james-revisoai

Haha reading your comment felt like finding someone who speaks my language. I absolutely agree that the context of spaced repetition with incremental learning and reading is key, if you are to apply it.

Studying the psychology of learning, I happened upon the fact this is a real "Learning Loop" with another masters student who explained their problems studying effectively to me.

Basically, when you need to study for something, you often have loads of content. But you want to be able to manageablly learn it - and, in order throughout the learning process to be most effective - do so with spaced repetition/active recall. That requires the incremental reading or context you mention!

I built a way of capturing this reading-flashcards-reading-flashcards mangeable l op into a tool. Instead of just generating flashcards from PDFs and showing them one at a time, it first groups the reading (in this case, for a lecture, by slides covering each topic), and it shows you say the 7 relevant slides for that top, then it moves on to ask you 2-3 flashcard questions for each of those relevant slides (so lets say 18 questions). Crucially, it will keep track of you mistakes and progress on those chunks of flashcards with SRS later and the next day bring up those you forgot. As a learner you can then rotate back into studying via reading, addressing another incremental chunk.

It's so much better than trying to consume a whole lecture or chapter or large concept en masse! By actually contextualising and always feeling like the next bit is do-able, you have enough momentum to study well but also feel progress. This is why I like incorporating spaced repition, but contextually. For other psychology reasons related to study help in exam conditions, it also offers a free recall which tests at the harder end, beyond what most apps do, and does useful things with the results later on.

I do host that learning mode - "Learning Loop Mastery" - on my site, Revision.ai - which can turn lectures into those incremental learning loops via flashcards/reading sections if you want to try it. As of tomorrow it will be widely available (again).

Learning depends on the environment and whether it is pursued in an auto-didactic sense (Even for a job, say) or whether you are learning for an exam/part of a cohort.

It's not wrong to say Curriculum does not matter. But the level of curriculum is also something that needs to adjust to your current level and related fields you have knowledge within, to prevent you becoming overwhelmed.

Most people stop learning being motivation dries up as Test Anxiety rises to the point where they are at a "low-performance" place in the eustress curve. A few days there and people pause until it becomes urgent. A lot of this is a lack of momentum, but also not dedicating or having access to judgements of learning about your own progress.

In other words, if you judge your learning at all, it helps you manage.

There is a natural tradeoff between the flow-state of "just one flashcard with one information principle at a time, endlessly" and the longer term state influencing your time in flow-state of "am I progressing, what don't I know, how do I feel about my learning and mistakes?"

Think about learning databases, or CSS. When did you really takeoff? Probably A) Practically copying others examples (existing queries ran in PhpMyAdmin, or codepen code) And then later B) Once you overcame a big mistake and saw progress - suddenly what "Display" did clicked for you, and you saw how useful it could be to use the "fixed" option, it unlocked your understanding of the items in A and confirmed or disconfirmed your understanding of how it works.

Again it all depends. Self-motivated learning, even for a job, is easier to work with than compulsory learning. Because there, you don't even have the motivation to gaze up to the horizon and gather any excitement or understanding for what the learning might later lead to. It doesn't feel like a path, it feels like a brick wall. In this regard, a list of subjects is somewhat skin to someone stacking bricks, rather than elucidating a path. Overwhelming anxiety while learning is a real thing. The context really matters as to whether this approach is always the wisest.

Please can you tell me more about the daf-2 gene and it's role in lifespan for nematodes? Any explainers on this? How does insulin relate to it - the metabolic aspect?

That makes a lot of sense! I do see value there. I guess it's the context of the class, but biased to the content which LLMs at test-time are bad at understanding/marking (e.g. the items/subjects in MMLU the LLMs still fail at).

As for the structure approach, which interests me - How can you find/note down that information, which you reference in your second paragraph?

I agree it is valuable. In times past, just identifying wrong answers (with Multiple Choice Questions / distractors) could give you great insight into misconceptions in a class (e.g. and that could fill the "Things definitely wrong to mention").

