BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.sciwork.dev//
BEGIN:VEVENT
UID:pretalx-sw26-EJEUHD@pretalx.sciwork.dev
DTSTART:20261205T090000Z
DTEND:20261205T093000Z
DESCRIPTION:Registration
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:Registration - 
URL:https://pretalx.sciwork.dev/sw26/talk/EJEUHD/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-HXXWPE@pretalx.sciwork.dev
DTSTART:20261205T093000Z
DTEND:20261205T094000Z
DESCRIPTION:Opening
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:Opening - 
URL:https://pretalx.sciwork.dev/sw26/talk/HXXWPE/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-QEBSV8@pretalx.sciwork.dev
DTSTART:20261205T094000Z
DTEND:20261205T101000Z
DESCRIPTION:Ten years after AlphaGo demonstrated the remarkable potential o
 f combining learning and search\, reinforcement learning has become an imp
 ortant framework for building agents that learn to make sequential decisio
 ns.\n\nIn this talk\, we will give an introduction to the core ideas behin
 d reinforcement learning and planning. We will start from rewards and the 
 basic formulation of sequential decision-making\, then introduce value fun
 ctions\, Q-value functions\, and policies. From there\, we will discuss po
 licy evaluation\, policy iteration\, and policy-based methods\, building a
 n intuitive picture of how an agent can learn to improve its behavior thro
 ugh interaction and experience.\n\nFinally\, we will move from learning to
  planning and introduce Monte Carlo Tree Search (MCTS)\, one of the key co
 mponents behind AlphaGo. By connecting these concepts\, the talk aims to p
 rovide a conceptual roadmap for understanding how reinforcement learning a
 nd search algorithm can work together to solve complex decision-making pro
 blems.
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:On the Tenth Anniversary of AlphaGo: An Introduction to Reinforceme
 nt Learning and Planning - Sheng-Wen Cheng
URL:https://pretalx.sciwork.dev/sw26/talk/QEBSV8/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-VTNVUA@pretalx.sciwork.dev
DTSTART:20261205T094000Z
DTEND:20261205T101000Z
DESCRIPTION:這場演講將帶您穿透 LLM 黑箱，探索機制可解釋
 性（MI）的最新進展，它本質上是一項結合多個研究領域
 的挑戰。我們將援引各個假說來剖析模型內部的概念運
 作，討論如何近似手術般精準去觀察、操控模型，並理
 性展望延伸問題。
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:打開 AI 的黑盒子：機制可解釋性（Mechanistic Interpretability）的最新進展 - Junli Xiao
URL:https://pretalx.sciwork.dev/sw26/talk/VTNVUA/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-88UBCZ@pretalx.sciwork.dev
DTSTART:20261205T102000Z
DTEND:20261205T105000Z
DESCRIPTION:三年前，聽到有人想用 AI 幫 GCC 加上新的指令支
 援時，我的第一反應是：X! 笑死!\n\n三年後，AI Agent 完全
 融入我的日常 compiler 開發流程：修 bug、加測試、研究效
 能問題、分析 IR 與 assembly。\n\n這場演講會從一位 Open Sou
 rce Compiler Developer 的角度，分享在 GCC、LLVM 等大型開源專
 案中使用 AI Agent 的經驗。
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:AI Agent x Open Source Compiler - Kito Cheng
URL:https://pretalx.sciwork.dev/sw26/talk/88UBCZ/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-8HGEEN@pretalx.sciwork.dev
DTSTART:20261205T102000Z
DTEND:20261205T105000Z
DESCRIPTION:PyVista is a VTK-based Python library for 3D data visualization
 . In this talk\, we introduce the technology stack\, implementation detail
 s\, and performance characteristics of `pyvista-wasm`\, which enables PyVi
 sta to run entirely on WebAssembly (WASM). By combining Pyodide and VTK-WA
 SM\, we can deliver full-featured 3D scientific visualization directly in 
 the browser—without requiring a backend server. We will share hands-on t
 echnical insights from development alongside live browser-based demos.\n\n
 ---\n\n**Background and Motivation**\n\nHistorically\, scientific visualiz
 ation has relied on local desktop environments or heavy server-side render
 ing pipelines. With the rise of WebAssembly\, heavyweight scientific libra
 ries like VTK are following NumPy and SciPy into browser environments\, op
 ening up entirely new architectural possibilities.\n\n**Talk Outline (30 m
 ins)**\n\n1. **Overview of PyVista and WASM (5 mins)**\n* What is PyVista?
