Mark (@MarcosFelipeRa7) header
Mark (@MarcosFelipeRa7) profile picture

Solo un fan de muchas cosas || BNHA || GINTAMA || BLACK CLOVER || JUJUSU || Y vamos a ver que mas 🥸 ...

Joined November 2021
Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)Photo by Mark (@MarcosFelipeRa7)
Take the easiest anime to slander (MHA) and the worst slanderposters of all time (One Piece fandom) and it apparently creates a slander so bad that it loops back around into being the best MHA glaze I've ever seen. Like yes actually Deku got that shit on he might be goated
Official art btw 💔 This yo goat?😭
30
777
7
13,421
127,320
Replying to @makisucks0
Yes, this is my goat why you ask?
3
10
2,629
29,189
友達にスタレ布教することが出来て今これ
8
5
356
4,681
Mark retweeted
HE GOT HIMSELF A GF BY JUST CHECKING IF IT WAS STILL RAINING 😭
22
105
6
3,508
112,362
So adorable
3
15
546
7,094
Alright i’m invested
11
27
1
743
11,086
Mark retweeted
AAAAAH
4
81
734
Mark retweeted
La IA en Desarrollo de Hoyoverse en las betas NO SON PARA JUGADORES Lo están desarrollando principalmente para solucionar el problema de las filtraciones y denunciar dichos problemas Así denunciar y llevar a juicio a los leakers es costoso y molesto
miHoYo ha entrenado un AI que juega mirando la pantalla y usando teclado/ratón como un humano. Basado en Qwen3-VL, con miles de horas de gameplay + razonamiento + RL, llega a ~50-80% de éxito en quests de Honkai: Star Rail. El paper “Gamer” detalla la arquitectura generalista VLA para juegos 2D/3D en tiempo real. ¿Es el futuro del testing automatizado o de compañeros de juego?
3
5
44
2,345
Mark retweeted
Deku was genuinely trying to break the fourth wall
4
174
16
2,368
94,879
Mark retweeted
アキヴィリ、こっちこっち! #HonkaiStarRail
1
247
4
2,250
14,704
Mark retweeted
Look at them 🥹
1
3
122
1,588
(草图
1
153
1
2,458
8,762
Did this with Triple Threat as the full team because it just felt right to draw the trio in this #THSC #HonkaiStarRail
Silly crossover art to celebrate or smth #THSC #HonkaiStarRail
12
271
3
1,699
9,570
210
3
2,150
10,110
Mark retweeted
This could be good for game testing i don't see any use for public use
Hoyoverse is developing an AI that can play games on its own. #HonkaiStarRail
1
1
6
255
☆〈English Translation〉 Never underestimate someone with 200 speed
🎭キャラpv
22
342
3,196
Mark retweeted
Oh it's indeed for Game Testing
MiHoYo/Hoyoverse AI Information–––– Apparently, miHoYo, the developer of “Genshin Impact,” has developed an AI that can “look at the screen” and “use a keyboard and mouse” to play games just like a human. For example, in a side quest in “Honkai: Star Rail,” it completed more than 80% of the objectives. The research team says that “having AI play new content lets us check whether quests and other features progress properly.” The recipe for developing this AI system is interesting, too. ① The development was based on Alibaba’s publicly available Qwen3-VL (8B) model ※In its unmodified state, the quest completion rate was 0% ② Taught it how to move the mouse and press keys using more than 10,000 hours of gameplay recordings across a wide range of games →The completion rate rose to 34% ③ Added explanations of “what it’s trying to accomplish right now” to the gameplay recordings and had it learn from them →It rose to 72% ④ Finally, had it play on its own and reinforced successful actions →The number of successful attempts at the task of finding and talking to Pom-Pom (a character) increased from 3 out of 16 to 15 As a result, it reportedly achieved a success rate of around 50% across seven minigames, surpassing GPT-6.1 Sol’s 17%. The developers are hopeful about its potential applications in automated game testing. Still, as players, we should also keep an eye on the possibility that AI capable of playing games alongside us will become even more advanced. #HoyoVerse #miHoYo #GenshinImpact #HonkaiStarRail #ZenlessZoneZero
1
1
21
2,151
"Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent" is a joint study presented by the "miHoYo Honkai AI R&D Team." The study includes contributors affiliated with the National University of Singapore, Tsinghua University, the University of Chinese Academy of Sciences, the University of Science and Technology of China, and Xi'an Jiaotong University. One of its supervisors is Jiang Dawei (David Jiang), the producer of Honkai Impact 3rd, Honkai: Star Rail, and Honkai: Nexus Anima. A number of contributors are miHoYo interns, making up 5/8 of the core contributors and 4/15 of the regular contributors. The miHoYo Honkai AI R&D Team developed "Gamer," an AI system designed to perform tasks in complicated environments, such as playing video games or interacting with digital and simulated worlds. Unlike ordinary AI systems that handle simple, one-step tasks, Gamer can perform a series of actions over an extended period. It can observe its surroundings, remember important information, make decisions, and determine what to do next. Gamer was trained using tens of thousands of hours of keyboard and mouse recordings. It was then further improved to use less computing power while still making good decisions when faced with complicated situations. Gamer was also trained to complete tasks that require multiple steps within 2D and 3D game environments. The results showed that it could perform well, adapt to different environments, and operate with efficiency comparable to that of skilled human players. It explored: — Genshin Impact, Honkai Star Rail, Zenless Zone Zero, PUBG: Battlegrounds, DOTA 2, Minecraft, Terraria; — Palworld, Stardew Valley, Monster Hunter Wilds, Hogwarts Legacy, Cyberpunk 2077; — GTA V, Assassin's Creed Syndicate, Apex Legends, DOOM Eternal, PEAK, Ori and the Will of the Wisps; — Fortnite, Counter Strike 2, and more. Gamer plays games much like a human does. It only looks at what appears on the screen and uses keyboard and mouse inputs to control the game. It doesn't directly access the game's internal data or modify its underlying programming to gain an advantage. The team also developed a system called "Streaming-vLLM," which helps Gamer learn and operate more efficiently. It allows the AI to interact with games in real time, reduces delays, and makes the training process faster and more reliable. However, the researchers use a separate system to check whether Gamer has successfully completed a task. This system provides feedback or rewards that help Gamer learn and improve. In conclusion, the study shows that even ordinary keyboard and mouse recordings, when collected in large amounts, can be valuable for training advanced AI systems. By learning from these interactions, AI can gradually develop more complex skills and become better at understanding and responding to different environments. Ultimately, the goal is to develop AI that can interact intelligently not only with digital environments, such as video games and computer applications, but also with physical environments, such as robots operating in the real world. Access the study: arxiv.org/abs/2611.18397
今後は、AIに新作のゲームを先に遊ばせるようになるかもしれません。 「原神」開発元miHoYoが、"人間と同じように「画面を見て」「キーボードとマウスで」ゲームを遊べる"AIを開発したそうです。 たとえば、「崩壊:スターレイル」のサブクエストで、途中の課題の8割以上を達成しています。 研究チームは、「新しいコンテンツをAIに遊ばせることで、クエストなどがちゃんと進行するかを確かめられる」としています。 なお、このAIシステムを開発したレシピが面白いです。 ①開発のベースはAlibabaの公開モデルQwen3-VL(8B) ※素の状態ではクエストの達成率は0%でした ②幅広いゲームの1万時間を超えるプレイ記録でキーとマウスの動かし方を覚えさせる →達成率34%になりました ③プレイ記録に「いま何を目指しているか」の説明を付けて学ばせる →72%まで伸びました ④最後に自分で遊ばせてうまくいった操作を強化する →パム(というキャラ)を探して話しかける課題の成功が16回中3回から15回に増えました この結果、ミニゲーム7種でも成功率は約50%で、GPT-6.1 Solの17%を上回ったそうです。 開発元は「ゲームの自動テストへの応用」に期待を寄せていますが、プレイヤーとしては一緒にゲームを遊べるAIがさらに高度化する可能性にも目を向けたいところです。
5
14
17
100
16,910
Mark retweeted
pay to win about a medal
They saw everyone had fun last patch with this gamemode and was do-able with Trial characters too and Hoyo was like: nope, time to make it p2w! Why do they keep doing this, we just wanna have fun. Thanks for punishing us for skipping ig.
11
10
451
10,733