But how do you without a human expert work out "what should the answer have?", in a way that doesn't knock out answers that are left-field but genuine/synthezizing?

My experience of marking exam papers in STEM (only a few courses mind, maybe 300 papers) is that markers are wildly different. Some generally give every pupil +5% higher grades than a more critical neighbour; some are "absolutist" to the course content (either as they learned it, or it is currently taught to the student), and more than would ever admit will freely move scores up 10-20% if the students answer simply novelly interests them, so long as it's not outright provably already wrong.

The available tools to remark and to get 1:1 marking feedback (as offered in Germany after exam season) allow rectifying and smoothing these differences, but in some ways, they are genuine. I just don't know how you can codify it. Even with human markers, there is so much disagreement. I doubt that making 3 different models mark with the same criteria and averaging results/synthesizing a "majority judgement" summarization would work either. It's too hard to make the models care or identify novelty relative to the other answers you've seen that day for the question, if that makes sense.

Hi! Yes we use something similar at Revision.ai to analyze student short, long and recall-based answers. The context of their studying content is considered when providing feedback in more ways than just the "correctness" of their answer.

When you say "leave small details around concepts that LLMs make mistakes on", do you imagine any way to do this for a variety of source material? Let's just say you have the course lectures, and several questions with curt answers. Can that be useful then?

Some of the teachers I have worked with do this; and there are some great teachers(lecturers) like Martin who teach effective studying techniques: https://www.youtube.com/watch?v=IlU-zDU6aQ0

Flashcards and Anki were incredibly effective for me. I am building a study tool which automatically generates them from lecture slides. Beyond that, it applies other psychology of learning techniques - It breaks up content into manageable chunks you can see your progress on (to boost your confidence of exam coverage) and has an exam-conditions like mode where it shows you relevant slides, then asks you questions and to freely recall by blurting onto a blank page everything you remember. It's then able to give you a realistic judgement of your learning and this way of learning by actively recalling freely is, though more intense, ultimately more effective than either reading, watching or quizzing in the long term.

The part of learning that other flashcard apps miss, even Anki, is that students must be motivated. They must enjoy, visually engage, and be confident as they study. They must feel they are on a sustainable realistic path as they gather knowledge when on the way to their goals. And they don't have to study every day to do that - they just need to genuinely learn and be able to see the state of their progress...

I am excited to be working on EdTech now[Revision.ai] which unveils and detects misconceptions to aid the student in eliminating them. Misconceptions are a massive opportunity - as much as spaced repetition itself [https://www.youtube.com/watch?v=eVtCO84MDj8]. We believe they will genuinely learn, feel comfortable and see their progress, when they overcome misconceptions with tech. Other tools really aren't identifying misconceptions; answering is still a "scale of wrong-to-right". And that's a big opportunity. To truly reduce Test Anxiety by studying and building your confidence. I think Tech can be useful in this way, even without teachers intervening, if enough of the learning content is there.

I agree motivation to learn and maintaining interest is key, to effect the learning rate you mention. The enjoyment itself, although a crucifying word to use in discussions when teaching students considered the "best of the best", still matters and predominates over the effectiveness factor for most of the semester for students, at least. Because of just being "told to learn" it.

Basically 70% of the semester most students are not studying 40 hours - they are doing 30 hours of real work, and perhaps only 15 hours effectively. And for good reason: Nothing is bridging or rewarding them in a way that interests/motivates them, for courses where the interest isn't natural.

A bridge to motivate them would be ideal. In 2021 I started using GPT-3 to generate motivational "reasons to learn a concept" cards for my flashcard app, - Revision.ai - which you can read about here in the 3rd item: https://www.revision.ai/articles/20ThingsRevisionAIDoesForBe... - the reason we disabled them was simple: we could never quite time the cards right to help the student when they needed it. When the app is closed, they aren't motivated - they don't see them. Mid study session? Showing such cards (or AI generated examples) interrupted the flow [https://www.instagram.com/p/CVVlIuVg31W/] We have also tried recommending relevant short/mid length youtube videos for visual/"breaks" from overwhelming learning. That did not boost student success either. I guess it doesn't address that you are still being told to do it, not naturally flowing into learning.