 \n* Why WASM? What changes for scientific workflows?\n\n\n2. **Technology 
 Stack & Implementation Details (12 mins)**\n* Emscripten builds of VTK\n* 
 Running PyVista on Pyodide\n* Integrating with WebGL rendering\n* Challeng
 es and solutions in local development and CI pipelines\n\n\n3. **In-Browse
 r Live Demo (8 mins)**\n* Running inside JupyterLite\n* Interactive 3D mes
 h manipulation\n\n\n4. **Performance & Limitations (5 mins)**\n* Performan
 ce comparison: WASM vs. native\n* Current WASM constraints and practical w
 orkarounds\n* Future roadmap
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:PyVista on WebAssembly: Enabling Serverless 3D Visualization - Tets
 uo Koyama
URL:https://pretalx.sciwork.dev/sw26/talk/8HGEEN/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-S97GU7@pretalx.sciwork.dev
DTSTART:20261205T110000Z
DTEND:20261205T113000Z
DESCRIPTION:Matrix multiplication may look like three nested loops or a sin
 gle BLAS call\, but implementing it efficiently in an array library requir
 es more than a fast computational kernel. Shapes\, strides\, broadcasting\
 , and operand reuse can change which execution strategy works best—even 
 when the arithmetic workload remains the same.\n\nThis talk explores the i
 mplementation and optimization of matrix multiplication in Solvcon’s Sim
 pleArray. We follow the path from direct computation to layout-aware execu
 tion\, covering BLAS-compatible layouts\, signed-stride operands\, and bat
 ched matrix multiplication. Particular attention is given to small-matrix 
 execution\, reuse of broadcast operands\, and the trade-offs between compl
 ete packing and streamed packing with reusable scratch storage. We also di
 scuss how separating operation semantics\, execution selection\, and compu
 tation makes these optimizations easier to integrate and maintain.\n\nTo e
 valuate the implementation\, we briefly demonstrate a Qt Benchmark Inspect
 or and explain how identical operands\, numerical comparisons\, and balanc
 ed timing help assess performance improvements. The central lesson is that
  efficient matrix multiplication depends not only on kernel speed\, but al
 so on how data layout\, reuse\, and execution strategy fit together.
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:Beyond Calling BLAS: Implementing and Optimizing Matrix Multiplicat
 ion - Chun-Shih Chang
URL:https://pretalx.sciwork.dev/sw26/talk/S97GU7/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-BPRMPT@pretalx.sciwork.dev
DTSTART:20261205T110000Z
DTEND:20261205T113000Z
DESCRIPTION:當 AI Agent 能夠協助產生大量程式碼，開發者如
 何確認一個跨越幾何、聲學與即時互動的應用，真的符
 合原先的設計？\n\n　　本演講以 OpenAI Build Week 期間開發
 的 Echo Canvas 為案例，分享如何與 Codex 協作，將自然語言
 描述與可編輯的空間場景，轉換成能在瀏覽器中操作與
 試聽的空間音訊原型。使用者可以調整牆面、開口、音
 源與聆聽位置，觀察聲音路徑，並比較遮蔽、早期反射
 與殘響帶來的變化。\n\n　　演講將說明兩個互相配合的
 設計：以規格、資料契約與測試約束 Codex 的開發工作；
 在應用內，讓 GPT-5.6 負責編譯候選場景及解釋結果，並
 透過結構與領域驗證，把模型輸出交給確定性的幾何與
 聲學計算，再由 Web Audio 完成瀏覽器 HRTF 渲染。\n\n　　
 透過架構拆解與 Demo，我們將討論如何界定 AI 的責任、
 檢查生成結果，以及區分自動測試與人耳驗收所能證明
 的事情。\n\n　　Echo Canvas 定位為用於空間音訊原型設計
 與預視的互動式聲學近似；演講也會交代近似方法與使
 用限制。聽眾將帶走一套把 AI 協作、可驗證計算與互動
 展示連接起來的工程實作思路。
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:從自然語言到空間音訊：用 Codex 打造可驗證的瀏覽器聲學原型 - Kevin
URL:https://pretalx.sciwork.dev/sw26/talk/BPRMPT/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-JJXHKS@pretalx.sciwork.dev
DTSTART:20261205T113000Z
