daktshh

[Dylan O'Sullivan @DylanoA4]

Joan Didion, every day is all there is

[cage 🇵🇸 @CAGEtheGEGEG]

does mamdani know the whistlestop summit ending nisc

[daktshh @daktshh]

@n0ats fête feeling & dis ain’t right🤞🤞

[︎໊ @fentanvl]

this new pz album

[daktshh @daktshh]

@chiefdosaa_ pz on a rok beat ??? oh yeah my neighbors are gonna hate me fs

[daktshh @daktshh]

@junyawtf would you say art *should* be separated from the artist, rather than it *can* be ?

[daktshh @daktshh]

@torgeirlysen 1000% agree,, in retrospect i learned the most (both abt myself and the subject) during my MS when i took 3 PhD level courses along w my regular course

brutal but i dont regret it in the slightest

[daktshh @daktshh]

@STRAPPEDUS it’s 8 songs now ? is the delay cuz of a ft🫪🫪

[daktshh @daktshh]

@n0ats PLEASE tell me we’re getting a lancey ft tomo

[daktshh @daktshh]

@wllsykes @STRAPPEDUS luc1us n earl

[daktshh @daktshh]

@lifemocksart everytime i think he can’t get any better i hear a new snippet mannnnnn

[tender @tenderizzation]

did the deepseek v4 model architecture team design this cookie

[biased estimator @selfattentive]

It’s in torchtitan. It’s literally implemented in deepspeed. Guys why build this ourselves it is up streamed in huggingface accelerate. Guys just use megatron. We can plug this into monarch, we are using fairscale. It’s all already done in lightning. Let’s just clone torchtune. We can drop this into axolotl, should take a few days max. I don’t understand why we can’t just build on mosaicml composer. I mean nanotron is right there. We can spin this up today with Nemo.

[kat💓 @wowkatherine]

remembering every bad thing that happened to me last year

[kalomaze @kalomaze]

ignoring the masculine urge to start a robotics sidequest

[Mihaela van der Schaar @MihaelaVDS]

What if a model could adapt its compute budget on the fly?

Today, we face a rigid trade-off: either maintain costly fleets of specialist models, or rely on dynamic methods that fail to scale to large pre-trained LLMs.

We introduce Nested Subspace Networks (NSNs) to break this.

[daktshh @daktshh]

@chiefdosaa_ needed more pz on here💔💔

[Maddie D. Reese @maddiedreese]

Gemma 4 running locally on a Nintendo Switch :)

1.5 tokens per second haha, but it runs!

@googlegemma
@googleaidevs
@GoogleDeepMind

[daktshh @daktshh]

@mithernet この閾値の結果として、モデルがドメイン外の質問に対して悪影響を及ぼすのは当然ではないでしょうか？私の理解では、閾値によって結果の活性化がゼロにクリップされ（ドメイン外の質問は結果のキーと必ずしも類似しているとは限らないため）,情報が流れなくなるはずです。お時間をいただきありがとう!

[daktshh @daktshh]

@mithernet 論文発表おめでとうございます！Google翻訳を使っているので、誤解があったらご容赦ください。あなたの論文は興味深かったですが、いくつか質問があります。このアーキテクチャでは、閾値s_rがトリム変換後に勾配が流れなくなるほど高い値に最適化されないようにするにはどうすればよいのでしょうか？