Very Quick Shaeda Update
Firstly I just want to say thank you to everyone for signing up and expressing interest in trying the app out.
To be completely honest the demand and the amount of signups and emails has been a little crazy - I definitely did not expect it to be this popular after just a month of adding the waiting list.
I appreciate some have been waiting for a while, so below is just a very quick update on where the app is in no particular order:
TLDR:
Beta release was pushed back slightly to fix and improve some parts.
After I add the main core feature (the ability to save cards to one’s own personal database for future review) and sufficiently test it, the app will be released.
Some new features, languages and ~STEM topics have been validated (and improved).
I am debating adding a “My Sources” feature.
The plan was for beta to be completely free but due to the (very) large amount of sign ups it would just be financially impossible. As such, there will likely be a very (very) small monthly fee that will be far cheaper than other AI-related study apps that lack the features shaeda does and have not been tested and validated extensively for accuracy.
Beta Release
The launch has been delayed slightly while the ‘save-to-database’ functionality has been fully implemented and is working as it should.
This ‘save’ feature is essential for the spaced repetition system that makes flashcard learning effective (and the bugs annoy me)
I could release beta now without this, but I’d prefer to wait and have the main feature in place to allow effective studying right away. Saving of cards that one deems beneficial and important is key to the effectiveness (and time-efficiency) of flashcards - and thus the app
This also gives more time for testing and further validating.
Development
Testing a “Choose your AI/Model” feature
For most users the default model will be more than enough, however the ability to have a smarter model may be desirable for those studying at higher proficiency levels for the most complex languages and/or topics at ~MSc level in very niche areas.
Select different model depending on preference and use-case
This comes with trade-offs of course (much slower, unknown if it will generate meaningfully more accurate content, much more expensive etc)
For context here, one of the most difficult languages for AI models to accurately translate is Arabic. If you invest a lot of time into the prompting and model settings though, you can get good results. I’ll provide the accuracy report for Arabic below:
Conversation Mode in Arabic. Here, both questions and the answers were in Arabic. 
‘Paperclip’ feature has been added
Generates related words or sentences based on the flashcard currently viewing
This will be really helpful for situations where there’s very similar words or ways to phrase something that you may prefer to save instead
For example, Paperclipping ‘words’ on a flashcard that says “Today I felt tired” would generate new cards such as: asleep; to sleep; bed; morning; night; to rest, etc.
Paperclip: Generate new cards that have related words or sentences.
Testing a “Check It” feature for both Language study or Academic study
Some cards will be generated intentionally containing 1 or more errors. Some cards will not contain any errors. You would then Check It (the card) and try to identify any potential error/s.
The ‘back’ will either reveal the error/s or state that there is none.
It’s designed to test actual understanding and comprehension rather than just rote memorisation of the card, an occasional worry if considering using flashcards.
Example for Spanish:
Front - El perro es muy bonitas.
Back - 2 errors: gender and number agreement. El perro es muy bonito.
Example for Physics:
Front - In a DC circuit, if you keep the resistance constant, the voltage is inversely proportional to the current.”
Back - Misstates Ohm's Law (V=I×R). In a DC circuit, if you keep the resistance constant, the voltage is directly proportional to the current
With Check It, the app will intentionally generate some cards with either word and/or sentence errors for language study, or factually incorrect statements for academic study.
Debating adding a “My Sources” feature
Here users simply upload either their own PDFs/notes and the app will generate flashcards based off of this content rather than it’s own.
Hundreds of apps of this kind already exist online and there is seemingly little to differentiate them (or justification for the price)
However, I see a few potential issues with this approach, some of which I’ll summarise below:
Hidden gaps due to technical details:
LLMs suffer from a little-known problem known as “Lost In The Middle”. As the name suggests, the model effectively remembers the ‘middle’ parts of a large input worse than it remembers the beginning and end. This affects the length of content one should send in. If the user is not aware of this and inputs, say, 20 pages of a textbook (or worse, 20 pages of hand-written notes) the app will generate a collection of cards, but it would likely have missed (or misremembered) the middle parts. The result of this is a collection of, say, 60 cards, of which 90% were from, say, pages 1-7 and 14-20 and the user would likely not have noticed. Or in the unlikely event they had, they now need to go back into the “submit” section, go back to pages 8-13 and now resubmit these middle pages and go through it again. This is all needless friction and wasted time.
Narrow Knowledge:
The user would likely assume that any cards generated reflect the entirety of what they need to (or should) know for that particular segment. As such, their knowledge essentially becomes overfitted on this very particular set of cards, or ‘spikey’. They learn this fixed set of cards incredibly well, but outside of this isolated knowledge they have (hidden) gaps. This ‘gap’ risk can also be seen in research.
This is part of the reason behind “Deep Dive”. Does the user know the topic or just know the card?
Opportunity Cost:
It’s not clear to me whether this way (creating cards from one’s own notes or sources etc) is even necessary or beneficial for many languages/subjects
As said, there’s a hidden opportunity cost built-in. That is, time spent rewriting/rereading (to produce the notes) is time that could have been better spent on more effective methods. If this time investment did not produce a meaningful improvement in quality (which is what I’m positing for the majority) then it’s actually had a less beneficial effect. Note that this does not mean the user would have not learned anything. Just not as much as they could have if they had instead re-read/wrote slightly less and ‘flashcarded’ more.
However a definite use-case for a “My Sources” could be for languages and/or very complex niche subjects that have not (yet) been validated.
Although even on this, below I will provide some feedback on niche subjects at MSc level and in other languages (i.e., essentially a 3-fold level of complexity: the subject, the level and the language all simultaneously)
Physics/Engineering fields at MSc level in Mandarin, Malay and English
(More on) Accuracy Validation
Many more languages and ~STEM fields have been validated
You can view the list on the updated landing page here (newer signups to the waiting list may already have seen these)
I still require the following languages: Czech; Finnish; Georgian; Greek; Hindi; Hungarian; Indonesian; Korean; Swedish; Dutch
If anyone is either a native speaker and/or teacher of any of the above and would be interested in testing please email me.
Some languages actually failed. For example, one teacher deemed the Mandarin to be ‘correct’ but only up to around B1/B2. After this it would sometimes word too literally rather than how natives would say. Fixing or improving this shouldn’t be too difficult though, so I’ll have the 2 teachers retest after making some updates.
Pricing
Initially I was planning on making beta release free for all but there is just far (far) too many signups now that it would be financially impossible, so I will clarify the plan for beta access in the coming weeks.
Some important notes here:
Very early members who do sign up to any paid tier will be grandfathered in, meaning as the app becomes more developed with better features and AI models gradually become smarter and faster etc the price you pay remains the same for as long as the subscription is maintained (along with of course having a greater say on future updates and features etc)
10% of proceeds will be donated to charities supporting, but not limited to, animal welfare, those with Down syndrome and those with Alzheimer’s or dementia
All remaining revenue will be going straight back in to development
You can view some additional pricing info on the landing site here
That is all, thank you for reading and again thank you for the (slightly crazy) level of signups.
If you know of anyone who may also be interested in trying beta, feel free to share this with them as well.
And if anyone has any questions at all please feel free to reply to this email.
Thank you!
PS: If anyone is interested in joining and supporting the development of shaeda, please do reach out.