Masato Takano

ソフトウェア業界24年目
受託開発会社(5年)→外資系日本法人(13年)→フリーランス(5年)→メディア系IT子会社(1年)
Java/C#/C++/Python/GO/JavaScript/他
AIに駆逐される恐怖に震える日々

[Masato Takano @norasuzu]

動画の画質差異を提示しようかと思ったけど、エロ動画なのでお見せできません

[Masato Takano @norasuzu]

ローカルでのMiniMax H3検証の続き。iGPUはエロ動画生成の夢を見るか？
ClipProjを入れてテキストエンコーダ分の使用メモリを10GB近く削減した状態で、SageAttentionを有効にして再検証

[結果]
20分→12分ちょい (768p、4step LoRA)

あれ？これなら実用範囲じゃね…？
メモリに余裕があり別作業もOK

[Masato Takano @norasuzu]

勿論、トレードオフはあるわけで全体的に動画が荒くはなっている
x2アップスケールを入れてみるか

[Masato Takano @norasuzu]

ローカル用のMiniMax H3用ワークフローを改善して、24分弱→20分前後に短縮
iGPU環境だとメモリ消費量が厳しすぎるのでClipProjを導入してテキストエンコーダのメモリ消費を抑えるのが一番効いた
SageAttentionはComfyKitchenと大差が無かったが、明日もう少し検証しよう

[Masato Takano @norasuzu]

「オルバンスおう戦はイベント戦闘扱いなので、キャラロストにはならない」のチェック

オルバンスおう戦→タイトル画面でゾンビ化→ふっかつ→PT再編成まで確認

あとは通常時のリング・アミュレット・アップグレードでのゾンビ化確認をしないと…

[Masato Takano @norasuzu]

イニシエダンジョンGodot勝手移植
ボスキャラの挙動実装・チェック中…
「オルバンスおう (本気)」を試してたら一部の武器スキルがおかしな事に気付いた
ジャンプアタックかスピニングチャージのはずだが、どっちだろう？
アトラス画像の切り出しがおかしい？

全チェックし直すか…アヘェ

[Masato Takano @norasuzu]

1年前のGPUレンタル価格よりも今のほうが高いという悲しみ
VastAIで4090環境をOn-demandで借りようと思ったら、軒並み$0.42/hを越えててびっくりする
5090だと$5/h～

インフレ・AI需要のどちらもあるんだろう

[Masato Takano @norasuzu]

VastAI、GPUの世代交代が起こってるのかRTX4090が$0.3/h未満のものが多くなり、5090も$0.3～0.36/hぐらいが多くなってきた
半年ぐらい前は4090は$0.3～0.33/hが相場だったので、全体的に5～10%ぐらい下がってる感じ
ただ、曜日・時間帯にもよるとは思う

[Masato Takano @norasuzu]

※ただしユーザはLLM

[Tibo @thsottiaux]

2026 is the year of linux desktop

[Masato Takano @norasuzu]

大変に興味深い
・思考レベル(variant)はコード品質に影響しない
・Terra(max)がAstraに匹敵する？
・一般的なコーディング用途ではTerra(medium)で十分

[Morgan @morganlinton]

Okay, woke up this morning and my Sol benchmark was finished, so I have now completed my full run, on the five core OpenAI models available today, across all effort levels.

Through this process as I shared a bit last week, I also decided to create a third eval suite, focused on routine engineering tasks.

This current suite, which I'm now calling VulcanBench Frontier, is really, most likely, harder tasks that regular engineers on engineering teams are giving models on a normal day.

What I'm testing with this eval suite is how these models do with hard stuff, things you might give a model, but not likely on a daily basis. Which means I would look at these results as providing signal for what model/effort to use for your hardest tasks.

Also as a quick reminder, or heads up for those just joining me here. With this v4 eval suite I also included a code quality scoring system that uses both Muse and Grok to score the code quality and include this as 33% of the score.

I think code quality output for models is going to matter more and more over time, not because models need to write good code for humans, but because models need to write code other agents can understand and work on too.

As for key insights I got from this, here's three:

1. Increasing effort level does not seem to impact code quality. So anyone thinking Low effort writes crappy code and Max writes beautiful code, that doesn't seem to be the case. Low and Max in pretty much every OpenAI model writes the same quality code.

