Chip Huyen
Engineer and writer on machine-learning systems; author of Designing Machine Learning Systems and AI Engineering. "I work to bring AI into production. I write about AI system design."
Chip Huyen did not write this page. What is this?
It collects the places they publish and what they have said there, each linked to the source. They have no account here. Is this you? Claim it, correct it, or ask us to remove it from ppll.
Where they publish
Blog huyenchip.com Posts on getting machine learning into production. Has a feed.
Long posts working through one system-design question at a time — evaluation, agents, inference cost, what an ML platform is actually made of.
Recent
- Common pitfalls when building generative AI applications 16 Jan 2025 As we’re still in the early days of building applications with foundation models, it’s normal to make mistakes. This is a quick note with examples of some of the most common pitfalls that I’ve seen, both from public cas…
- Agents 7 Jan 2025 Intelligent agents are considered by many to be the ultimate goal of AI. The classic book by Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach (Prentice Hall, 1995), defines the field of AI res…
- Building A Generative AI Platform 25 Jul 2024 After studying how companies deploy generative AI applications, I noticed many similarities in their platforms. This post outlines the common components of a generative AI platform, what they do, and how they are implem…
Show 7 more
- Measuring personal growth 17 Apr 2024 My founder friends constantly think about growth. They think about how to measure their business growth and how to get to the next order of magnitude scale. If they’re making $1M ARR today, they think about how to get t…
- What I learned from looking at 900 most popular open source AI tools 14 Mar 2024 [Hacker News discussion, LinkedIn discussion, Twitter thread] Update (Feb 2026): The full list of open source AI repos is hosted at Good AI List, updated daily. It’s balooned to 15K repos, and you can submit missing rep…
- Predictive Human Preference: From Model Ranking to Model Routing 28 Feb 2024 A challenge of building AI applications is choosing which model to use. What if we don’t have to? What if we can predict the best model for any prompt? Predictive human preference aims to predict which model users might…
- Generation configurations: temperature, top-k, top-p, and test time compute 16 Jan 2024 ML models are probabilistic. Imagine that you want to know what’s the best cuisine in the world. If you ask someone this question twice, a minute apart, their answers both times should be the same. If you ask a model th…
- Multimodality and Large Multimodal Models (LMMs) 10 Oct 2023 For a long time, each ML model operated in one data mode – text (translation, language modeling), image (object detection, image classification), or audio (speech recognition). However, natural intelligence is not limit…
- Open challenges in LLM research 16 Aug 2023 [LinkedIn discussion, Twitter thread] Never before in my life had I seen so many smart people working on the same goal: making LLMs better. After talking to many people working in both industry and academia, I noticed t…
- Generative AI Strategy 7 Jun 2023 I had a lot of fun preparing the talk: “Leadership needs us to do generative AI. What do we do?” for Fully Connected. The idea for the talk came from many conversations I’ve had recently with friends who need to figure…
Link verified 20 Sept 2026. Recent items update automatically from the channel.
GitHub @chiphuyen Code, and the notes that became the books.
Recent
- chiphuyen — I'm a writer and computer scientist. I grew up chasing grasshoppers in a small rice-farming village in Vietnam. I spend a lot of time with chickens and alpacas. 🔭 I'm getting back into b… 24 Jun 2026 CommitsFix link in README.md for GoodAIList · Fix typo in README · Revise personal details and project descriptions in README
- dmls-book — Summaries and resources for Designing Machine Learning Systems book (Chip Huyen, O'Reilly 2022) 9 Jun 2026 CommitsAdd translations
- sniffly — Claude Code dashboard with usage stats, error analysis, and sharable feature 8 Aug 2025 CommitsAdd MIT License file (#18) · Release 0.1.5 (#16) · Fix error analysis (#15)
Show 3 more
- ml-interviews-book — https://huyenchip.com/ml-interviews-book/ 21 Mar 2025 CommitsAdd book cover
- metaflow-transformers-tutorials — Metaflow tutorials for ODSC West 2021 16 Nov 2021 CommitsUpdate instructions · Clean up some code · Clean up the repo
- python-is-cool — Cool Python features for machine learning that I used to be too afraid to use. Will be updated as I have more time / learn more. 27 Dec 2019 Commitsdecorators, timeit, and cache
Link verified 20 Sept 2026. Recent items update automatically from the channel.