I welcome any technical or conceptual ideas you have to improve this and help increase interest amongst students. We have found that turning lecture slideshows into sets of exercises with clear visuals[https://www.instagram.com/p/C5ByftwiJ00/], breaking content up, and showing progress does motivate students to study more(and in a semi-related dissertation I wrote, possibly reduce Test Anxiety and tension). Please let me know of any ideas you have!

I also recommend https://www.learningscientists.org/posters from scientists in the field, which covers additional scientifically-effective approaches that a course designer or more long term approach might take. For example, not only applying Spaced Repetition, Interleaving and Active Recall (all possible through automated spaced repetition apps based on simply input), but elements like Dual coding (related to "Varied practise") - mixing visual and other elements when learning (which requires effort to create them), and elaboration practise (like free recall). To the end of applying these - and other critical elements like focusing on the motivation of the learner and enabling them to understand what they are missing with progress & identified "blindspot" misconceptions (on their own incredibly powerful), I've been developing Revision.ai since before GPT-3, through a Psychology MSc.

Yes, I have often thought if it wasn't got GPT-3, this work would have been more recognised and more powerful. However generative embeddings are still more useful for more abstract cases and do embed a different space to these similarity/contrastive kind of embeddings.

I think it is. But there are levels to learning and approaches that tech can take. Speaking as a Psychology of learning and Comp Sci graduate.

For a long time we have had the capability to use student learning record data to recommend optimal questions to maximise student grades in a given subject(adaptive learning); we've known the best exercises and methods to apply this with, like quizzes/active recall (see Roediger & Karpicke, 2006) and increasingly, made interfaces that keep people in the zone.

The problem is, such systems have 3 challenges compared to a human experienced tutor/small group teacher:* - 1. It's not abstract. It can't identify misconceptions in your wrong answers like a human would and refocus new activities on these gaps. Doing this is very effective. I believe this is possible in the future, hence why I developed blurt mode, a learning exercise on my website Revision.ai which asks students to blurt content they have learned on a specific subject, rates their coverage and accuracy and most important identifies, clarifies and stores for future exercise content their knowledge gaps, by using AI tech to identify misconceptions in a way we could not do 4 years ago. - 2. An app will never truly check up on you socially or with social expectations of appearing in real life like a tutor; it's never as urgent. Staying engaged for 6-7 hours rarely happens either unlike real life with it's physical location constraints (which also have a role in memory formation and retrieval) - 3. The perfect app needs millions of student learning records that account for your specific experience and knowledge. This is not possible. While humans cannot do this either, intuition of teachers is lost with apps.

*Small groups/tutors can provide customised, targeted and long-term feedback which you feel a sense of accountability for responding to according to the research from Bloom, and is largely responsible for why wealth correlates with education with higher attenuation at the top quartile than the differences between the bottom three quartiles - paying for private tutoring really is a cheat to achieving more on grades (and therefore more opportunities), although whether it improves intelligence, is debatable(nobody agrees what intelligence is really, beyond some early G factor work that personally I do not see as that valid).

I think you should be careful using the word "personalized" too - one, it is a popsci buzz word, and two, and consider "adaptive" instead - they have different meanings in the psychology of learning and like I address above, the personal aspect are mostly confined to humans.

# People they delighted or prevented from feeling frustrated? The problem is, this is rather hard to measure.

I own touchba.se and use it to set up meetings/calls. @Circleback team you could take it on the cheap, but I think Circleback makes a lot more sense for your company.