DTEND:20261205T125000Z
DESCRIPTION:Lunch
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:Lunch - 
URL:https://pretalx.sciwork.dev/sw26/talk/JJXHKS/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-XRJ7SE@pretalx.sciwork.dev
DTSTART:20261205T113000Z
DTEND:20261205T125000Z
DESCRIPTION:Lunch
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:Lunch - 
URL:https://pretalx.sciwork.dev/sw26/talk/XRJ7SE/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-NBDZ3J@pretalx.sciwork.dev
DTSTART:20261205T125000Z
DTEND:20261205T134000Z
DESCRIPTION:Project Sharing
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:Project Sharing - 
URL:https://pretalx.sciwork.dev/sw26/talk/NBDZ3J/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-CPDGQT@pretalx.sciwork.dev
DTSTART:20261205T135000Z
DTEND:20261205T163000Z
DESCRIPTION:小時候是不是也曾幻想過寫程式控制火箭升空？
 現在，你的 code 真的能主宰火箭的飛行軌跡！隨著臺灣
 入軌火箭計畫的推進，我們正徵招「導航、制導與控制
 （GNC）」的跨領域軟體人才，並分享國家太空中心（TASA
 ）與陽明交通大學前瞻火箭研究中心（ARRC）共同推動的
 「[火箭導控軟體競賽（Balloon Popping Challenge）](https://www.t
 asa.org.tw/en-US/events/detail/273c7841-ce66-400a-a501-1bf0f7fdc17e)」。
 \n\n這是個結合 [Gymnasium](https://github.com/Farama-Foundation/Gymna
 sium) 訓練環境與 [ActiveRocketPy](https://github.com/ARRC-Rocket/Ball
 oonPoppingChallenge) 模擬系統(基於開源專案 [RocketPy](https://gi
 thub.com/RocketPy-Team/RocketPy)) 的火箭導控競賽。活動中會手
 把手帶領與會者嘗試控制火箭的推力向量控制（TVC）系
 統，並進一步探討如何應對陣風、感測器雜訊等複雜的
 真實情境。\n\n「Be a User -> Be a Contributor」是這場工作坊
 的核心精神！我們非常鼓勵大家參與開源貢獻：無論你
 想發揮 Computer Science 背景幫忙打造飛行軌跡的資料視覺
 化工具，還是對物理學有熱忱，想發 PR 改善火箭 Python 
 模組的模擬特性，這裡都有你大展身手的舞台。\n\n本工
 作坊無需事先報名，敬請攜帶筆電，一起成為太空開源
 專案的貢獻者！
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:體驗火箭導控 Balloon Popping Challenge - William Mou
URL:https://pretalx.sciwork.dev/sw26/talk/CPDGQT/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-STY3KU@pretalx.sciwork.dev
DTSTART:20261205T135000Z
DTEND:20261205T142000Z
DESCRIPTION:Werilog is a browser-based environment for Hardware Description
  Languages that removes the "local toolchain tax" of hardware design — n
 o installs\, no simulator setup\, just coding in the browser.\n\nAt first 
 glance\, Werilog looks like another EDA tool: write Verilog\, see it simul
 ate. But the real question we're exploring is the opposite direction — c
 an you draw a diagram and get working codes back out? Most tools only go o
 ne way: code in\, diagram out\, read-only. We're building toward a loop wh
 ere the diagram is not just a visualization but an editable\, valid repres
 entation of the code itself — and where structured\, regular designs (li
 ke datapaths and control logic) can be constructed visually and compiled d
 own to correct HDL.\n\nThis talk covers why we think bidirectionality matt
 ers for how people learn and design hardware\, what we've built so far (li
 ve HDL-to-diagram visualization\, in-browser simulation\, an early diagram
 -to-HDL path for structural designs)\, the architectural choices and const
 raints that shaped it\, and where the project is headed — including wher
 e AI assistance fits into closing the loop.\n\nThis talk covers why we thi
 nk bidirectionality matters for how people learn and design hardware\, wha
 t we've built so far (live HDL-to-diagram visualization\, in-browser simul
 ation\, an early diagram-to-HDL path for structural designs)\, and the arc
 hitectural choices and constraints that shaped it.
