Letters from Andrew Ng

Personal messages to the AI community

Robots extract colorful data streams from silo towers, highlighting data silos being broken.

Tear Down Data Silos!: Many software-as-a-service vendors aim to hold their customers' data in silos. Their customers would do well to open the silos so AI agents can use the data.

AI agents are getting better at looking at different types of data in businesses to spot patterns and create value. This is making data silos increasingly painful.

Robot bakes pizza at 1000 degrees for 5 hours, causing a fire, illustrating mistake in error analysis.

Technical Insights

Improve Agentic Performance with Evals and Error Analysis, Part 2: Best practices for error analysis in agentic AI development, and how LLMs make them easier

In last week’s letter, I explained how effective agentic AI development needs a disciplined evals and error analysis process, and described an approach to performing evals.

A man at a computer says AI ordered pizza, while a delivery man outside holds a fruit basket, highlighting a mix-up.

Technical Insights

Improve Agentic Performance with Evals and Error Analysis, Part 1: When AI agentic systems go astray, it’s tempting to shortcut evals and error analysis. But these processes cas lead to much faster progress.

Readers responded with both surprise and agreement last week when I wrote that the single biggest predictor of how rapidly a team makes progress building an AI agent lay in their ability to drive a disciplined process for evals...

Cartoon of developer fixing failing AI tests by marking them as passed without solving the code.

Technical Insights

Agentic Coding and Agentic Software Testing Go Together: Agentic coding can make mistakes, but agentic testing can find and fix them.

Automated software testing is growing in importance in the era of AI-assisted coding.

Comic showing tech interviews: 2022 asks “Can you code FizzBuzz?” vs 2025 asks “Can you build an e-commerce platform?”

Technical Insights

AI Skills Are Redefining What Makes a Great Developer: The job market for software developers requires knowing how to use AI

There is significant unmet demand for developers who understand AI.

Cartoon robots with square heads and antennae sit in rows on an assembly line, each smiling while assembling gears, boxes, and tools.

Technical Insights

Agents Running in Parallel Get There Faster: Parallel agents can accelerate AI systems as test-time compute scales up.

Parallel agents are emerging as an important new direction for scaling up AI. AI capabilities have scaled with more training data, training-time compute, and test-time compute.

Colorful LEGO bricks labeled for AI concepts: prompting, agentic, guardrails, evals, RAG, fine-tuning, computer use, async programming.

Technical Insights

Meet The New Breed of GenAI Application Engineers: A new breed of software engineers is building more powerful applications faster than ever, thanks to generative AI. Here’s how to identify them in job interviews.

There’s a new breed of GenAI Application Engineers who can build more-powerful applications faster than was possible before, thanks to generative AI.

Code snippet showing ‘Keep Building!’ printed in multiple programming languages including Python, Java, JavaScript, and C++.

Technical Insights

How to Become a Multilingual Coder: AI makes it easy to code in any programming language — especially if you know just one.

Even though I’m a much better Python than JavaScript developer, with AI assistance, I’ve been writing a lot of JavaScript code recently.

Cartoon of two coworkers coding; one struggles with evaluations, the other iterates quickly through model updates and test cases.

Technical Insights

We Iterate on Models. We Can Iterate on Evals, Too: Building automated evals doesn’t need to be a huge investment. Start with a few quick-and-dirty examples and iterate!

I’ve noticed that many GenAI application projects put in automated evaluations (evals) of the system’s output probably later — and rely on humans to manually examine and judge outputs longer — than they should.

Cartoon of a relaxed man saying “Relax! I’m lazy prompting!” while lounging under a beach umbrella near a stressed coworker at a desk.

Technical Insights

The Benefits of Lazy Prompting: You don’t always need to provide context when prompting a large language model. A quick prompt can be enough.

Contrary to standard prompting advice that you should give LLMs the context they need to succeed, I find it’s sometimes faster to be lazy and dash off a quick, imprecise prompt and see what happens.

