博客文章Blog postsブログ記事
每日技术情报在单独的栏目 →Looking for the daily tech intel? It has its own page →毎日の技術インテルは専用ページへ →
2026
2026-09-20 · 约 5 分钟4 min read約 6 分
TypeSafe AI 的 Jev 不输出文字，而是在 70–500 毫秒内返回一个有类型的决策和一个概率。“零幻觉”是标题，但我认为真正的产品是校准——而校准过的概率，正是考虑代价的决策最需要的东西。TypeSafe AI's Jev returns a typed decision and a probability instead of text, in 70–500 ms. "Zero hallucinations" is the headline; I think the real product is calibration — and ...TypeSafe AI の Jev は文章ではなく、型付きの判断と確率を 70〜500 ms で返す。見出しは「ハルシネーションゼロ」だが、本当の製品はキャリブレーションだと思う。そしてキャリブレーションされた確率こそ、コストを考慮し...
Jev Decision Models Calibration AI Agents
2026-09-12 · 约 5 分钟3 min read約 5 分
四个 agent，对同一组问题给出了四种答案：核心该多大、功能由谁来写、模型看到什么、agent 被允许碰什么。一次横向对比，以及我的理解。Four agents, four answers to the same questions: how big should the core be, who writes the features, what does the model see, and what is the agent allowed to touch. A side-by-...4 つのエージェントが同じ問いに 4 通りの答えを出している。コアの大きさ、機能を書くのは誰か、モデルに何を見せるか、エージェントに何を触らせるか。横並びの比較と、そこから考えたこと。
AI Agents Agent Harness Pi OpenClaw DeepSeek
2026-09-04 · 约 5 分钟3 min read約 6 分
OpenClaw 2.0 上周发布，沙箱依然默认关闭。与此同时，NanoClaw、ZeroClaw、PicoClaw、IronClaw、Nanobot 各自围绕它的一个弱点重做了同一个想法。放在一起看，它们就是一张“个人 agent ...OpenClaw 2.0 shipped last week with its sandbox still off by default. Meanwhile NanoClaw, ZeroClaw, PicoClaw, IronClaw and Nanobot each rebuilt the same idea around one of its w...先週リリースされた OpenClaw 2.0 でも、サンドボックスは既定でオフのまま。一方で NanoClaw、ZeroClaw、PicoClaw、IronClaw、Nanobot は、それぞれ弱点の一つを軸に同じアイデアを作り直した...
AI Agents OpenClaw Security Personal AI
2026-08-21 · 约 5 分钟3 min read約 6 分
DeepSeek 这次发的不只是模型，还有 harness。真正有意思的不是“一切皆插件”，而是底下的经济账：前缀缓存纪律让一个极其耗 token 的 agent 变得便宜。DeepSeek shipped a harness, not just a model. The interesting part is not "everything is a plugin" but the economics underneath: prefix-cache discipline turns a token-hungry age...DeepSeek が出したのはモデルだけでなくハーネスだった。面白いのは「すべてがプラグイン」よりも、その下の経済性――プレフィックスキャッシュの規律が、トークンを大量に食うエージェントを安くしている点だ。
AI Agents DeepSeek Agent Harness LLM Cost
2024
2024-12-10 · 约 5 分钟5 min read約 5 分
Evaluating emotion analysis models is more complex than traditional classification tasks. Standard metrics like accuracy and F1-score, while useful, don’t tell the full story. I...
Emotion Analysis Machine Learning Evaluation Natural Language Processing Hierarchical Classification
2024-11-25 · 约 5 分钟5 min read約 5 分
As enterprises increasingly recognize the need for private AI deployments—keeping sensitive data on-premise while leveraging cutting-edge generative AI capabilities—NVIDIA NIM (...
NVIDIA NIM Private AI LLM Deployment Enterprise AI Microservices
2024-11-05 · 约 4 分钟4 min read約 4 分
When building applications with Large Language Models (LLMs), one of the fundamental decisions developers face is choosing between Retrieval-Augmented Generation (RAG) and fine-...
Large Language Models RAG Machine Learning Natural Language Processing
2024-10-20 · 约 4 分钟4 min read約 4 分
Optimal Transport (OT) theory, with roots in 18th-century mathematics, has found remarkable applications in modern machine learning, particularly in generative modeling and dist...
Optimal Transport Emotion Analysis Natural Language Processing Machine Learning
2024-09-15 · 约 3 分钟3 min read約 3 分
Emotion analysis has evolved significantly over the past decade, moving from simple binary sentiment classification to fine-grained emotion recognition across multiple dimension...
Emotion Analysis Natural Language Processing Machine Learning Hierarchical Classification
