Letters from Andrew Ng

Personal messages to the AI community

Welcoming Diverse Approaches Keeps Machine Learning Strong: What technology counts as an “agent”? Instead of arguing, let's consider a spectrum along which various technologies are “agentic.”

One reason for machine learning’s success is that our field welcomes a wide range of work.

We Need Better Evals for LLM Applications: It’s hard to evaluate AI applications built on large language models. Better evals would accelerate progress.

A barrier to faster progress in generative AI is evaluations (evals), particularly of custom AI applications that generate free-form text.

Technical Insights

Project Idea — A Car for Dinosaurs: AI projects don’t need to have a meaningful deliverable. Lower the bar and do something creative.

A good way to get started in AI is to start with coursework, which gives a systematic way to gain knowledge, and then to work on projects.

Technical Insights

Building Models That Learn From Themselves: AI developers are hungry for more high-quality training data. The combination of agentic workflows and inexpensive token generation could supply it.

Inexpensive token generation and agentic workflows for large language models (LLMs) open up intriguing new possibilities for training LLMs on synthetic data. Pretraining an LLM

Technical Insights

Why We Need More Compute for Inference: Today, large language models produce output primarily for humans. But agentic workflows produce lots of output for the models themselves — and that will require much more compute for AI inference.

Much has been said about many companies’ desire for more compute (as well as data) to train larger foundation models.

Proposed ChatDev architecture, illustrated.

Technical Insights

Agentic Design Patterns Part 5, Multi-Agent Collaboration: Prompting an LLM to play different roles for different parts of a complex task summons a team of AI agents that can do the job more effectively.

Multi-agent collaboration is the last of the four key AI agentic design patterns that I’ve described in recent letters.

Technical Insights

Agentic Design Patterns Part 4, Planning: Large language models can drive powerful agents to execute complex tasks if you ask them to plan the steps before they act.

Planning is a key agentic AI design pattern in which we use a large language model (LLM) to autonomously decide on what sequence of steps to execute to accomplish a larger task.

Technical Insights

Agentic Design Patterns Part 3, Tool Use: How large language models can act as agents by taking advantage of external tools for search, code execution, productivity, ad infinitum

Tool use, in which an LLM is given functions it can request to call for gathering information, taking action, or manipulating data, is a key design pattern of AI agentic workflows.

Technical Insights

Agentic Design Patterns Part 2, Reflection: Large language models can become more effective agents by reflecting on their own behavior.

Last week, I described four design patterns for AI agentic workflows that I believe will drive significant progress this year: Reflection, Tool use, Planning and Multi-agent collaboration.

Technical Insights

Agentic Design Patterns Part 1: Four AI agent strategies that improve GPT-4 and GPT-3.5 performance

I think AI agent workflows will drive massive AI progress this year — perhaps even more than the next generation of foundation models. This is an important trend, and I urge everyone who works in AI to pay attention to it.

Technical Insights

Life in Low Data Gravity: With generative AI, data is bound less tightly to the cloud provider where it’s stored. This has big implications for developers, CIOs, and cloud platforms.

I’ve noticed a trend in how generative AI applications are built that might affect both big companies and developers: The gravity of data is decreasing.

Technical Insights

The Dawning Age of Agents: LLM-based agents that act autonomously are making rapid progress. Here's what we have to look forward to.

Progress on LLM-based agents that can autonomously plan out and execute sequences of actions has been rapid, and I continue to see month-over-month improvements.

Illustration of a Python inside a cardboard box

Technical Insights

The Python Package Problem: Python packages can give your software superpowers, but managing them is a barrier to AI development.

I think the complexity of Python package management holds down AI application development more than is widely appreciated. AI faces multiple bottlenecks — we need more GPUs, better algorithms, cleaner data in large quantities.

Technical Insights

How to Think About the Privacy of Cloud-Based AI: How private is your data on cloud-based AI platforms? Here's a framework for evaluating risks.

The rise of cloud-hosted AI software has brought much discussion about the privacy implications of using it. But I find that users, including both consumers and developers building on such software

Technical Insights

Outstanding Research Without Massive Compute: Researchers at Stanford and Chan Zuckerberg Biohub Network dramatically simplified a key algorithm for training large language models.

It is only rarely that, after reading a research paper, I feel like giving the authors a standing ovation. But I felt that way after finishing Direct Preference Optimization (DPO) by...

Letters from Andrew Ng

Welcoming Diverse Approaches Keeps Machine Learning Strong: What technology counts as an “agent”? Instead of arguing, let's consider a spectrum along which various technologies are “agentic.”

We Need Better Evals for LLM Applications: It’s hard to evaluate AI applications built on large language models. Better evals would accelerate progress.

Project Idea — A Car for Dinosaurs: AI projects don’t need to have a meaningful deliverable. Lower the bar and do something creative.

Building Models That Learn From Themselves: AI developers are hungry for more high-quality training data. The combination of agentic workflows and inexpensive token generation could supply it.

Why We Need More Compute for Inference: Today, large language models produce output primarily for humans. But agentic workflows produce lots of output for the models themselves — and that will require much more compute for AI inference.

Agentic Design Patterns Part 5, Multi-Agent Collaboration: Prompting an LLM to play different roles for different parts of a complex task summons a team of AI agents that can do the job more effectively.

Agentic Design Patterns Part 4, Planning: Large language models can drive powerful agents to execute complex tasks if you ask them to plan the steps before they act.

Agentic Design Patterns Part 3, Tool Use: How large language models can act as agents by taking advantage of external tools for search, code execution, productivity, ad infinitum

Agentic Design Patterns Part 2, Reflection: Large language models can become more effective agents by reflecting on their own behavior.

Agentic Design Patterns Part 1: Four AI agent strategies that improve GPT-4 and GPT-3.5 performance

Life in Low Data Gravity: With generative AI, data is bound less tightly to the cloud provider where it’s stored. This has big implications for developers, CIOs, and cloud platforms.

The Dawning Age of Agents: LLM-based agents that act autonomously are making rapid progress. Here's what we have to look forward to.

The Python Package Problem: Python packages can give your software superpowers, but managing them is a barrier to AI development.

How to Think About the Privacy of Cloud-Based AI: How private is your data on cloud-based AI platforms? Here's a framework for evaluating risks.

Outstanding Research Without Massive Compute: Researchers at Stanford and Chan Zuckerberg Biohub Network dramatically simplified a key algorithm for training large language models.

Subscribe to The Batch