TL;DR — Please use Sterling Chat to Import or Generate Expressions for the time being. Over the course of the next few months, you will be able to upload any audio, video, or text file into Sterling Chat, and we will automatically create a Material for you (complete with live audio and video clippings if they are provided)

Phrasing as a product is broken down into 3 pillars:

1. Create (creating expressions, segmenting and explaining, multilingual)

2. Review (humane-srs, graph based reviews, multilingual)

3. Materials (books, movies, tv shows, podcasts, music, multilingual)

Realistically, all three of these came together this year, in a very MVP fashion.

~Q2 of this year was around reviews - getting the algorithm working smoothly so it was fast, effective, and enjoyable to use. After months of tweaking, it got to the point where I’m always caught in that _just one more card_ state of mind that I was aiming for.

~Q3 of this year was around expressions - making sure that the explanations were on point, testing and fixing several different languages (looking at you Japanese), improving the alignments, and adding new ways to examine Expressions that I didn’t grok. After months of poking and prodding, everything is in place - I can understand every expression, and I constantly go down rabbit holes understanding how each Expressions works.

~Q4 of this year, I will give Materials that same treatment. This is a little write up of what to expect, what went wrong, and what I’m looking forward to.

What to expect from Materials 1.0

Drag and drop any video, audio, or text file into the chat interface, and get back an interactive, readable copy of the file with synced audio & video.

You’ll be able to ask Sterling for recommendations, and it will scan all of your Materials to provide to you the best (highest leverage) recommendations to learn.

I’ll be working on features in roughly the following order:

• uploading private audio + video to parse it into Quotes with native Audio + Video

• uploading private epub, mobi, txt and pdf books (or other unstructured content)

• improved reader view

• extracting public subtitles

• a few extra surprises :)

Materials 1.0 will be faster, cheaper, and also largely private. There will be a massive de-emphasis on shared materials in the beginning, until the feature is stabilized, then sometime next year I have some plans for how to share the materials without infringing on copyright.

What went right with Materials Beta

Currently, there is an extraction process in place that in March, was able to process millions of events with a 100% success rate, despite the chaotic nature of subtitle tracks. It’s able to take unformatted text, and produce bilingual texts, with deduplicated words, and provide recommendations based on what you know, and what is important to the show.

There are a lot of issues when it comes to scale that were sorted out - how do you load millions of words and lookups in an efficient fashion? How can you serve recommendations for users across billions of words in tens of thousands of documents? How can you parse thousands of sentences without slowing down the whole system? How do quotes link and load into Polaroids? How do you deal with multilingual translations of the same content, and even further, multiple translations in a single language? How do you search a large catalog of data quickly, filtering by language, etc? How can you do the three way merge required between various data sources? These were are worked out and dealt with in the initial beta.

The design was also polished, and got to a point that I’m quite happy with. Some UI patterns emerged that looked great on desktop and on mobile, and adapted to fit the varying content available.

So the good news is not much will need to change on the backend, as it’s mostly just clean up and refactoring. Almost all of the updates will be UX related - how it works as a user.

What went wrong with Materials Beta

When I set out to build Phrasing, I really wanted the experience of a user coming in, typing in their favorite show, and seeing a bunch of expressions ready to be learned. So this is what I built, and this is what existed in Materials Beta. The expectation was that over time, users would extract more and more movies and tv shows, eventually populating the library to the point where users rarely needed to spend credits.

However, there are a few issues with this approach:

1. The extraction process, both for Materials and for Expressions, are constantly improving. This means that all of the public extractions, especially those over a year old at this point, are incredibly dated.

2. The quality of subtitles is, frankly, a nightmare. I think they’re largely good enough to learn from — but I do not think learning from them will neither make you fluent, nor are they close enough to unlock comprehension of the actual material as fast as I would like.

3. Almost every single user preferred generating expressions to start with than finding suitable materials. Despite the fact that this was not a pleasant experience, and was only ever intended as a testing feature for developers, users still always started with this.

Is Phrasing still useable in the meantime?

Absolutely! One thing that I was very surprised about when I started learning Phrasing was just how little I actually used Materials Beta. Perhaps once a month I would go into an Material and get some new content to learn, but besides that, 99.99% of my time is spend in reviews or Sterling Chat.

Humane SRS is the best on the market, and Expressions are capable of explaining even the most complicated of text to users of any level.

How do I get Expressions for now?

There are three things that I would recommend:

1. Copy & paste Expressions you would like to learn into import. If you’re learning from printed material, you can use your phone’s camera to copy the text and paste it into Sterling Chat (there is an “OCR” option in the import section just for this case)

2. Give your content to an LLM (Grok, Claude, ChatGPT) and ask it to create flashcards from quotes. These days you can upload entire transcripts and get pretty good recommendations. You can copy and paste this into the “Import” feature in Sterling Chat, and we will turn them into expressions for you.

3. Lastly, you can generate Expressions with Sterling Chat. The quality of this will vary wildly by language, and it always helps to get them verified by a native speaker if you can.

In our testing, the recommendations LLMs provide are on par with what Phrasing provides new users. Where Phrasing’s recommendations really shine is after you’ve been using the product for a few months, and Phrasing has learned what you know, what you struggle with, what you’re already studying, and what you’ve never seen.

By the time you get to that stage, Materials 1.0 will be launched. Huzzah!

In the meantime, I recommend everyone to just start using Phrasing for reviews. In the Sterling section, you can either generate expressions (have an LLM come up with them) or import expressions you get from elsewhere (textbooks, classes, or even conversations with an LLM). We use an LLM to parse the text, so as long as the message is within the acceptable message length, we can parse any format into Expressions ready to be extracted and learned.