Odin @ohdin021

Joined January 2010

Tweets

17K
Followers

402
Following

505
Likes

24K

Odin @ohdin021

5 days ago

@DeathyDL I have good news for you

0 0 0 12 0

View Details

honestly i'd probably play Tifa, I'm totally ok with her being in Street Fighter but.. this game has been on what feels like life support for 3 years now, nothing substantial has changed despite MASSIVE majority feedback. You can argue "casuals aren't complaining", but that is a failing for professional game studios to fall for, casuals don't generally complain because they're not invested enough in the game to give substantial or meaningful feedback - this isn't always true, and often they have some good insight for the teams to carefully pick up on, but they're not focused or targeted in it. I get it, they want to capture a large casual audience and they have done that with accessibility features like modern and game balance/design choices.. but for me the game just kind of sucks, it feels stale, it bothers me that major pain points haven't been addressed in 3 years. Kind of has me hoping Harada's new game makes serious waves.

ストリートファイター / STREET FIGHTER @StreetFighterJA

4 weeks ago

#スト6 Year 4キャラクター発表‼ #FF7R よりティファ参戦!! ニューファイターとなるヤスミン、アルジュン、ゲストファイターとしてティファ、そしてワールドツアーからボシュが堂々参戦💥 2026年8月3日：ヤスミン 🗡️ 2026年秋：アルジュン 💨 2027年春前半：ティファ ☄️ 2027年春後半：ボシュ 👊

1K 35K 74K 18.8M 5K

0 0 1 89 0

View Details

Odin @ohdin021

4 weeks ago

@Bri4nF x.com/ohdin021/statu…

Odin @ohdin021

4 weeks ago

0 0 0 33 0

0 0 0 20 0

View Details

Odin @ohdin021

a month ago

Need @AleksLeVO to dub this

MΔDDiE @MiyanEatWorld

a month ago

Half of what I see about Ingrid is just this #SF6 #StreetFighter6

34 2K 14K 140K 1K

0 0 2 140 1

View Details

Odin @ohdin021

2 months ago

@dieworkwear thank you

0 0 0 8 0

View Details

Robby Soave @robbysoave

2 months ago

"At least you can eat almonds, you can't eat a data center." I was genuinely not prepared for how stupid the people on the other side of this argument would be.

112 123 3K 35K 55

View Details

Smirkley @Smirkley

2 months ago

A 5-gallon water cooler jug of water produces about $132 in economic output from data centers, but only about 2 cents from almonds.

333 582 6K 377K 962

View Details

Steve Everley @saeverley

2 months ago

This entire story has been debunked from every angle since the original article was published over a week ago, so of course People magazine is like “let’s just re-publish all the false stuff!”

People @people

2 months ago

Georgia Residents Didn't Know Why Their Water Pressure Was Low. A Data Center Was Gulping 30M Gallons people.com/georgia-reside…

28 8 11 56K 5

10 137 2K 48K 52

View Details

Brett Winton @wintonARK

2 months ago

on AI and water use. You could ask a frontier model 4 questions per day every day for a year. Or you could eat a single solitary almond.

127 340 4K 127K 394

View Details

Terminally Online Leftist 🥂 @terminallyOL

2 months ago

marcus aurelius already got this one covered for your dumbass, don't worry

Tokyo @otokyo

2 months ago

Dear atheist, What if, after you die, you find out that God is real all along? You lose.

10K 232 3K 4.1M 346

75 4K 27K 387K 3K

View Details

witcheer @witcheer

2 months ago

follow-up on my local agent stack (RTX 4060 Ti 8 GB, Qwen3.6-35B-A3B) found the hidden VRAM killer: context checkpoints. llama.cpp creates a checkpoint every 8192 tokens during prefill (~63 MiB each). at 26K context that's 5 checkpoints = 315 MiB silently eating GPU memory. on 8 GB VRAM this pushes past the 7 GB cliff. >with default checkpoints, 26K context → 14.6 tok/s >with --checkpoint-every-n-tokens -1 → 30.3 tok/s full sweep without checkpoints: - 3.5K → 34.2 tok/s - 20K → 31.5 tok/s - 26K → 30.3 tok/s - 35K → 29.6 tok/s - 50K → 27.7 tok/s smooth degradation, no cliff anywhere. the ~2 tok/s drop per 10K tokens is from attention scaling on Qwen3.6's 10/40 attention layers. I also switched from Hermes to Pi as the agent harness. ~1.9K token system prompt vs Hermes's ~13.5K. in thiscase, Pi is more suitable here. Hermes is still my main agent on my mac mini tho. final config: ``` turboquant fork, -ncmoe 30 -c 65536 -np 1 --cache-type-k q8_0 --cache-type-v turbo3 --no-cache-prompt --checkpoint-every-n-tokens -1 ```

witcheer @witcheer

2 months ago

update on mylocal agent stack (RTX 4060 Ti 8 GB, Qwen3.6-35B-A3B Q4_K_M) my initial problem was that 64K context on standard llama.cpp killed speed. V cache q4_0 pushed graph splits from 62 → 82, and Hermes decode dropped from 31 → 9-11 tok/s. unusable for real agent work.

6 8 123 23K 126

12 10 165 11K 193

View Details

stevibe @stevibe

2 months ago

Unsloth just published MTP-enabled quantized GGUFs for Qwen3.6-35B-A3B. huggingface.co/unsloth/Qwen3.…

23 53 675 48K 461

View Details

Odin @ohdin021

2 months ago

I mean I despise DI, and maybe the community has overcompensated and stigmatized DI as a scrub trap to cope with how volatile the mechanic actually is, but it is extremely uncommon at pro tournament level because those players are usually very risk averse. Funny enough it makes me wonder why daigo didn't just throw it out more considering how often he used reversals

0 0 0 15 0

View Details

AboveSpec @above_spec

2 months ago

"You need a 24 GB GPU for serious local LLMs in 2026." Everyone repeats this. It's not true anymore. Just ran a 35B-parameter model on an RTX 4060 Ti 8 GB: • 41 tok/s at 16k context • 24 tok/s at 200k context Recipe + benchmarks below 🧵

136 230 3K 284K 3K

View Details

Odin @ohdin021

2 months ago

@thefleecer @emezie I promise you nobody at the pro level is pressing random drive impacts to catch pokes, it is extremely high risk and the reason it's annoying as a mechanic is that high risk medium reward mechanics are a ranked ape's favorite meal

2 0 1 61 0

View Details

Odin @ohdin021

2 months ago

@JWonggg Mena is the best but level 2 is busted, both can be true. Blanka mental stack is notoriously difficult for even the best players to deal with hence why everyone complains about him. Daigo winning the majority of the first rounds says a lot about how powerful level 2 is.