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:Werilog: Toward a Bidirectional Coding Architecture - Sivan
URL:https://pretalx.sciwork.dev/sw26/talk/STY3KU/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-MMHXE3@pretalx.sciwork.dev
DTSTART:20261205T143000Z
DTEND:20261205T150000Z
DESCRIPTION:我們會在這場 talk 裡會以 bbgo 為基礎，希望以深
 入淺出的方式，介紹以下主題：\n- 回測引擎的元件與開
 發\n- 可測試交易策略開發\n但不會包含：\n- 交易策略相
 關指標介紹與優化
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:Shall We Test? BBGO 可測試交易策略開發 - Dboy Liao
URL:https://pretalx.sciwork.dev/sw26/talk/MMHXE3/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-CSADBJ@pretalx.sciwork.dev
DTSTART:20261205T150000Z
DTEND:20261205T160000Z
DESCRIPTION:Tea Break
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:Tea Break - 
URL:https://pretalx.sciwork.dev/sw26/talk/CSADBJ/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-EFCHZR@pretalx.sciwork.dev
DTSTART:20261205T160000Z
DTEND:20261205T163000Z
DESCRIPTION:大型語言模型對大多數開發者來說\,常常是一個
 「輸入文字、吐出答案」的黑盒子。我們每天呼叫 API、
 串接 Chatbot\,卻很少真正拆開來看:這些模型內部到底在算
 什麼?當我們說「微調一個模型」\,實際上又是在調整什
 麼、可能會踩到哪些坑?這場分享想從這兩個角度\,帶大
 家從「懂原理」走到「真的做得出來」。\n\n這次分享主
 要會分成兩個段落：\n\n第一段會用一個極簡版的 GPT 實
 作(MicroGPT)\,搭配一份刻意設計過的小型測試資料\,現場展
 示一個語言模型從「什麼都不懂」到「學會特定規律」
 的完整過程。測試資料來自 python-docs-zh-tw 的術語列表\,
 並刻意把原本的命名機器改成翻譯機器(例如把「access」
 對應到「存取」)\,讓資料規模小到可以在一般筆電上、
 幾分鐘內完成訓練\,同時保留真實的語言對應關係。透過
 這個極簡模型\,我們可以清楚看到:文字是怎麼被切成 toke
 n、token 怎麼變成向量、注意力機制怎麼決定「這個詞該
 參考哪些上下文」。因為模型夠小、資料夠少\,整個過程
 是完全透明、可以逐行程式碼解釋的。\n\n第二段會實際
 用開源預訓練模型\,微調出一個「Python 技術文件英翻中
 」的專用翻譯模型。訓練資料完全來自 python/python-docs-zh-
 tw 這個開源翻譯專案:透過解析專案裡的 .po 翻譯檔案\,把
 每一組已經由社群人工翻譯、審核過的英文原文與繁體
 中文譯文\,轉換成訓練用的格式。\n\n在模型選擇上\,會分
 享實際嘗試過的幾個 Hugging Face 開源模型並使用 LoRA(透過
  Unsloth 加速)這種參數高效微調技術\,只調整模型裡一小
 部分參數\,就能讓模型學會特定領域的翻譯能力\,而不需
 要重新訓練整個模型；以及在模型訓練時實際踩過的各
 種坑。
DTSTAMP:20261122T200526Z
LOCATION:403
SUMMARY:從 MicroGPT 到真實模型:拆解 LLM 原理\,再用 LoRA 微調打造 Python 文件繁中翻譯 LLM - Timmy
URL:https://pretalx.sciwork.dev/sw26/talk/EFCHZR/
END:VEVENT
BEGIN:VEVENT
UID:pretalx-sw26-DSWD9R@pretalx.sciwork.dev
DTSTART:20261205T165000Z
DTEND:20261205T170000Z
DESCRIPTION:Closing
DTSTAMP:20261122T200526Z
LOCATION:402
SUMMARY:Closing - 
URL:https://pretalx.sciwork.dev/sw26/talk/DSWD9R/
END:VEVENT
END:VCALENDAR
