Maddy Zhang

Local AI sounds complicated. RAM, VRAM, quantization, model sizes, runtimes, and a bunch of tools just to run a model that you could use instantly with ChatGPT or Claude.

So why bother?

In this video, I break down why local AI is becoming increasingly useful in 2026, what actually happens when you run a model on your own machine, how quantization and memory work and how to get your first local model running.

🚀 What you'll learn
✅ Why local AI is becoming an important software engineering skill
✅ The biggest advantages of running AI locally: privacy, cost, control, and offline access
✅ What model weights, runtimes, RAM, VRAM, and memory bandwidth actually mean
✅ How quantization makes large models practical on consumer hardware
✅ How to run your first local model with LM Studio or Ollama
✅ Which model sizes make sense for 8GB, 16GB, 24GB, and 32GB+ machines
✅ How to connect a local model to your existing applications through an API
✅ How local coding agents work with your files, terminal, and editor
✅ The difference between retrieval and fine-tuning
✅ How LoRA and lightweight fine-tuning work
✅ When local AI is better than cloud AI, and when it isn't
✅ How to think about local AI for real-world products and AI workloads

👉 Watch the full breakdown here: https://youtu.be/6hQL9dbnNqA

15 hours ago | [YT] | 300

Maddy Zhang

Went to OpenAI DevDay 2026 in SF!

With 20+ announcements, here's what stood out to me:

-Dots: always-on agents! When you’re not actively working with one, it does proactive research in the background, and it can connect to 4,000+ apps through plugins.
-GPT-6.1 Sol: offers close to Astra-level performance on agentic coding and computer use, at a fifth of the token price.
-Ultrafast: a new premium speed tier, up to 8x faster token generation in Codex.
-Codex expansion: you can run it in the cloud from any device, steer the CLI with your voice, and have it take a first pass on PRs before you look at them.
-MCP Events support: plugins can now kick off automations when something happens in a connected app.

My favorite moments were watching Romain Huet have Codex build a raffle app on stage that picked six attendees for tickets to next year’s DevDay, and seeing Thibault Sottiaux press the button to reset everyone’s ChatGPT account usage 😮

And of course, the people were the best part! Got to chat up with so many friends, new and old.

Thanks for having me, ‪@OpenAI‬!

🏷️ #OpenAI #conference #Codex #AI #AgenticAI #SoftwareEngineering

Skip the Carousel

1 2 3 4 5 6 7 8 9 10

1 week ago | [YT] | 225

Maddy Zhang

AI agents feel easy when you're building a demo. But once you try to ship one to production, you realize an agent isn't just an API call. It reasons, calls tools, keeps state, and takes multiple steps, and every step is another way for it to fail.
In this video, I build the architecture of an agentic AI system from the ground up, using a customer support agent as the example.

💡 Here's what you'll learn:
✅ The Agent Loop: why an agent is a loop, not a prompt, and how reliability erodes with every step
✅ Model Routing: how to balance cost and latency against quality by matching models to tasks
✅ Tools & MCP: why tools should be designed like strict APIs, with read and write actions kept separate
✅ Memory & State: how to treat context as a scarce resource and avoid context rot
✅ RAG: why bad retrieval is behind so many bad AI answers
✅ Orchestration: when to use a defined pipeline, a single agent, or multiple agents
✅ Evals & Observability: how to test non-deterministic systems and debug individual agent runs
✅ Security & Production Controls: how to limit blast radius and defend against prompt injection

👉 Watch the full breakdown here: https://www.youtube.com/watch?v=UqnLq...

1 week ago | [YT] | 33

Maddy Zhang

As a senior software engineer, I have spent the last few years working extensively with AI and software engineering.

In this video, I break down my entire agentic AI workflow from setting up agents with the right context and memory, to planning, parallel agents, autonomous loops, graph-based workflows and reviewing AI-generated code safely

Here’s what I cover 👇
✅ How I set up memory files and skills so agents understand my codebase and conventions
✅ Why you should plan before letting an agent write code
✅ How I use plan mode to catch expensive mistakes before implementation
✅ How meeting context and requirements can flow directly into coding workflows
✅ How I use terminal-based workflows, agent harnesses, and Git worktrees
✅ How to run multiple AI agents in parallel with isolated contexts
✅ Loop engineering: designing agents that can write, test, fix, and repeat on their own
✅ Graph engineering: combining multiple agent loops for research, implementation, and review
✅ How to move from manually prompting agents to building more autonomous workflows
✅ How I decide how deeply to review AI-generated code based on risk
✅ When a quick diff review is enough and when I manually inspect every line
✅ Why testing, staging, screenshots, and real-world validation still matter with AI-generated code

👉 Watch the full breakdown here: https://youtu.be/tSF6eZW8ZMU

2 weeks ago | [YT] | 672

Maddy Zhang

Working at Microsoft completely changed how I think about software engineering.

I got to work on Roslyn analyzers and developer tools used by millions of developers, and being inside such a massive engineering organization taught me lessons that go far beyond writing code.

In this video, I break down the 7 biggest things I learned at Microsoft, from how to approach learning and criticism to designing for users you’ll never meet, maintaining backwards compatibility, building secure software, and adapting to AI.

