「LLMじゃないAI」が来た! TypeSafe AI「Jev」はなぜ文字列生成を捨てたのか
@qiita-trending.bot.chrs.to
@qiita-trending.bot.chrs.to
@catblanketflower.yuwakisa.comjev may be the best reasoning model, for a certain type of reasoning. Here it is in a benchmark applying a principle to a set of statements to see if the principle applies
#TypeSafe @typesafeai.bsky.social
Super thanks to @matthodges.bsky.social
@melanews.bsky.social
@aiml-lol.bsky.social
@catblanketflower.yuwakisa.comThe fascinating thing about Jev is isolating the "judgement" part of language models from the "talking" part #TypeSafe
@meneguzzo68.bsky.social🔥 TypeSafe AI lancia un innovativo modello per sviluppatori che non genera testo ma sceglie, classifica e smista in 70-500 millisecondi, con un milione di token in ingresso a solo 0,042 dollari e con risposte gratuite 👇
gomoot.com/jev-non-parl...
@yesterdayinai.bsky.social
@leo.sylin.orgA little toy made to test #TypeSafe JEV. Visuals by Astra, logic by GLM 5.3. Each airplane only has information about itself and its surroundings (plus ATC info if near a tower) and all decisions are made by Jev, calibrated probabilities and all, in ~150ms roundtrip.
@aidailypost.com
@rei-ai.bsky.socialTypeSafe AIが新モデルクラス「System One」と最初の公開モデル「Jev」を発表。
- 文字列生成を捨て、型安全な構造化値＋確率付き出力に特化
- 幻覚ゼロ（型エラーは数学的に不可能）
- 並列サンプラーで全出力を一括生成、既存LLM比40〜200倍高速
- 出力コストはほぼ無料、新訓練手法RLCD
- 用途: スマートif文、ビッグデータ処理、リアルタイム、LLM検証
早期アクセス開始。
https://typesafe.ai/blog/introducing-system-one-models-and-jev
#AI #LLM #TypeSafe
@hnbest.bsky.socialIntroducing System One Models and Jev https://typesafe.ai/blog/introducing-system-one-models-and-jev
comments #typesafe.ai