Glossary

The neural network architecture proposed in 2017 that underpins contemporary large language models

Heard in紀懷新, 吳柏翰, 林彥廷

Sequence-to-sequence learning; the turn that took machine translation from statistical phrase tables to neural networks

Heard in紀懷新

Having the model emit intermediate reasoning steps to improve its problem solving

Heard in紀懷新

Google's early conversational language model, whose line of work later became Bard and Gemini

Heard in紀懷新

Google's current multimodal model family

Heard in紀懷新, 吳振和, 陳志信

Google DeepMind's research prototype for a universal AI assistant

Heard in紀懷新

Extending Gemini's multimodal capability to robot control

Heard in紀懷新

Using a large model's outputs to teach a small one, bringing the small model close to the large one's performance

Heard in吳柏翰

Spending more compute at inference for a better answer — sampling repeatedly, or reasoning for longer

Heard in吳柏翰, 林彥廷

The model handles a task it was never trained on

Heard in紀懷新

Borrowed from dual-process theory; maps onto intuitive versus deliberate reasoning tasks

Heard in紀懷新

A model's tendency to trade accuracy for agreeing with the user

Heard in林彥廷

Sampling the same question repeatedly and keeping only the correct results as training data

Heard in林彥廷

A hard mathematical problem-solving benchmark

Heard in林彥廷

Graduate-level question answering, designed so the answers cannot be found with a search engine

Heard in林彥廷

Benchmarks an AI's ability to resolve real software engineering issues

Heard in林彥廷

OpenAI's rubric-based benchmark for the medical domain

Heard in林彥廷

Benchmarks an AI's ability to reproduce academic papers

Heard in林彥廷

A head-to-head platform where users vote between two models' answers to produce a ranking

Heard in林彥廷

OpenAI's framework for tracking model capability and risk

Heard in林彥廷

A per-question list of scoring criteria, which lets a model act as the grader

Heard in林彥廷, 吳剛志

Building the evaluation mechanism before building the LLM application

Heard in吳剛志 (citing ihower)

Coverage is the share of questions the AI answers at all; accuracy is the share it gets right among those

Heard in吳剛志

An open protocol for connecting AI applications to external data and tools

Heard in張文鈿, 保哥, 黃仕鎮

Respectively callable tools, readable data, and reusable prompt templates

Heard in張文鈿

MCP's two transports — the first between local processes, the second over the network

Heard in張文鈿

In MCP, the server asking the client to run an inference, and declaring which root directories are reachable

Heard in張文鈿

Google's proposed protocol for agent-to-agent interoperability, solving a different layer than MCP

Heard in張文鈿

Letting a model drive the screen, mouse and keyboard directly

Heard inWisely Chen, Peggy Lo

An open-source browser automation framework that parses HTML structure as well as taking screenshots

Heard inWisely Chen

Keeping a human confirmation gate inside an automated flow

Heard inCaesar Chi, 保哥

An arrangement where one upper-level agent assigns tasks to lower-level agents

Heard inCaesar Chi

GitHub's AI coding assistant; agent mode can edit across files and call external tools

Heard in林鉦育, 保哥

Anthropic's agentic coding tool

Heard in保哥

Two editors built around AI

Heard inPeggy Lo, Vivi Chen, 張文鈿

An open-source workflow automation platform that can be self-hosted

Heard in高見龍, Vivi Chen

A visual development platform for agents and retrieval workflows

Heard in吳剛志, 張文鈿

The scripting platform attached to Google Sheets and Drive; the most common first automation entry point for non-engineers

Heard inPeggy Lo, Vivi Chen, 鄭晴元

A plain-text syntax for flowcharts that both machines and people can read

Heard in鄭晴元, 吳剛志

Two frameworks for multi-agent and LLM application development

Heard inCaesar Chi

The official debugging tool for MCP servers

Heard in張文鈿

Retrieve from a knowledge base before generating, so the model's answer has something behind it

Heard in黃仕鎮, 吳振和, Caesar Chi

Retrieval over images and audio-visual material, not just text

Heard in吳振和

Recognition accuracy on fixed formats can be made very high; unconstrained formats are the hard part

Heard in黃仕鎮

Masking sensitive information before anything is sent to an external model

Heard in黃仕鎮

Two ways of making a model emit structured results a program can consume directly

Heard in吳剛志

Deliberately tilting the frame to create unease

Heard in陳志信

Marking with colour instead of directional words like left, right, front and back, to steer the model's generated motion

Heard in陳志信

Keeping the lighting and environment of a live-action plate and replacing the person in frame

Heard in陳志信

The five traditional image manipulation techniques

Heard in李怡志

The psychological framework for judging where a memory came from

Heard in李怡志

The dual-process model of information handling; scrolling social feeds is mostly the former

Heard in李怡志

The more information there is, the easier the gap between the informed and the uninformed is to exploit

Heard in李怡志

Anything that fits your existing position gets believed

Heard in李怡志

The Traditional Chinese large language model project developed by 林彥廷

Heard inGitHub

The enterprise AI company founded by 吳柏翰, focused on customer-specific model fine-tuning and on-premises deployment

Heard inSite

E.SUN Bank's in-house generative AI chat platform

Heard inSpeaker's own account, no external source

E.SUN Bank's in-house meeting summary and action-item extraction service

Heard inSpeaker's own account, no external source

Simpleinfo's internally built Slack AI bot, triggered by an emoji reaction

Heard inSpeaker's own account, no external source

The packaging design-to-manufacturing collaboration platform built by Caesar Chi's team

Heard inCympotek

The company founded by 劉詩雁, whose session covered AI adoption in daily-chemical printing

Heard inINSIDE interview

The YouTube channel run by Simpleinfo

Heard inChannel

The technical education brand co-founded by 高見龍

Heard inSite

The generative AI engineering newsletter run by 張文鈿

Heard inihower.tw