This makes some great points, and Martin Lombell has covered something very similar in a rather timeless one hour engaging lecture here: https://www.youtube.com/watch?v=IlU-zDU6aQ0

Some other diamonds I've uncovered or found while working on learning psychology are:

- that students who organise fun and exercise on their schedule before studying, get better grades than the students who start by deciding when to study;

- that most of us can learn much better with confusing, misconception-revealing content, over simple content - we all assume too much, as Veritasium covered: https://www.youtube.com/watch?v=eVtCO84MDj8&list=PL772556F1E...

- in principle, a desirable difficulty of varied content given in a medium that isn't lazy (read: you need to be able to get feedback, like a quiz, or rewriting, or tutoring) is the key way to establish viable long term learning; several other effects like Spaced Repetition only have benefits if you can stick to your motivation for learning, so you do need to prioritise motivation first. The article mentioned learning styles not being proven; in fact, one can argue the opposite is true; people who learn with mixed, varied content perform the best. But we also like the learning type which is comfortable for us; this boosts motivation. The problem is that studying students across a whole semester and multiple courses rarely happens, and studies focusing on how varying the learning content within a 10 minute window changed how much students learned, does not replicate the value of being familiar with a learning content style, nor the cost that things within it might blur. There are pros and cons to sticking to one style; variation certainly helps in the short term.

- enjoyment(or motivation) matters. Because almost every student reports procrastinating early in the Semester (95% in one study, and I guess the other 5% didn't fill in the form). A real issue with learning research is that it tends to replicate exam period, exam conditions, and not the 60-80% of students lives where many put off studying, gently attend lectures, socialise and work between semester exam seasons. It is in that time that results are predetermined, not in the final weeks. That's what my learning app for students focuses on, and achievable regular cadence of studying in that period; motivating and fun enough to engage before your peers are. It splits content like your lectures up into memorable emoji-represented exercises, and then sends you notifications using the emoji for one exercise each day, to get you to just study a little each day. This has big downstream effects.

- the quality of learning content is as variable as the quality of laptops; some is poor. Key concepts like the "Minimum Information Principle" - that questions based on building definitions, linking concepts and memorizing should have only one short clear answer - are almost necessary if you want to learn a serious amount without forgetting what matters and blurring answers. Long answers, although they seem better because of detail, will drop your grades as you accidentally confuse or forget part of them, as compared to separating them into different or say follow up cards.

What's the Ed Tech startup?

50k-200k LOI* in January would be unusual for an EdTech startup given the timing of semesters and university/locality/school contract signing periods.. is there more to the story, a distribution channel, or hyper-local focus?

I'm the technical founder of Revision.ai, feel free to email for a chat or advice. An AI Ed Tech startup made widely available in 2019 with traction. I had a similar choice at one point.

*(Letters of Intent? For seats? or is this MAU?)

I'd be interested in discussing this with you - we think alike, although, having background in psychology too, I was at first sceptical of your use of terms like "Rearranging". But I like your summary and how you orient it to the programming domain but also reference examples of good learning app techniques.

I am working on an app that makes many decisions in this area, and is truly trying to attempt the psychology of learning ethically to university studying (including parts of computer science you can learn without actually "doing" the coding etc). I have a few ideas and new learning user flows I would love to get your feedback on as well as have a wider discussion and nerd out a bit on psychology studies I think you'll find revealing, would you be up to talk?

As a quick peak: My own thesis on student learning broadly begins with three e's: "Effectiveness, Enjoyment(or motivation), and Environment", as I believe the first two are necessary, and the third one an augmentation, as properties of effective learning systems for people. This is based on combining my psychology knowledge (mostly the first and last two E's) with my experience volunteering and being a TA and hearing from students who through 90% of the semester struggle with the middle e - Enjoyment - more than others, or to the point where it prevents applying the others(not motivated to use effective techniques, not motivated to go to the library when procrastinating, etc) correctly. I am super interested in combining this with the teachers view and how they work too (e.g. curriculum design, personalised tutoring)

Is this monthly revenue or total so far? Recurring? What do you costs look like?

