Creation of new landing page

8-core processor is fairly understandable, though I might be biased. Knowing what the hell RMSE is requires some background in math, plus we use our own homebrew method of calculating RMSE, which you won’t find in literature. We can show log loss instead, but that also requires some background in math. Also, about the whole “20-30% fewer reviews” thing - it’s a lie. Well, not a lie, but it’s based on simulations. We haven’t measured it directly because it would be mighty difficult.

One more thing, maybe this works because for most people there is an understanding of processor even though not for cores. So they just assume this means “good processing power”.

BTW, we don’t need to tell people about RMSE, just tell them how fsrs is better at memory prediction. A little bit of marketing mumbo-jumbo. Showing a percentage is also a good idea.

I’ll copy what I said on Discord

Honestly, I don’t have a good idea how to “sell” FSRS. Showing graphs with RMSE/log-loss is not a good idea, unless you expect 99% of the users to be STEM nerds. The “20-30% fewer reviews” part is somewhat questionable. I think you would be better off not showing any numbers at all.

I’ll copy what I said on discord.

not reviews. that’s not working. but say “our SOTA algorithm can give you good retention with 30%-50% less workload”

1 Like

I checked “ABC of FSRS” because I forgot the exact wording. It says “With FSRS, users have to do 20–30% fewer reviews than with Anki’s default algorithm to achieve the same retention level.”
…it was mostly written by me btw

Btw, “our SOTA algorithm can give you good retention with 30%-50% less workload” immediately makes me want to ask “less workload compared to what?”
(also, just say “state of the art” instead of SOTA)

1 Like

@landingpageguage I wonder what’s your opinion

Yes we will put a link to the wiki page for the curious. For the non curious, they don’t really ask these questions. But “compared to other apps” can be included, but it sounds weird because other apps also can use FSRS. other algorithms maybe?

Edit: Let’s first see what dae thinks.

I think that there’s not much point in comparing Anki to “competitors” and their algorithms. I think we should be building the case for Anki through it’s own merits. By showing the effectiveness of FSRS v4.5 and Anki as a whole through statistics that can be easily understood by all audiences would be an effective way to promote Anki.

The following examples show how this could be represented provided we can settle on a suitable selection of these stats:

Oh I like that. We are the Apple of SRS world. Anki has no competitors. Anki is THE SRS.

As for the images, I like the last one, we will probably have to change a lot of that. Instead of “we”, saying “our volunteers” something like that. Also the “Number of Users” might be nice.

I don’t think data like “Revenue growth” is something Anki would need :sweat_smile:

One more thing, just like that 55% faster coding we might wanna include something similar if dae agrees. Only problem is do we have data on the average person learning with their sub-optimal methods. If we did then something like “55% faster retention” would look really good. @Expertium what do you think? I remember Piotr Wozniak used to note down his retention data. If there’s similar data available plus data on time spent we can get a very rough estimate. I’ll keep my hopes low though.

“55% faster coding” is, like, shampoo commercial level of bad. 55% of what, what are we measuring here? Faster than who?

Anyway, I don’t really have good suggestions. The number of users (or the number of times Anki was downloaded) would be nice, but we need Dae’s help for that. RMSE and log-loss are for nerds, we need something simple and intuitive. I don’t know what numbers to use to show that FSRS is great.

My feddback is purely on the writing. I would not agree to descirbe Anki as a tool to make learning easier. The learning process is done outside of Anki. You use Anki to review what you have learned in order to keep it in the long-term memory.

1 Like

My question is does that make sense to people who don’t know about FSRS? For example would it make sense to a person if I tell him to “only review things when you’re at the brink of forgetting them”. But now I also realise the problem with the word “learn”. It may not be clear to some people that when we say learn we mean long-term retention of understood facts, would be terrible if people start thinking Anki will spoon feed them everything.

Also @Expertium I think I found where I got the 50% less review data from. I said this in a different post a few months ago.

That’s not very useful without knowing their retention before and after. A comparison is meaningful only when retention is the same both with FSRS and with SM-2.
In order to achieve that, you either need to simulate review history, or get a bunch of people and tell them “Follow my instructions, set Anki up in this specific way”. Sadly, the latter is really difficult unless you can get help from some research institution.

I think people have mistaken the examples I shared for examples of wording or figures to use. These were purely to show what this section might look like. Either way I agree that in its “raw” form, the FSRS RMSE isn’t suitable for all audiences

One thing that might be useful to show tell here is that FSRS personalises the scheduling for you. It adapts to the way your memory works. I agree that RMSE shouldn’t be there. @L.M.Sherlock Do you have better ideas on how to demonstrate FSRS’s strong points for an audience that doesn’t understand RMSE/log loss?


Might be relevant:

[scheduling algorithms] automate the tracking of memory states and the efficient scheduling of reviews.

According to my previous paper, the threshold method (like the current FSRS) reduces 16% cost compared with Anki (SM-2).

1 Like

We’re comparing Anki with Anki? :laughing:

That only makes sense for people who are already using Anki’s SM2. If you were trying to sell Anki (with FSRS) to someone who has never used Anki (and by extension, other software) how would you go about it?

Then we can compare it with RANDOM. 1 - 511/902 * 100% = 43%.

1 Like

Can you explain it in simple terms so that everyone may understand? Is random referring to reviewing at random intervals (aka the most popular method)?

RANDOM, which chooses a random interval from [1, Half-life] to schedule the review.

The half-life is the interval when the retention decays from 100% to 50%.

1 Like