Learn by Doing with Steven

新文章：风险应该决定你的 AI 架构

AI 现在很容易搭起来。也正因为如此，风险比以前更重要。

低风险推荐工具可以接受轻量解释；但临床分诊、公共福利、信贷、招聘、欺诈检测、公共安全这些高风险场景，需要完全不同的架构：可追踪、可监控、可审计、可人工介入、可申诉。

模型不是架构。真正的架构，是围绕模型建立的一整套 operating system。

阅读全文：
learnbydoingwithsteven.github.io/
linktr.ee/learnbydoingwithsteven

Ref: IBM Technology
https://www.youtube.com/watch?v=3G6AV...

1 week ago | [YT] | 0

Learn by Doing with Steven

New article: Risk Should Determine Your AI Architecture

AI is easy to build now. That is exactly why risk matters more.

A low-risk recommendation tool can live with lightweight explainability. A high-risk system used for clinical triage, public benefits, credit, hiring, fraud detection, or public safety needs a different architecture: traceability, monitoring, auditability, human override, and contestability.

The model is not the architecture. The architecture is the operating system around the model.

Read more:
learnbydoingwithsteven.github.io/
linktr.ee/learnbydoingwithsteven

Ref: IBM Technology
https://www.youtube.com/watch?v=3G6AV...

1 week ago | [YT] | 0

Learn by Doing with Steven

AI Terms You Should Know | Episode #11

Loss Function: How Does AI Know It Is Wrong?

A loss function measures the difference between a model's prediction and the correct answer.

Loss tells the model: How wrong am I?

Prediction -> Calculate Loss -> Backpropagation -> Optimization -> Repeat

Common loss functions include MSE, MAE, Binary Cross-Entropy, Categorical Cross-Entropy, and Hinge Loss.

What should we unpack next: Optimizer, Parameter, Epoch, or Batch Size?

#AI #LossFunction #MachineLearning #DeepLearning #NeuralNetworks #AITraining #AITerms

1 week ago | [YT] | 0

Learn by Doing with Steven

AI Terms You Should Know | Episode #12

Optimizer: The Algorithm That Decides How the Model Learns

Backpropagation calculates gradients.
The optimizer decides how to turn those gradients into parameter updates.

Loss Function -> measures the mistake
Backpropagation -> calculates gradients
Optimizer -> updates parameters
Learning Rate -> controls step size

Common optimizers include SGD, Momentum, RMSProp, Adam, and AdamW.

Should Episode #13 be Parameter, Epoch, Batch Size, or Inference?

#AI #Optimizer #MachineLearning #DeepLearning #Adam #AdamW #SGD #AITraining #AITerms

1 week ago | [YT] | 4

Learn by Doing with Steven

AI Terms You Should Know | Episode #10

Learning Rate: How Big Should Each Learning Step Be?

Gradient Descent tells an AI model which direction to move.

The Learning Rate tells it how far to move.

Too high -> unstable.
Too low -> painfully slow.
Well chosen -> efficient learning.

Sometimes, one tiny number can determine whether an entire AI training run succeeds or fails.

What should we unpack next: Loss Function, Optimizer, Parameter, or Epoch?

#AI #LearningRate #GradientDescent #MachineLearning #DeepLearning #AITraining #NeuralNetworks #AITerms

1 week ago | [YT] | 1

Learn by Doing with Steven

New article: Why AI Benchmarks Are Breaking

AI progress is real, but our measurement systems are under pressure. Benchmarks can be contaminated, harness changes can make the same model behave differently, inference providers can trade accuracy for speed, and agents can learn to exploit reward signals.

The real question is no longer just “which model scored highest?” It is whether the whole stack remains truthful, useful, and controllable in production.

Read more:
learnbydoingwithsteven.github.io/
linktr.ee/learnbydoingwithsteven

中文：
新文章：AI Benchmark 正在失效

AI 的进步是真的，但测量体系正在承压。Benchmark 可能被污染，harness 的小改动可能让同一个模型表现完全不同，推理服务商可能为了速度牺牲准确率，agent 也可能学会利用 reward signal 的漏洞。

真正的问题不再只是“哪个模型分数最高”，而是整套 stack 离开 demo 以后，是否仍然真实、有用、可控。

阅读全文：
learnbydoingwithsteven.github.io/
linktr.ee/learnbydoingwithsteven

Full articles / 全文：
learnbydoingwithsteven.substack.com/p/why-ai-bench…
learnbydoingwithsteven.substack.com/p/ai-benchmark…

1 week ago | [YT] | 0

Learn by Doing with Steven

AI Terms You Should Know | Episode #07

Training: How an AI Model Learns

Training is how an AI model learns from data by making predictions, measuring mistakes, and adjusting itself repeatedly.

Prediction -> Error -> Adjustment -> Better Prediction

Training builds or improves the model. Inference uses the trained model to produce responses.

Which part should I explain next: loss functions, backpropagation, parameters, or optimizers?

#AI #AITraining #MachineLearning #DeepLearning #NeuralNetworks #AITerms #LearnByDoingWithSteven

3 weeks ago | [YT] | 1

Learn by Doing with Steven

The Engineer's Verdict: Why Your Name on the Code Matters More Than Ever

Agents can produce evidence. They cannot carry answerability. Engineering now means owning the verdict: ship, block, redirect, or accept risk.

工程师的 Verdict：为什么你的名字比以往更重要

Agent 可以产生证据，但不能承担责任。工程师的关键角色，是对最终 verdict 负责。

Read more / 全文：
open.substack.com/pub/learnbydoingwithsteven/p/the…
open.substack.com/pub/learnbydoingwithsteven/p/ver…

learnbydoingwithsteven.github.io/
linktr.ee/learnbydoingwithsteven

#AI #AIAgents #SoftwareEngineering #AIEngineering #CodeReview

3 weeks ago | [YT] | 0

Learn by Doing with Steven

The Context Layer: The Missing Infrastructure for Production-Grade AI Agents

Production AI needs more than model intelligence. It needs a shared context layer for business definitions, domain skills, norms, permissions, and learning loops.

Context Layer：生产级 AI Agent 缺失的基础设施

生产级 AI 不只需要模型智能，还需要共享的 context layer：业务定义、领域技能、组织规范、权限和学习循环。

Read more / 全文：
open.substack.com/pub/learnbydoingwithsteven/p/the…
open.substack.com/pub/learnbydoingwithsteven/p/con…

learnbydoingwithsteven.github.io/
linktr.ee/learnbydoingwithsteven

#AI #AIAgents #ContextEngineering #EnterpriseAI #LLM

3 weeks ago | [YT] | 0

Learn by Doing with Steven

Stop Burning Tokens: Build Domain Expertise Into the AI Loop

The next phase of AI development is not better prompt wording. It is better loop design: high-signal feedback, expert judgment, and validation.

别再烧 Token：把领域专家放进 AI 改进循环

AI 开发的下一阶段，不是更漂亮的 prompt，而是更好的循环：高信号反馈、专家判断和验证。

Read more / 全文：
open.substack.com/pub/learnbydoingwithsteven/p/sto…
open.substack.com/pub/learnbydoingwithsteven/p/tok…

learnbydoingwithsteven.github.io/
linktr.ee/learnbydoingwithsteven

#AI #AIEngineering #Evaluation #LLM #AIAgents

3 weeks ago | [YT] | 1