I was running one of them, and entering kaggle competitions throughout 2021 and 2022 using them. Many efforts and uses of Sentence-transformers (and new PhD projects) were thrown in the trash with Instruct GPT models and ChatGPT. I mean it's like developing a much better bicycle (lets say an ebike) but then cars come out. It was like that.

The future looked incredibly creative with cross-encoders, things like semantic paths, using the latent space to classify - everything was exciting. A all-in-one LLM that eclipsed embeddings on all but speed for these things was a bit of a kill joy.

Companies that changed existing indexing to use sentence transformers aren't exactly innovating; that process happened once or twice a decade for the last few decades. This was parents point I believe, in a way. And tbh, the improvement in results has never been noticeable to me; exact match is actually 90% of the solution to retrieval(maybe not search) already - we just take it for granted because we are so used to it.

I fully believe in a world without GPT-3, HN demos would be full of sentence transformer and other cool technology being used for demos and in creative ways, compared to how rarely you see them.

What features struggle to work self-hosted?

It was my (perhaps outdated) experience that in fact some elements of JS (like using scraper libraries) would not work on Vercel at the time due to memory of serverless function and other limits. But locally they worked on my dev machine.

Good points... In many ways, before LLMs, vectors were getting so exciting, Sentence Transformers and BERT embeddings felt so instrumental, so powerful... work by the txtai author (especially things like semantic walking) felt incredible and like the next evolution. It's a shame in a way that all the creative and brilliant uses of text embeddings from similarity embeddings didn't really have any time to shine or go into product before ChatGPT made so much except search use cases obsolete..

Sounds like you may be asking the wrong questions. As you may be assuming the product direction is correct and uniform for your customer base - but is it? The pricing sounds ok, I just doubt you'll get useful information if you have small scale and no current users. You may confirm some people will pay, but they may be wanting to pay for different functionality to what you build. Some evidence of willingness to pay without churn data is pretty useless for improving your chance at success unfortunately. Now it does decrease your chance of building something nobody -ever- wants, but... that's not the same thing, and you should care about the former really.

In a third of the time you'd spend setting up payments and describing the idea with marketing, you should perhaps ask people about their problems, then mention your solution contextually, and then listen carefully to what they say.

Do all the responses talk about it doing what it does now? Then charge.

Do they talk about mostly one clear functionality/vision they all want? Great, charge and build.

Do the responses cover multiple ideas/functionalities? That's bad, it means you have to find a niche with a plurality of expectations, or this will never work because the expectations are too varied for a cohesive profitable offering, or tend towards integrated visions that you cannot compete with on the build power of a small team. Sometimes people don't ask for what they turn out to want (e.g. iPhone, Notion itself with great design and low stress), but this is the case for very few successful startups, and if you look carefully, they all tend to rapidly respond to feedback post-launch anyhow.

I heard this general idea from Jason Cohen - invalidating - and it's incredibly useful. This sounds like the right scenario to apply it.

I do think Googles strategy for UI is pretty strong, and avoids their products feeling outdated or changing too much to flow. However, I am particularly sad that now rather than "Inspect Element" being right next to my cursor as the bottom option when I right click (if the menu appears above), I now need to hover over a small area to get it to scroll down into view, often overshooting, and then fuss about just to inspect the element. I did grin a little when I realised the only other change I noticed in my workflow was that the context menu now had rounded corners.

If you can afford to optimise for your "Recall rate" over your "Precision" as such(e.g. your current time or number of opportunities/ideas is high), then go one step further than the Mom test and what people here say:

You need to listen to the diversity of what people think your solution will do and enable and you want a clear plural answer. If your problem is "urgent" and the solution is "obvious" the way people think it will work is too segmented, you can still have something that fails on unit economics because too many people have their expectations broken. (Learnt this from Jason Cohen). Really you need to talk to at least 10-20 people and try different questions, different approaches. Some ask about the problem. Some review paper designs with. Some, ask them to sit down and show you how they do the task. Generally very few ideas will be both an urgent problem people would pay for that people ALSO think will work and output in a fairly uniform way amongst the stakeholders.