🚀 What you’ll learn
✅ Why Microsoft values being a “learn-it-all” instead of a “know-it-all”
✅ Why great engineers design for users they’ll never meet
✅ What backwards compatibility teaches you about building software that lasts
✅ Why security should be part of engineering from the very beginning
✅ Why shipping and learning from real user behavior beats relying on assumptions
✅ How design reviews and constructive criticism can save you weeks of wasted work
✅ Why adaptability matters more than mastering any single AI tool
✅ What Microsoft’s changing AI strategy reveals about the future of software engineering
✅ How to stay useful even when the tools and technologies keep changing

👉 Watch the full breakdown here: https://youtu.be/2CN02BZY59A

3 weeks ago | [YT] | 657

Maddy Zhang

AI was supposed to replace jobs, transform industries, and become the biggest technological shift in history. But the numbers tell a different story. Big Tech is spending hundreds of billions on AI while many projects struggle to deliver real returns.

In this video, I break down the 2026 AI boom, where the money is going, why AI projects are failing, and how rising costs, energy demands, hallucinations, and legal risks are exposing the cracks.

So, are we heading toward one of the biggest tech bubbles in history? And what does it mean for software engineers, AI engineers, and tech careers?

🚀 What you’ll learn
✅ How much Big Tech is actually spending on AI in 2026
✅ Why the AI spending boom may be getting ahead of the real value
✅ What the $7+ trillion AI investment thesis means
✅ Why AI infrastructure and energy demand are becoming major constraints
✅ Why so many enterprise AI pilots fail to deliver ROI
✅ Why companies are rolling back AI customer-service agents
✅ How AI hallucinations create serious legal and business risks
✅ What the AI bubble could have in common with the dot-com crash
✅ Why AI isn’t necessarily failing even if the AI bubble bursts
✅ Which skills and roles could become more valuable after the correction
✅ How software engineers can position themselves for what comes next

👉 Watch the full breakdown here: https://youtu.be/oGCb9gGBBrw

1 month ago | [YT] | 948

Maddy Zhang

I solved 400+ LeetCode problems and interviewed candidates at Google. And here's the truth I wish I knew when I started:

You DON'T need to solve hundreds of LeetCode problems to crack a Big Tech interview.

What you actually need is to recognize the patterns that keep appearing across problems.

In this video, I break down the 9 most important DSA patterns you need to master in 2026, the clues that tell you when to use each one, and the LeetCode problems you should practice to build real problem-solving intuition.

🚀 What you'll learn:
✅ Why solving 400+ LeetCode problems isn't necessary
✅ Why DSA and coding interviews still matter in 2026
✅ The 9 DSA patterns that appear again and again
✅ How to recognize when a problem requires hashing
✅ When to use two pointers and fast/slow pointers
✅ How sliding window problems actually work
✅ How to spot binary search even without a sorted array
✅ When to use a monotonic stack
✅ DFS vs BFS and when to use each
✅ How heaps solve Top K problems efficiently
✅ When backtracking is the right approach
✅ How to recognize dynamic programming problems
✅ The key clues hidden inside problem statements
✅ The best LeetCode problems to practice for each pattern
✅ How to approach problems you've never seen before

👉 Watch the full breakdown here: https://youtu.be/Q5QoGocSnjo

1 month ago | [YT] | 1,266

Maddy Zhang

I spent 3.5 years at MIT studying computer science, one of the most intense academic environments in the world which helped me land internships and full-time offers at companies like Google, Amazon, Microsoft, IBM, and Snap.

But what does a 4-year computer science degree actually teach you—and which parts still matter when you're working as a software engineer?

In this video, I break down all four years of computer science at MIT, the most important skill each year teaches you, what I actually use at work, and what you can learn for free outside of university.

Here’s what I cover 👇
✅ What you actually learn in Year 1: programming and discrete math
✅ Why Python is disposable but learning to think precisely isn't
✅ How data structures and algorithms change the way you solve problems
✅ Why Big O and complexity analysis matter even more in the AI era
✅ What computer architecture, operating systems, networking, and databases teach you about real-world systems
✅ Why correct code can still fail in production
✅ How to develop systems-thinking and debug problems one layer below the symptoms
✅ What Year 4 electives and capstone projects teach you about real engineering work
✅ The biggest gap in a traditional CS degree: Git, terminals, debugging, profiling, and developer tooling
✅ The one resource I consider the highest-ROI addition to your CS education

👉 Watch the full breakdown here: https://youtu.be/GfqGmEKK_Ns

1 month ago | [YT] | 959

Maddy Zhang

Most people learn machine learning the wrong way. They jump straight into algorithms and math without understanding the core ideas that make ML systems actually work.

In this video, I break down 10 essential machine learning concepts that every software engineer and machine learning engineer should understand, from the foundations of training a model to the ideas powering modern AI and LLM systems.

Whether you're learning machine learning for the first time, preparing for ML interviews, or trying to understand how modern AI works under the hood, this video will help you build the right mental models.

💡 Here’s what you’ll learn:
✅ Training vs Inference — how models learn during training and generate outputs during inference
✅ Loss & Gradient Descent — how models measure mistakes and update their weights
✅ Generalization & Overfitting — why memorizing training data doesn't make a useful model
✅ Embeddings — how AI represents meaning as vectors for search, recommendations, and retrieval
✅ Tokens — how language models actually process text and why tokenization matters
✅ Attention — the mechanism that powers modern Transformer-based models
✅ Context Windows & Context Rot — why bigger context doesn't always mean better reasoning
✅ Pre-training & Post-training — how models go from raw knowledge to useful AI assistants
✅ RAG — how retrieval-augmented generation brings external knowledge into model context
✅ Evals — how to systematically measure whether an AI system is actually getting better


👉 Watch the full breakdown here: https://youtu.be/zVX7lOf5ER4

1 month ago | [YT] | 825

Maddy Zhang

Which video should I make next?

1 month ago | [YT] | 64