2. Terra Max matches Astra at half the price. I'll just leave that here 👀

3. You never need to use GPT 5.5, and if you are, in any workflows, stop, you're wasting money.

Comparison model cards below, and if you want to see detailed reports on each specific model run, you can find those here:
https://t.co/zM6xmollED

[Masato Takano @norasuzu]

[忘備録]
・(エロ)動画生成用の参照画像は、[Anima Element 7]の[Nexus Prism 4]を使って生成。陰影が綺麗＋調整が出来るため
https://t.co/MYDQXEUvgo

・非エロの場合はgpt-imageを使うのが楽

・VastAIでのGPUは4090(24GB)がベター
　5090は割高…のはずなんだがたまに逆転する。謎。

[Masato Takano @norasuzu]

プレビューとwebmエンコードを外した状態で22分30秒
MiniMax H3をiGPUで遊ぶのは絶望的なのでおとなしくクラウドGPUを使おう
VastAIで4090(24GB)を3時間使い倒しても$1未満

iGPUでの動画生成は寝る時に放置なら使えるかなぁ

[Masato Takano @norasuzu]

DaSiWaV2のワークフロー＋minimax_h3_fl2va_bf16_turbo_multistepで4stepを試した結果、24分弱
プレビューとかwebmエンコードを削ればもう少し速くなりそうではある
780M＋64GB環境の限界を感じる…こんなことならEVO-X2買っておけばよかったと思ってしまう

[Masato Takano @norasuzu]

日頃はVastAIで試しているMiniMax H3を、無謀にもローカル環境で試してみた
7840HS(780M)、64GBメモリ、ComfyUI0.36、ROCm10.0

4step LoRA適用状態でウェイトのロードから5秒動画の出力完了までに28分
実用的…とは言い難いが一世代前のiGPUでもなんとかなるもんだ

[Masato Takano @norasuzu]

Jevで既存画像に対して高速Danbooruタグ付けできんかな
既にあるタガーで十分過ぎるとは思うが

[Masato Takano @norasuzu]

もういい加減WordPress滅びてくれないかな…(WordPress保守案件を見つつ)

[pwn.ai @pwn_ai]

So... we decided to hack WordPress Core, AGAIN! 🔥 ⛓️ Click2Shell is a one-click unauthenticated remote command execution chain (Preauth RCE) affecting every WordPress website. Wordpress rolled out a fix yesterday! The story about how one preview link made WordPress click Install, load an inactive theme's PHP, and hand us RCE is below. Happy Friday!

https://t.co/91D2F2lSdb

[Masato Takano @norasuzu]

やはり1.58bit量子化は性能低下が厳しいものがある模様
はい撤収

[Benjamin Marie @bnjmn_marie]

Here is the answer I was looking for
Buried in the report

Bonsai 2 27B's accuracy on terminal bench and swe bench drops by nearly 20 points.

It's all I needed to know and what they should have published from the beginning in their main benchmark table.

[Masato Takano @norasuzu]

いまいち詳細がわからんが、BitNet(1.58Bit量子化)の亜種みたいなものか？
これも週末に試してみるか…と思ったが、PrismML専用のllama.cppが必要なのか
めんどくさ

[PrismML @PrismML]

Today, we’re announcing Ternary Bonsai 2 27B.

Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance.

Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use.

Ternary Bonsai 2 27B is available today under Apache 2.0.

[Masato Takano @norasuzu]

DBや外部サービスへのリクエストが処理時間において支配的なら、バックエンドの言語なんてなんでもいいわな
でもJavaScript・Python・PHPは滅びろ
静的型付け言語以外は許さん

[Alexander Belanger @abelanger5]

We've had a number of Go vs Rust debates over the years on our team. We think a lot about performance since we're building an infra service, and the performance question basically comes down to: where does your software spend the most time?

It's classic Amdahl's law: for our case, 90% of our time is spent in the database layer. Even if we got a 2x improvement on the 10% of the time we spend in the actual application runtime, which is written in Go, we're still only improve a little over 5%. Our total performance improvements from an expensive rewrite would be bounded at a little over +10%.

On the other hand, if we throw all of our resources into optimizing the database layer, we could get massive improvements, up to 9x! It would even make more sense to throw tons of resources at something like pg_rust from @mmalisper -- which we're watching closely -- over rewriting our Go app in Rust. We're obsessive about the database layer for this very reason, and less obsessive about the language.

This type of 90/10 split, where 90% of the time is spent on the durability layer, is the sign of the healthy application which needs no rewrite.

[Masato Takano @norasuzu]

ComfyUI 0.36.0でデフォルト状態でMiniMax H3関連がかなり高速になったそうで、Ryzen AI 370(iGPU:890M)でも使えなくもない速度になったそうな
週末にローカル環境(7840HS、iGPU:780M)で試すか…5秒に15分ぐらい掛かりそうで怖い

[Roberto Maurizzi @RobertoMaurizzi]

@sep_is_heim On a Ryzen AI 370, H3 went from 45 minutes for 5 seconds at 0.3 megapixel to 12 minutes for 5 seconds at 0.5 megapixel... 😱😱😱

[Masato Takano @norasuzu]

Union Alphaは不安定すぎて話にならない
評価しようにも使えなければどうにもならない