I appreciate your and the other commentators scepticism, and agree for the premise of generating text. However to play devils advocate - Can't inference of single layers be quite useful sometimes for embeddings? and the larger models could embed capabilities only achievable at large parameter size? (i.e. you don't need to finetune the last layer like the BERT days or necessarily use a certain layer for relevant embeddings, you could use the first layer outputs and finetune a custom matrix projection to make some useful embeddings that might include properties of the input that the larger models show capabilities for that smaller models don't?)

(ofc this assumes that those capabilities considered "emergent" at larger sizes do not require later layers which they probably do). There is a weird situation where we have the sentence transformer, the bge, e5 and etc kind of embeddings, and then a big jump up to generative model embeddings the model providers provide, but not much widespread adoption of e.g GPT-J or neo-20B embeddings (even though at time of release, they had notable usecases over sentence transformers)

Not necessarily; much smaller models like T5 which in some ways introduced instructions (not RLHF yet) did have to include specific instructions for useful translation - of similar format to those you find in large scale web translation data, but this is coincidental: you can finetune it with whatever instruction word you want to indicate translation - the point is, a much smaller model can translate.

The base non-RLHF GPT models could do translation by prefixing by the target language and a semi colon, but only above a certain amount of parameters are they consistent. GPT-2 didn't always get it right and of course had general issues with continuity. However, you could always do some parts of translation with older transformer models like BERT, especially multilingual ones.

Larger models across different from-base training runs show that they become more effective at translation at certain points, but I think this is about the capacity to store information, not emergence per say (if you understand my difference here). You've probably noticed and it has always seemed to me 4B, 6B and 9B are the largest rough parameter sizes with 2020 style training set ups that you see the most general "appearance" of some useful behaviours that you could "glean" from the web and book data that doesn't include instructions, while consistency seems to remain the domain of larger models or mixed expert models and lots of RLHF training/tricks. The easiest way to see this is to compare GPT-2 large, GPT-J and GPT-20B and see how well they perform at different tasks. However the fact it's about size in these GPTs, and yet smaller models (T5 instruction tuned / multilingual BERT) can perform at the same level on some tasks implies that it is about what the model is focusing it's learning on for the training task at hand, and controllable, rather than being innate at a certain parameter size. Language translations just do make up a lot of the data. I don't think it would emerge if you removed all cases of translation / multi language input/outputs, definitely not at the same parameter size, even if you had the same overall proportion of languages in the training corpus, if that makes sense? It just seems too much an artefact of the corpus aligning with the task.

Likewise for code - Gpt-4 generated code is not like arithmetic in the sense of the way people might mean it for code (e.g. branching instructions / abstract syntax tree) - its a fundamentally local text form of generation, this is why it can happily add illegal imports etc to diffs (perhaps one day training will resolve this) - it doesn't have the AST or compiler or much consistent behaviour to imply it deeply understands as it writes the code what could occur.

However if recent reports about arithmetic being an area of improvement are true, I am very excited, as a lot of what I wrote above - will have to be reconceptualised... and that is the most exciting scenario...

Self-hosting Supabase is pretty hard when you consider mail set up, and all the config like setting your own domain/secrets/config without the full dashboard Supabase managed service provides, doesn't it? Aka you can't just get-up-and-go fill in a nice UI/form on an EC2 instance AMI template for Supabase...

I have tutored 3 students in Math for a term or more in the UK, here's my thoughts:

- Student motivation is crucial. Your first duty is not to be a cardboard tutor, to be interesting, different, to stir something. Without this nothing later really matters (to me it is equivalent to pursuing a relationship where there is attraction, but not laughter, fun and relaxation - it's a base ingredient)

- Tutors who uncover and help students see misconceptions and knowledge gaps do demand more mental focus, but are much more helpful; these (e.g. visiting last weeks homework mistakes) are incredibly useful mid session too to break up sessions and to illustrate progress.

