Steve💙🇨🇦 retweeted
Just to put today's crash into perspective:
After today's crash, Nvidia, $NVDA, is now trading at a Forward P/E ratio of ~13x.
That's 7 turns below the S&P 500's ~20x Forward P/E and less than one-third of Walmart's ~43x.
Ironically, the name at the forefront of the AI buildout is now one of the "cheapest" stocks in the market.
Truly incredible.
I love that we're all on here now just making our own Inference engines, chasing every last token for new records, while normal people are still just using Lmstudio.
I've been using my own inference engine this week. Required little polish to be totally usable for actual work.
YASAKANIエンジンで、RTX 5090 の単一デコード 1,220 tok/s を達成!!
今回は世界最速のエンジンである証明をしました。
RTX 5090 1枚・Qwen3.8-27B・1本の生成で、公開されている最速記録を5本すべて上回った(定格・同じ土俵の条件で、最大2.5倍)。
「5090で最速」を名乗る MegaCapybara 1.39 とは、同じPC・定格・同じ問題で直接対決。動画は実際の生成を、記録した時刻どおりに再生しています。
■ 公開記録 → YASAKANI(tok/s)
- mkyt(JSON)500 → 1,220(2.44倍)
- NInfer(構造化3問)377 → 937(2.49倍)
- TensorFold(英語)298 → 424(1.42倍)
- ignis(コード)491 → 532(1.08倍)
- MegaCapybara(コード)540※ → 498(同じPC・定格では466、1.07倍)
※540はメモリ+20% OCと自前のプロンプトでの値
■ 同じPCで直接対決(MegaCapybara 1.39 Tiny、思考オフ・greedy)
- JSONL 160行:1,220 vs 635(1.9倍)
- 英語の推理小説:424 vs 253(1.7倍)
- 中→英 翻訳:613 vs 366(1.7倍)
- SQL:745 vs 603(1.2倍)
- コード編集(ファイル全文の書き戻し):1,241 vs 440(2.82倍)
- JSONL 8,192トークン:1,260 vs 666(1.89倍)
12問の平均でも、構造化1.4倍・散文1.6倍・翻訳1.6倍。
■ エージェントの仕事ほど速い
コード編集は「約4,700トークンのファイルを1か所直して、全文を返す」課題。AIエージェントが毎日やっている作業です。
3つのエンジン(YASAKANI 高速版・標準版・MegaCapybara)の出力は完全に同じ(4,711トークンが一致)で、違うのは速さだけ:1,241 / 1,136 / 440 tok/s。
■ Qwen推奨設定(思考あり・温度1.0)でも
雑談・説明・コード・数学・JSON・長文入力の6課題の中央値:YASAKANI 381、SGLang 184、vLLM 183、llama.cpp 147(10/18、同じPC、他社は課題ごとに一番速い設定)。
この条件で250 tok/sを超える公開記録は、10月中旬の調査では見つかっていません。
■ 長い文脈でも
25万トークンの文書を140秒で読み、続きの返事は0.3秒で開始。25万トークンを積んだ状態でも152 tok/s。262,144トークンの窓でVRAM 24.9GB(32GBのカードに収まる)。(10/30計測)
■ 速いだけではない
元モデルとのズレ(KLD):標準版 12.4GB で 0.021。MegaCapybara Tiny は 0.058(16.6GB)。
1,220 は高速版(11.7GB、KLD 0.044)。標準版でも 1,114。
■ 正直に書いておくこと
- ゼロから書くコードは僅差(1.04〜1.08倍)。540はOC込みの値で、定格では向こうも406〜533
- MegaCapybara以外の相手の数字は各自の公表値(別のPC・別のプロンプト・別の日)。同じPCで測ったのはMegaCapybaraだけ
- 中国語の物語は英語より遅い(176〜302 tok/s)
- 第三者の再現は公開後。それまでは「この条件で測って」と言われれば、同じPCで測って出します
■ 測り方
RTX 5090 ×1(定格600W・OCなし)、Windows、2026-11-03。NInferが公開している12問(コード3・物語3・翻訳3・構造化3)+自作のコード編集・長いJSONL。上限4,096トークン(長いJSONLは8,192)、思考オフ、greedy。速度=(出力トークン数−1)÷生成時間(読み込みは除く)。MegaCapybaraは公開版1.39を公式の1会話設定で起動し、向こうのサーバーの表示とこちらの計測は一致。
エンジンもモデル形式も自作。オープンソース・無料で公開予定。日程はこのアカウントで。
#YASAKANI
I almost regret not getting the bigger 32GB version when I could have grabbed it used for $800 CAD, but like, apparently that isn't needed.
It needs a bit more polish, as it likes to still interrupted me at times. Also, it has has a special mic connected to is, which was $90.
Steve💙🇨🇦 retweeted
Day 25/
We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today.
We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage.
And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.
If the other SWE on your team are using Claude, be a hero and start doing code reviews with Codex + Sol-6.1-high instead.
Claude is blind to its own bugs, and if your company is heavy on the Claude, it's automated testing systems are also likely going to be Claude based.
Did you know Qwen 27B can decompile console games on its own? It's that capable.
On a completely unrelated note, an IRL frend just put up a repo that compiles Playstation games into web-playable WASM binaries.
He's very smart but shy, so give his repo a Star to annoy him. 👇
The coolest part is it lets you MOD THE GAME WHILE PLAYING IT. Like Game Genie on steriods!!!
Their repo: github.com/dpchamps-vibe-cit…
My fork of their work, with a compile Playstation Pong game to play: steveseguin.github.io/playst…
Youtube explainer by them: youtu.be/6qEwwaCxalc
Steve💙🇨🇦 retweeted
Introducing Gemini 4 Carbon, it's a big jump over Argon which we released just 2 weeks ago.
Gemini 4 Carbon is SOTA on cyber, long horizon tasks, ranks #1 on Artificial Analysis V5, significantly improves on usability in domains like coding, and is more aligned.
if interested, I dump my optimizations, research, and whatever else here.
github.com/steveseguin/b70-o…
I suppose a downside of running your own engine though is it's harder to share optimizations?
There's so many model variations at this point, it often feels like I'm starting of from scratch each time.
Soon enough AI will just be able to one-shot the optimizations I'm betting.








