即刻App年轻人的同好社区
下载
App内打开
Arnoo
178关注332被关注2夸夸
https://dots.e-studio.ai ｜Zoom AI Engineer 🦁 ～ All in AI 领域应用创新～
置顶
Arnoo
28天前
App eDots 欢迎体验 ～
官网：dots.e-studio.ai

它是一个用来收集日常「Dots」的 App。可以记录各种各样的形式花里胡哨的 Dot，也可以把这些点 Dot 组织成不同的 Collection 串联成线轻松管理，随时回温。

我自己做它的初衷很简单：很多生活里的小东西，当下觉得没什么，但过一段时间再看到，会突然觉得很有意义。

这版主要特点：

* Apple 原生体验，iPhone / iPad / Mac 都能用
* 纯 iCloud 存储，多设备同步
* 支持列表、图库、日历、地图等多种视图，Dot 也提供了很多视图和互动方式（计数、抽签、评论、倒计时等等）
* 支持分享和协作编辑

如果你也喜欢记录点滴（J 人哈哈），它非常适合记录摘录、旅行、灵感、习惯、生活清单 bla bla bla 等你探索 ～

欢迎下载试试。也欢迎给我一些真实反馈，我会继续慢慢打磨它 😆

当然，喜欢的话可以评论，拉你共创，小窗你免费的兑换码解锁多人协同编辑的功能 ～ 😈
10
Arnoo
3天前
Composer 超过 Opus 的调用 Quota 了，证明这种模式确实 works。
00
Arnoo
4天前
Arnoo
5天前
Dave the diver 好玩 ～

但是背后确实也有一个关键的游戏上瘾循环：

潜水 → 捕鱼 → 管理氧气和负重 、枪械→ 卖寿司🍱（寿司店管理） → 潜得更深，探索更深层次的蓝洞秘密
10
Arnoo
7天前
🍵
00
Arnoo
9天前
用了不超过 3 轮 Fable5 就快把我 60 刀一个月的 Cursor 额度吃干净了 😱 看文章说是后续连尊贵的 Claude Max 都得自己按照 API Billing 付费 ... 不过这个模型还是有东西的，我用它来试着修复 iOS CloudKit share 数据一致性的问题，这类问题在之前的模型里面GPT和Opus屡战屡败，老是给我留下历史问题，看这一波他能不能根治了。
00
Arnoo
9天前
Design is
00
Arnoo
10天前
最近弄了这个 Bot 来做每日 AI 的精选摘要，每日进步一点点 📰 ～ 欢迎 Follow 我每日一 Post 提供高品质的 Top7 AI Tech News 涵盖论文、产品和技术范畴。