Beliefs
Korrents What they believe 37 beliefs — each backed by an exact quote.
Each is a — compiled by korrents.com, not by them: the one-line wordings are korrents', the quotes are theirs.
Recent
Because anyone can now rebuild whatever you make, the ordinary incentive to build software has been knocked out from under us.
I feel like now I can build anything I want. But at the same time, anyone can build anything I want. So, what is the incentive structure for me to do anything?
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
Anything a person can describe, AI can now produce — software has reached the point image generation reached with style.
if you can describe a style, AI can generate it. And the same with software. If you can describe a software, then AI can build it for you.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
Every product put into the world removes the need for imagination, because it becomes something anyone can point AI at and copy.
And the more you build, right? The more you put things out there, you remove the need for imagination. You can say, "Okay, now I like that website. Do that for me."
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
Show 34 more
The reason to keep building is that building is practice at problem-solving, and the supply of problems will never run out.
The product you build is just solving a problem. So, is that like is the more you do that, the better you become at problem solving. And one thing I do believe that it will never change that there will always be problems to solve.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
Problems follow a long tail: AI will keep getting better at the common ones, and the edge cases will never go away.
So, I think of them as the top of the long long tail long tail problem. Like, something very common issues that a lot of people experience. AI will get really good at it. And over time AI will cover more more edge cases, but the edge cases will never go away.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
The problems worth a small builder's time are the ones too small for a big AI company to be motivated to solve.
Like if it's a big problem, like everyone can see, then all these big companies will get into it. But whereas it's like there are a lot of problems that's like smaller, then maybe OpenAI won't be motivated to solve it, but maybe I can like all a lot of people can.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
Human preference is not an equation you can package by asking which of two answers people prefer; it is personal, cultural, geographic and generational.
I don't think that human preference is just an equation that people can just like package nicely and like, "Hey, ask people like which which of these two answers people would prefer?" It's very very personal, very culturally dependent, geographically dependent, age dependent.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
Which interface an AI product should have is a cultural fact, not a technical one — Vietnamese companies ship voice bots before chatbots because people are on motorbikes.
But in Vietnam and also in a lot of other Asian countries, people are on the move all the time. Like people are on the motorbike all the time. So, they actually really don't like typing. So, the voice like a lot of the companies in Vietnam actually deploy voice bots before they do uh do chatbot.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
We are still retrofitting the tools we already had onto workflows we have only imagined, instead of designing for how people will actually work with AI.
So, I I think there's a lot of things we still try to imagine. What would be an AI-driven world look like, right? I think people are trying to like retrofit what already exists to fit what they think is a new workflow.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
Line-by-line code review has become an outdated ritual: senior engineers should be giving feedback on how their team instructs AI, not on the code AI wrote.
So, I think that the whole workflow of like reviewing code is very outdated. Like, I don't think the I think that the senior member, instead of like giving feedback on the code, they should be giving feedback on like how you give instruction to AI to produce better.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
The frightening part of giving agents more access is not bad code but irreversible action, so the guardrails to build next are guardrails on reversibility.
You cannot just reverse that, right? It's just not working. So, I do think that we need to build out the whole guardrails for the reversibility of actions because that actually where things get really really scary.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
AI web search is imitating a human habit for no reason — revisiting the same pages for snippets instead of taking the whole page once.
And I feel like it's a very human way of doing web search, right? Because we enter things in the Google and we see all the like citations, the quotations part. But like why would we limit AI to that? If AI already visit a page, why don't just pull the entire page out?
