Pinned Tweet
長文を読むために、モデルの中身を組み替える。
Qwen3.8-27Bをベースに、64層のうち16層をGated DeltaNetへ変換。日本語・英語のWikipediaを使い、層ごとの調整とLoRA蒸留を重ねて、Longbook27-v2を作りました。
huggingface.co/summerAIO/Qwe…
OpenAIでDotsとChatGPTに携わるsamirさん(@_samirism)を紹介!以前はSnapに在籍。新機能の情報だけでなく、ユーザーに率直な感想も聞いている方です。Dotsがどう進化していくか追いたい人はぜひチェック!!
Meet @_samirism: working on Dots & ChatGPT at OpenAI, previously at Snap. Samir shares product updates and asks users for candid feedback. Worth following if you want to keep up with Dots and how it is evolving!
1
253
summer retweeted
soon, your dot will have its own email address.
this means you can just cc it and say “find a time that works for everyone” or forward it to a receipt and ask “help me negotiate a refund for this order.” no need to leave your inbox.
we’re starting testing now
excited to get it into more hands soon.
📊 Step-count benchmark|ステップ数を増やすとどうなる?
・SUN 24→40 steps: CER 1.5%(元モデルと同等)but speed 2.67×→2.38×。クローンCERの改善も消えるので 24 steps 推奨
・LUNA 4→8 NFE: MOS 3.128→3.338, CER 2.0%→1.7%, clone MOS 3.111 🔥 still 5.06× faster than original
Measured on Grok Bot VM (CPU-only, 4 threads) / Grok BotのVMで計測
huggingface.co/summerAIO/Iro…
teapawt.org/i/status/2122383512917…
🌞🌙 Irodori-TTS-v4.1-Small を魔改造しました — modded entirely on a Grok Bot VM (CPU-only, 8 cores)
Weight-space programming(task vector移植・マージ・bf16化・weight_norm folding・RoPE fusion)で2モデルを作成👇
☀️ SUN — quality-safe
RTF 2.96 → 1.11(2.7x faster)
MOS 3.431 → 3.452、clone CER 6.2% → 4.9%
huggingface.co/summerAIO/Iro…
🌙 LUNA — turbo(MeanFlow 4 NFE)
RTF 2.96 → 0.60(4.9x faster, faster than real-time on CPU)
DiT only: ~21x faster
clone CER 6.2% → 4.1%、speaker sim 0.937 → 0.941
huggingface.co/summerAIO/Iro…
Benchmark: 40 plain + 10 voice-clone sentences, kotoba-whisper CER / UTMOS22 / WavLM sim
MIT license. Credits: @Aratako_LM Irodori-TTS, takuma104 Yomi
#IrodoriTTS #TTS #GrokBot
3
356
🌞🌙 Irodori-TTS-v4.1-Small を魔改造しました — modded entirely on a Grok Bot VM (CPU-only, 8 cores)
Weight-space programming(task vector移植・マージ・bf16化・weight_norm folding・RoPE fusion)で2モデルを作成👇
☀️ SUN — quality-safe
RTF 2.96 → 1.11(2.7x faster)
MOS 3.431 → 3.452、clone CER 6.2% → 4.9%
huggingface.co/summerAIO/Iro…
🌙 LUNA — turbo(MeanFlow 4 NFE)
RTF 2.96 → 0.60(4.9x faster, faster than real-time on CPU)
DiT only: ~21x faster
clone CER 6.2% → 4.1%、speaker sim 0.937 → 0.941
huggingface.co/summerAIO/Iro…
Benchmark: 40 plain + 10 voice-clone sentences, kotoba-whisper CER / UTMOS22 / WavLM sim
MIT license. Credits: @Aratako_LM Irodori-TTS, takuma104 Yomi
#IrodoriTTS #TTS #GrokBot
一回消したの見えてんだぜ👀
We’re expanding free access to our frontier models for up to 200,000 scientists worldwide, with training and support.
Applications open Nov. 23. By default, we don’t use data shared through the program to train our models.
openai.com/index/expanding-c…
241
We’re expanding free access to our frontier models for up to 200,000 scientists worldwide, with training and support.
Applications open Nov. 23. By default, we don’t use data shared through the program to train our models.
openai.com/index/expanding-c…
summer retweeted
Today we're launching Perplexity Frontier Research.
The goal of Perplexity Frontier Research is to create a system that can perform autonomous research.
The work will focus on multi-agent orchestration, reinforcement learning, continual learning, systems, and pre-training.
We will develop algorithms, models, and evaluations, released openly, and tested in Perplexity products. Our work will be supported by more than 10,000 GPUs.
The Frontier Research team will regularly publish papers, models, data, and benchmarks.
We're actively hiring researchers and engineers for Perplexity Frontier Research in San Francisco, Palo Alto, New York, and London.
research.perplexity.ai
Astraーーーー!when
We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards.
These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems.
openai.com/index/our-first-p…
6
529
We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards.
These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems.
openai.com/index/our-first-p…
EmbeddingGemma 2、文章だけじゃなく画像・音声・動画も同じベクトル空間に置けるのか。答えを喋るモデルではなく「似たものを探す」側のAI。
公式カードでは不要な画像・音声エンコーダを読み込まない構成にもできる。全部入りを常に載せなくていいの、ローカル用途にはうれしい設計。
ゲーム素材を「雨の夜の街」みたいな言葉で横断検索する仕組みに使えたら楽しそう!!
huggingface.co/google/embedd…
これ面白そう!!d1.2は画像の中の位置や、文章の抜き出したい部分まで返せるのか。長文で説明してもらうより、そのままアプリで使える判断や抽出結果が欲しい場面って多い。ゲーム画面から必要な情報を拾う用途にも試してみたい。
Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities.
> #1 on the @ArtificialAnlys Decision Index (75)
> Read a value from the input (text, number or date)
> Find every passage of a kind in a text (spans)
> Locate things in a picture (box for one, boxes for every instance)
> Estimate a number (number with a range)
> Pick every option that applies (choice)
1/ 🧵
1
247
d1.2-ultra ranks first among decision models on the @ArtificialAnalysis Decision Index v2.1, and for less than half its price, d1.2 offers the best trade-off between quality and cost. Both models are available today on the Liquid API, at $0.02 for d1.2 and $0.042 per million input tokens for d1.2-ultra.
/2
Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities.
> #1 on the @ArtificialAnlys Decision Index (75)
> Read a value from the input (text, number or date)
> Find every passage of a kind in a text (spans)
> Locate things in a picture (box for one, boxes for every instance)
> Estimate a number (number with a range)
> Pick every option that applies (choice)
1/ 🧵
LFM2の構造を見直すと、「Attentionを全部なくしたモデル」ではないんだよね。350M版は公式カードで16層のうち短い畳み込みが10、Attentionが6。ゲート付き畳み込みとAttentionを組み合わせるハイブリッド。全層を同じ仕組みにする以外の作り方があるのが面白い!!小型だから何でも任せるというより、公式も用途を絞った調整を勧めている。モデルのサイズだけじゃなく、中の配分を見たい。huggingface.co/LiquidAI/LFM2…
18
598