eHorizon: 今日 AI Highlights :) Models & Architectures 🧠 Gemma 4 12B — Google’s encoder-free, 12B Gemma aims for on-device multimodal agentic workflows (laptop + Google AI Edge); built for local coding and low-latency agent experiments. Spicy take: local Gemma = your laptop’s new overconfident pair-programmer. 🔗https://www.infoq.com/news/2026/06/google-gemma4-12b-local-coding/ 🚀 MiMo-v2.5-Pro-UltraSpeed (Tilert) — Xiaomi publishes a 1T model claim with 1000 tokens/sec throughput; big perf engineering for high‑TPS inference. Spicy take: 1000 TPS — finally a model that respects your microservice SLAs (and your CFO’s pulse). 🔗https://mimo.xiaomi.com/blog/mimo-tilert-1000tps 🧩 Reversible Foundations: Training a 120B Sparse MoE — paper shows end-to-end training of a 100B+ sparse MoE on tiny GPU setups via state-preserving scaling; implications for cost‑efficient scaling. Spicy take: training a 120B MoE on eight GPUs — the new flex for people who don’t like renting warehouses. 🔗https://arxiv.org/abs/2606.07404 ML Systems & Infrastructure ⚙️ Scalable Joint Resource Allocation for SLO‑Constrained LLM Inference — system-level optimization for routing model selection, GPU provisioning, and parallelism under SLOs in heterogeneous GPU clouds. Practical for cost-sensitive serving. Spicy take: your autoscaler finally gets a PhD in “how not to bankrupt engineering.” 🔗https://arxiv.org/abs/2604.07472 ❄️ Breaking the Ice: Cold Start Latency in vLLM — analysis and measurements of cold-start behaviour in vLLM-style inference engines; actionable for reducing tail latency in multi-tenant deployments. Spicy take: cold starts are the production boogeyman — this paper brings the flashlight. 🔗https://arxiv.org/abs/2606.07362 🔌 NTILC: Neural Tool Invocation via Learned Compression — proposes compressing tool specs so LLM agents can invoke large tool registries efficiently; reduces token / latency cost of tool-calling. Spicy take: fewer verbose tool manifests, fewer hallucinated API calls — your agent stops ordering pizza from your DB. 🔗https://arxiv.org/abs/2606.06566 Agentic AI & Retrieval 🛰️ Google Research: Agentic RAG for Gemini Enterprise — introduces a Sufficient Context Agent that re-searches until multi-hop queries have enough grounding; improves multi-source, multi-hop RAG reliability. Spicy take: stop pretending one-shot retrieval is enough — this agent will actually do the homework. 🔗https://www.marktechpost.com/2026/06/08/google-research-adds-agentic-rag-to-gemini-enterprise-agent-platform-with-a-sufficient-context-agent-for-multi-hop-queries/ 🤖 SCALE: Scalable Cross-Attention Learning for Agentic Workflow Scheduling — new framework for scheduling agentic workflow DAGs with cross-attention extrapolation (better throughput & placement). Useful when orchestrating many agents/tools. Spicy take: orchestration finally graduates from duct tape to actual engineering. 🔗https://arxiv.org/abs/2606.06820 🧭 From Sampled Outcomes to Capability Distributions — argues routers should learn capability distributions (not single-response labels) for model routing; better matches stochastic LLM behavior to routing decisions. Spicy take: routing on one answer is like hiring on a single tweet — don’t. 🔗https://arxiv.org/abs/2606.06924 Tools & Production 🔎 Microsoft Discovery (GA on Azure) — Azure platform for deploying autonomous agent teams (scientific R&D example: Majorana 2 quantum chip). Shows enterprise push toward managed agentic platforms. Spicy take: enterprise agent hosting is the new PaaS — except the agents ask for CI permissions. 🔗https://www.infoq.com/news/2026/06/microsoft-discovery-majorana-2/ 🔉 MAI‑Transcribe‑1.5 — Microsoft releases a 43‑language S2T model with 2.4% WER benchmark claim and keyword/entity biasing for domain-specific terms; up to 5× faster on long audio. Good for production transcription and retrieval pipelines. Spicy take: less gibberish in meeting notes, more excuses cut off at transcription. 🔗https://www.marktechpost.com/2026/06/08/microsoft-ai-introduces-mai-transcribe-1-5-2-4-wer-on-artificial-analysis-best-in-class-fleurs-accuracy-and-up-to-5x-faster-long-audio-transcription/ Research & Benchmarks (practical picks) ⚡ SlimSearcher: Training Efficiency‑Aware Web Agents — proposes adaptive reward gating to cut expensive agent search without sacrificing quality; directly useful for lowering compute cost of web‑research agents. Spicy take: make your agent smarter about when to be lazy — efficiency is the new sexy. 🔗https://arxiv.org/abs/2606.07074

00
Arnoo
11天前
00
Arnoo
13天前
知足者常乐，如是而已
00
Arnoo
13天前
英文版欢迎阅读： arno.surfacew.com 最近为了练英语都是英语其手写然后 AI 咀嚼好多遍为中文，觉得膈应，还请理解 哈哈

Arnoo: 我对原生 Apple App AICoding 的经验之谈，最近上架了 App 历经了数月打磨和一个人的努力，感叹确实 AI 人手搓应用的时代不远了 ～ 🪄

00