Cartoon of a man playing violin saying “I’m fine-tuning!” while a woman at her desk covers her ears, replying “Did you try prompting?”

Technical Insights

When to Fine-Tune — and When Not To: Many teams that fine-tune their models would be better off prompting or using agentic workflows. Here's how to decide.

Fine-tuning small language models has been gaining traction over the past half year.

Illustration of a programmer at a computer displaying PyTorch code, while a smiling colleague gives a thumbs-up in approval.

Technical Insights

Learn the Language of Software: AI won’t kill programming. There has never been a better time to start coding.

Some people today are discouraging others from learning programming on the grounds AI will automate it.

Diagram of an RQ-Transformer speech system with Helium and Depth Transformers for audio processing.

Technical Insights

Wait Your Turn! Conversation by Voice Versus Text: Text interactions require taking turns, but voices may interrupt or overlap. Here’s how AI is evolving for voice interactions.

Continuing our discussion on the Voice Stack, I’d like to explore an area that today’s voice-based systems mostly struggle with: Voice Activity Detection (VAD) and the turn-taking paradigm of communication.

Diagram comparing direct audio generation with a foundation model vs. a voice pipeline using STT, LLM, and TTS.

Technical Insights

What I’ve Learned Building Voice Applications: Best practices for building apps based on AI’s evolving voice-in, voice-out stack

The Voice Stack is improving rapidly. Systems that interact with users via speaking and listening will drive many new applications.

“Responsible AI” written on a wall, with “Safety” crossed out in blue paint.

Technical Insights

The Difference Between “AI Safety” and “Responsible AI”: Talk about “AI safety” obscures an important point; AI isn't inherently unsafe. Instead, let’s talk about “responsible AI.”

At the Artificial Intelligence Action Summit in Paris this week, U.S. Vice President J.D. Vance said, “I’m not here to talk about AI safety.

Letters from Andrew Ng

Tear Down Data Silos!: Many software-as-a-service vendors aim to hold their customers' data in silos. Their customers would do well to open the silos so AI agents can use the data.

Improve Agentic Performance with Evals and Error Analysis, Part 2: Best practices for error analysis in agentic AI development, and how LLMs make them easier

Improve Agentic Performance with Evals and Error Analysis, Part 1: When AI agentic systems go astray, it’s tempting to shortcut evals and error analysis. But these processes cas lead to much faster progress.

Agentic Coding and Agentic Software Testing Go Together: Agentic coding can make mistakes, but agentic testing can find and fix them.

AI Skills Are Redefining What Makes a Great Developer: The job market for software developers requires knowing how to use AI

Agents Running in Parallel Get There Faster: Parallel agents can accelerate AI systems as test-time compute scales up.

Meet The New Breed of GenAI Application Engineers: A new breed of software engineers is building more powerful applications faster than ever, thanks to generative AI. Here’s how to identify them in job interviews.

How to Become a Multilingual Coder: AI makes it easy to code in any programming language — especially if you know just one.

We Iterate on Models. We Can Iterate on Evals, Too: Building automated evals doesn’t need to be a huge investment. Start with a few quick-and-dirty examples and iterate!

The Benefits of Lazy Prompting: You don’t always need to provide context when prompting a large language model. A quick prompt can be enough.

When to Fine-Tune — and When Not To: Many teams that fine-tune their models would be better off prompting or using agentic workflows. Here's how to decide.

Learn the Language of Software: AI won’t kill programming. There has never been a better time to start coding.

Wait Your Turn! Conversation by Voice Versus Text: Text interactions require taking turns, but voices may interrupt or overlap. Here’s how AI is evolving for voice interactions.

What I’ve Learned Building Voice Applications: Best practices for building apps based on AI’s evolving voice-in, voice-out stack

The Difference Between “AI Safety” and “Responsible AI”: Talk about “AI safety” obscures an important point; AI isn't inherently unsafe. Instead, let’s talk about “responsible AI.”

Subscribe to The Batch