- You really face a choice depending on the students motivation. I have tutored 3 maths students, 2 of which were on path to fail at their age 16 qualifications, who I did my utmost to stir interest in with relation to video gaming, science and literature respectively, and, by way of being more relatable than their teachers, helped them pass; and one who (long before ChatGPT) I recognised immediately that they were incredibly talented academically and talented at listening, but it did not show in grades because of a lack of confidence and their social beliefs about themself. After a few sessions, I introduced them to machine learning libraries (at the time sci-kit learn/nltk) far beyond anything on their syllabus and showed them beautiful ways to use maths later on to do incredibly interesting things like nearest neighbour search for conversation sentences in Friends episodes. I barely mentioned the maths until after the fact in these stirring few sessions... but once I did they knew they learned eagerly...

They latched on with such energy that their work was compared to that of someone 4 years their senior, and they (from a below average UK London school) got a place in Oxbridge (I cannot express how proud I was of them).

I think it goes to show self-belief and social-beliefs are so important to focus on. As a tutor you -can- make a difference, you -are- the fresh and more relatable start compared to a teacher. And a last important fact is be ready to use metaphors.

To really excel as a tutor I think you should attach math with wonder to something they like. For a fan on the new Matrix movie, it could be the mysterious RSA and P - NP problem. You are a revealer, a magician. Where possible hold some information in apprehension, for the next session. Make it a journey. The teacher must address every student, your work is for them. You are showing them a door to a new way of learning the world, and that door is made of a new material, that they will naturally want to learn about so they can walk through it!

Have a look at "Consensus" App.

I really like the points about symbols to the end, but am not sure the Music notation hits the right note. The symbols - besides Quarter note - for a rest/not playing are intuitive to almost every learner. They are just so different, so square, so alien compared to the notes in melody, that you don't make the mistake of playing a rest note in general.

Yes, it's hard to learn the symbols that involve playing something (especially clefs and timing not being linear - see: https://www.youtube.com/watch?v=Eq3bUFgEcb4 for an amazing "documentary" on notation systems), but often finding the pitch/right note is the larger challenge (for beginners) than understanding the meaning of notes. Is this because we do not teach people to play songs with only one note, but many rhythms? I don't know.

But I do not think the point of affordances in notation is a great example, especially given all the alternatives (see the video above) which, although in some cases clearer (e.g. a instrument-morphic notation), struggle in other ways (too much space used, too unable to represent timing, not transferrable between instruments, etc).

The reality is also that our mind can adapt easily in such distinct systems as music notation to shifted realities. Sheet music for different instruments sometimes uses the same clefs, bars and all, but is several notes a part (eg... the music of Violin and Viola are notated differently, such that the notes of the respective instruments more or less span the whole bar without going too far above or below for the extreme notes of the instrument). Musicians usually have no trouble learning this abstract switch, even though the affordance, is just one symbol at the beginning of each bar - the groove of playing a different instrument adapts it. I know this does not seem the case for UI, which are perhaps just "too similar" to each other and merge and feel unnatural, or maybe the issue is that while using a new UI - like the github comment selector being expanded - other sites may keep the old UI and confuse our mental model of afforances - but for music, it's just not the problem I see modelled in the article...

Available yes, but many are "premium/aftermarket" pricing now. There is such a high proportion of these names/prices popping up when you do domain search compared to the dominant "Unavailable" of .io or .com.

Source: I've registered .ai names since 2019. I use a 5 and 8 letter fairly common english word for two .ai domains; both were standard pricing originally, but comparable available domains are now >$10k. However, I could probably at best resell my domains for < $1k each, as resell value for .ai is hard once you use them. I guess investor dollars and being seen to do something might due the demand for these premium domain options.

Thanks for clarifying. I think I am just misremembering how terms were used because of the cohort of people I was amongst and our focus on business models, perhaps.