
he/him | [email protected] | views are my own | building experiences at @FarFlashGroup & building socials at @buildcanada
Greater Toronto Area
Joined February 2017
- Tweets7.6K
- Following1.6K
- Followers759
- Likes26.1K
Pinned Tweet
stunned. 10 years in the scene. every second worth it.
❤️
We're proud to present the #GOMLForever Brad Durish Community Award recipients! 🏆
Meet your winners: @bend0ll, @DopeSSBM, and @LukeFarag! ☁️
This year's winners have demonstrated exceptional leadership in the Canadian Smash scene, and we're so proud of them! ❤️
I recently went down a rabbit hole about the OSC closure and I never realized just how scandalous it all was.
I thought it was a convenient conspiracy theory but it really does feel like the OSC was sabotaged by Infrastructure Ontario - intentionally or not.
🧵
September 2024:
The Ontario gov tells CTV that the four-panel rainforest repair was still active. That permit application was later listed as cancelled.
A replay is worth a thousand stats.
Try PIVOT, a deep learning model that predicts the outcome of Melee matches, shows you near misses, alternative outcomes, player tendencies, momentum, and more.
pivot.lukefar.ag/
@iBDWSSBM @LiquidHbox @ZainNaghmi @Wizzrobe @moky_dokie @LilTrickshot @s00nsay @RapMonsterSSB @SaltSSB @jakedirado @KrudoKyle @SsbmGinger @NotAklo @n0ned @SSBMagi
Would love to have you guys check it out!
128
Is this generative AI?
No. It's a machine-learning model. Machine learning is a broad technique where computers use data to learn patterns and make predictions. Generative AI is a specific subset of machine learning focused on creating brand-new content. Additionally, the model has been trained only on publicly available Slippi files shared by tournaments.
Why do this?
In broadcasts of card games like poker, viewers who are unfamiliar with the ins and outs of the game are aided by visuals detailing the odds of each player's ability to win based on their hand. That way, even someone who has never played poker can appreciate watching a player's win odds rise from 5% to 100% as the cards on the table form a straight for them.
The idea here is similar. Melee is an extremely complex game to not just play, but also watch. Stocks and percentages are huge helps, but so much more goes into understanding a lead such as advantage state, character weights, matchups, stage picks, even a player's "momentum". The methodology is different, but the end result is supposed to be the same. Create a way to understand what's on screen, with or without a full understanding of Melee's intricacies.
Does it work?
Yes! PIVOT is trained on 2.5M publicly available tournament sets to create a 210k parameter model you could run on an old Android phone. It accurately predicts the winner with 75.5% accuracy by about one minute into a game using gameplay alone, without knowing player identities or rankings.
Even on modest hardware, within 0.1 ms of a frame state being passed to the model, an odds calculation can be formed. Currently, we can predict the result of a game with ~75% accuracy in about 60 seconds without any consideration of who the players are.
Is the model itself available?
Yes! huggingface.co/lukefarag/piv…
In the near future, I plan to release tools so you can more easily add this to your personal or tournament stream or integrate PIVOT in prepackaged ways into other Melee tools.
1
1
174
When smart people make their great work publicly available, it invites more people to build. 🙌
Credits for work we built off of:
- Fizzi and the Slippi team (duh)
- Slippi Lab: Frank Borden
- Slippi Viewer: Graham Preston
- libmelee: AltF4 and contributors
- Public replay dataset: AltF4, Nikki, and yashichi
- Frame-data extraction: pfirsich
- meleedb: BroccoliRaab
- Training Mode references: UnclePunch
- 20XX documentation: Achilles1515 and contributors
- Tournament replay archive: Nikki
- Replay dataset mirrors: Erick Martinez (erickfm)
- Ranked replay dataset: JoeyDonuts
- Lunar replay archive: madenney / Lunar Melee
- slpz decompression: Alexander Harrison
- GameCube controller artwork: Alphathon and Frank Razenberg (zzattack)
- slippi-visualiser: schmooblidon
- react-slp-viewer: vinceau
- HSDraw / HSDlib: Ploaj
116







