.

Thanks to the community here at Coda, there is a

pack⁠

that keeps me in Pipedream’s web far more than it should given the costs of using it.

I choose

Playwright⁠

for my site scrapping, which in hindsight I regret.

Not only was it complicated, each article required a manual trigger for automation anyway.

Scraping requires a template for each different source, and would be better suited to monitoring a consistant output - ie a hashtag on x, an RSS feed, *.CSVs

⁠

Mitigating your database getting populated with unnecessary data requires a human editor.

Mitigating your database getting populated with unnecessary data requires a human editor.

⁠

I highly recommend manually coping and pasting research into your database yourself.

By doing so, email mailing lists proved a valuable source, which can be ‘off-the grid’, lonely without a URL, on a spreadsheet.

Additionally, human editors best handle ‘news-jacked’ summaries - duplicate reports without a new treatment or source.

⁠

Classifying your data

Select lists

⁠

⁠

For ‘more tangible’ fixed values- ie places, people, events, publishers, companies, assets... I highly recommend starting out using select lists.

⁠

⁠

I was super impressed how quickly AI auto select learnt the sources of data from the text alone.

There where a few stubborn sources, such as Music Business Worldwide, which took seven manual interventions until the AI took over -

Anecdotally, this might happen when you abbreviate things, in my case MBW.

Select lists also a great place to start as they can be converted to a table later.

⁠

Table Relations

For more analytical classification such as sentiment I had to provide a description of information of what I wanted to measure - which was done in separate table called Sentiment.

⁠

Coda's AI can be forgiving

Coda's AI can be forgiving

⁠

⁠

❗making changes on tables relations after your database is populated will require a lot of AI compute.

If you haven’t upgraded your price plan yet, changes here will tip it over the edge.

⁠

Field Properties

Most fields in Coda have editable AI properties that can populate values from other values. Here I populated the date field from our content.

⁠

⁠

⁠

I also used AI field properties to calculate the word count in the source content.

⁠

Summarisation

AI Block

Coda’s stand alone ‘AI block’ has many use-cases. Here I have created ‘Question to ask’ field where I could prompt answers from the content database.

⁠

The value of the context window is populated the word count of how the content is filtered by topics and date

The value of the context window is populated the word count of how the content is filtered by topics and date

⁠

Summarisations options, especially length and type, allow you to create new data that can be repopulated back into your database.

Learning how to ‘repopulate-to-optimise’ performance is an essential skill-set to scaling up and working with much bigger data sets.

⁠

⁠

Context Window

The context window for Coda is about 12,000 words. Below you’ll note I have word count calculated as a field. Filtering the content by topic and a range of dates, I was able to manage how full the context window.

⁠

⁠

I ‘hard-coded’ the word count instead of using AI, a kin to a formular on an excel spreadsheet, as its a computationally less intensive.

⁠

Learn how to manage you context window.

Learn how to manage you context window.

⁠

⁠

I then linked this to a context window slider next to question to ask, visualising how full the window is.

⁠

SongTradr and Values

Consciously Collaborative

You understand ‘team’ and collaborate with kindness and respect.

When SongTradr launched its first product in 2017, it introduced a two-way marketplace. Both musical creators and advertising agencies were required to subscribe and purchase a form of ‘token’ power-up. This setup allowed demos to be pitched to specific briefs. Everyone had to pay, fostering a healthy environment for a new business. This model, I speculate, likely introduced a self-policing mechanism against excessive briefs and dubious pitches.

Humbly Confident

You’re confident but are never arrogant.

This concept was challenging for AI to categorize, I’ll grant it a ‘hall pass’ here though. I was particularly interested in identifying aspects linked to cash flow positivity. SongTradr’s approach to acquisitions exemplifies humble confidence as the purchase prices are never disclosed, a practise we have become less accustomed to in recent sync-tech investment times.

Diligently Driven

You’re self-motivated, hard-working, and get results.

When you research SongTradr, it could just be my TV background but a story becomes quickly apparent and it is one of its founder, CEO Paul Wiltshire.

The ability to tell a story is a staple fund-raising disciple and AI’s classification of diligence often happened in tandem with statements related to financial stability.

Songtradr, straight out of the gate, was a global business and in decade it has grown into just that, - pretty diligent on Paul’s part -as he has done what he said.

Innovative Solutions

You’re positive, open, and passionate about finding innovative solutions.

I could write excesses granular detail whats insights I’ve learnt about SongTradr’s PR and comms strategy - however, ‘Innovative Solutions’ is central part that- that is the message they want everyone to hear.

It’s one that holds up too when you compare it to Songtradr acquisitions strategy - as these have been driven by technology offerings and applying them to new markets & distribution pipelines.

Good Business for All

You do what’s good for the business.

Another hard for AI to classify and was frequently applied to content that was external opinion and market speculation rather than something more tangible.

‘Customer Success’ stories where another area I struggled to measure and I anecdotally think they are related - gaining reach is challenging when tell those, nor did I sample any Songtradr marketing materials.

‘Humble Confidence’ is a poetic example of how to conduct yourself in the music industry, known for its ephemeral, every transitory cultures - ‘Good Business’ was an area I identified where SongTradr could improve their external messaging around.

⁠

Final Thoughts

I wanted to end this with an AI generated summary about SongTradr, completely unedited by me.

I actually can’t do that in my current template, owing to the limitations of how I implemented the

context window⁠

and the amount of content I populated in the database.

In my experience using the stand-alone

AI Block⁠

, the notion that AI is a ‘probability machine’ becomes quickly apparent.

For example, when I asked my database about SongTradr and blockchain technology, it generated text suggesting that BandCamp is being used for a product launch to learn more about the technology.

This isn’t implausible, for a ‘fastest-way-to-fail’ blockchain product for the purposes of ‘social-impact’ in our community, as the AI suggested- BandCamp would be the obvious choice.

In this instance, the AI was drawing from a ‘thought-piece’ written by a 3rd party, published by DJMag with BandCamp being a party mentioned as part of a wider market landscape context.

To mitigate how the AI block synthesises data using probability, I’d need to think about how to classify opinion by an external author and what is a tangible factual statement..

indexing tangible facts would also help using Coda to calculate market sizing, something I also wouldn’t attempt to do given my current understanding.

Classification was the unexpected winner though, I just cannot believe how quickly Coda’s AI picked up the reigns and starts reliably doing its jobs. It’s incredible.