@retoor I'm not sure, it probably depends on the model configuration/implementation and equipment. But in the browser, for that "homemade GPT", I see that training on WebGPU is around x100 - x1000 times faster than CPU

@retoor I've tried to train ~80M GPT parameters on a single GPU in the browser so far. Pretty heavy. It is interesting to see how 1.5B parameter will behave...

@jonas-w Ah! Cool!

@Lasoloz thanks for pointing that out! The typo should be fixed after the Showcase cache is reset.

@hack I believe the internal implementation and the data model can't be changed, but the state can.

With the example of a mixer and its attachments, we can say that the mixer "can rotate any attachment that you give to it". Different attachments may be treated as a state change, however, the "rotation engine" stays hidden and not changed.

@electrineer let doctors decide if that "Sportsman <-- Powerlifter " inheritance is valid :D

@Hazarth as you've mentioned "I feel like the high-level module Plug should depend on the abstraction of wire" - that was exactly what I wanted to show by using the Socket as an abstraction of wires. The user doesn't care what wire types are there behind those two socket holes. But maybe semantically the word "Socket" was a bit misused. Those may be replaced with "Aluminium Wires", "Coper Wires", and just "Wires" (which is the "Socket" on the sketch)

@Kampouse yes it is similar to Excalidraw and Tldraw. However, one distinguishing feature is the ability to have nested drawings. The example in this rant illustrates it. You can click on S.O.L.I.D. letters and go deeper into other nested drwaings.

@Lasoloz nice sketch! :)

Yes, good idea not to hide the "Undo/Redo" buttons on the desktop. I'll probably change that.

With the Color button, yeah, I agree, the "minus-like" looking issue is there. The reason to have a "minus" is that the button does two things: shows the currently selected color AND the currently selected stroke size. But maybe on the desktop, I need to also expand this button into two separate ones... I need to try

But thanks for the ideas and especially for the sketch! :D

@atheist @hitko Hey, folks, I think we’re talking about different things

Your comments are valid, but it is not about me “trying to get rid of all collisions in a hash table" or me "thinking of a hash table too much like an array". That's why there is a hash function, collision resolution section, and linked lists described in the Sketches.

The comment I addressed with my a bit artificial example (with XXX-XX-XX numbers) was: “I'd suggest revisiting the hash table at least” because "even a perfect hash function will result in collisions” and thus using the phrase “what if a hash function isn't perfect" in Sketches is invalid.

What forbids us from coming up with artificial constraints (i.e. limited set of keys XXXXXXX, enough memory cells. etc.), where we can actually say that 0% collision is possible? It is rare, maybe it is not a real-life example. However, the wording of “what if a hash function isn't perfect" looks not that invalid in this case, doesn't it?

* - ah, yes, about the collison comment above, I meant "concatenated" to a number 4565623 (not a string) and than mapping this number to the address 1:1

@coolguy101 Cool! I hope this app could help with that! :)

@atheist yeah, the performance is a big issue, for now, it is pretty slow (on the phone especially) and there are some issues with zooming/panning. I plan to investigate and optimize it.

About the perfect hash function... This is a matter of wording, isn't it? I mean... under certain restrictions, depending on the task we're solving the function may be perfect, right? Let's say I want to store the Names of the users for every telephone number in my city. With these restrictions we may say, the telephone number has the format of XXX-XX-XX (i.e. 456-56-23). The hash function might be as simple/stupid as concatenation of all digits into string "XXXXXXX" (i.e. it may produce the "4565623" as a key). Let's say we have 10.000.000 slots of memory (each may contain the Name of the person, if we say the max Name length is 100). This way we will not have any collision, right? And the function is perfect in this case I guess

@netikras yeah, you're right, the performance is a big issue there. The task to do something with it is in our roadmap.

@green-portal nice :D

@irene the steering wheel limits are configurable in code, so technically it is possible to make the car more maneuverable, yes.

@iSwimInTheC wow, 0.1 score!

@asgs Cool! I'm glad that it was easy to grasp

Thanks for the good advice, folks! I'll try to find time to extend the article with more examples that you've mentioned

@c3r38r170 ah, I guess I've misunderstood :D I'm glad you liked the article! And by the way, the "anti-intuitive" part I really liked so much that I couldn't resist adding this phrase to the article. It is a bit more fun now :)

@c3r38r170 haha, sorry to hear that it was anti-intuitive. I hope you didn't waste too much time on it, since the article is pretty short :)

@codecrow yes you're right the app searches specifically for the prefix https:// and if it doesn't see it it won't recognize the link. Also, the dataset that I used for the training consisted only of the images of just one book, with white background and dark text, so the model is very immature now (thus there is an "Alpha" prefix on the app logo and the link to the Issues in the app, just to show that this is just a first experimental launch). But thanks for checking it anyways!

@bittersweet thanks! Yeah, Google Lens is way to go. The Links Detector app is indeed in alpha stage and has many issues so far. The one thing that this app gave me that Google Lens couldn't is all the fun creating the model and UI for it :D

@theabbie I've actually seen many links in modern books (about ML for example)

@coffeeholic it should be Safari on iOS or Chrome on Android. Also if it is opened in a WebView there is also might be some issues.