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
People no longer read books the way books assume they will — they jump around, whatever they tell the author.
Like I I write books and as much I wish that books as a format will be like will survive. I do think that people don't read books the same way anymore.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
The hard question is no longer how to build something but what to build, because anything already described can be replicated.
And the question nowadays is less about like how to build because if you can describe the problem in the solution you want, usually AI can do it. Maybe not today, but maybe like two or three years from now on they can do a lot of those. The question is like what to build because we talk about like yes, if something exists, right? AI can replicate it.
Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit Said 29 Mar 2026
As AI products converge on the same underlying models, product design becomes the real differentiator, not the model itself.
Because everyone uses the same models nowadays, the AI components of AI products are similar, and the differentiation is product.
Common pitfalls when building generative AI applications Said 16 Jan 2025
Building a working AI demo is far easier than turning it into a reliable production product.
This is perhaps one of the first painful lessons anyone who has built an AI product quickly learns. It's easy to build a demo, but hard to build a product.
Common pitfalls when building generative AI applications Said 16 Jan 2025
AI judges used to evaluate AI outputs need ongoing validation and iteration, just like any other AI system.
AI judges must be evaluated and iterated over time, just like all other AI applications.
Common pitfalls when building generative AI applications Said 16 Jan 2025
Planning is fundamentally a search problem: explore paths, predict outcomes, and pick the best one.
Planning, at its core, is a search problem. You search among different paths towards the goal, predict the outcome (reward) of each path, and pick the path with the most promising outcome.
Agents Said 7 Jan 2025
Most real-world AI agent systems are actually multi-agent systems made of multiple components.
Because most agentic workflows are sufficiently complex to involve multiple components, most agents are multi-agent.
Agents Said 7 Jan 2025
A self-driving car is more trustworthy than a random human driver.
Personally, I would trust a self-driving car more than the average stranger to give me a lift.
Agents Said 7 Jan 2025
Preparing context for a language model plays the same role that feature engineering played for classical machine learning models.
Context construction for foundation models is equivalent to feature engineering for classical ML models.
Building A Generative AI Platform Said 25 Jul 2024
In-context learning is a form of continual learning for language models.
In-context learning, learning from the context, is a form of continual learning.
Building A Generative AI Platform Said 25 Jul 2024
The probabilistic nature of AI models makes their outputs inherently unreliable.
AI models are probabilistic, making their outputs unreliable.
Building A Generative AI Platform Said 25 Jul 2024
A person's life tends to change character in cycles of roughly three to six years.
I have this theory that life has a circadian rhythm. Every 3-6 years, you become a different person.
Measuring personal growth Said 17 Apr 2024
Growing older tends to expand rather than shrink the number of achievable dreams and options.
As I grow older, I have more dreams. I now know many things that I didn't know before, and I have access to more resources than I ever did.
Measuring personal growth Said 17 Apr 2024
Under uncertainty, the best decision is the one that preserves the most future options.
In the face of uncertainty, I lean towards the decision that would give me the most future options.
Measuring personal growth Said 17 Apr 2024
The lower a layer sits in the software stack, the harder it is for an individual developer to build in it.
the lower we go in the stack, the harder it is for individuals to build.
What I learned from looking at 900 most popular open source AI tools Said 14 Mar 2024
AI will enable one-person companies to reach very high valuations.
Several people have speculated that we'll see many very valuable one-person companies (see Sam Altman's interview and Reddit discussion). I think they might be right.
What I learned from looking at 900 most popular open source AI tools Said 14 Mar 2024
Routing prompts to different AI models can improve response quality while lowering cost and latency.
Model routing has the potential to increase response quality while reducing costs and latency.
Predictive Human Preference: From Model Ranking to Model Routing Said 28 Feb 2024
As more AI models are developed with varying capabilities and costs, routing between them carries real economic value.
With more and more models being developed, each with different capabilities and a cost structure, model routing has clear economic values.
Predictive Human Preference: From Model Ranking to Model Routing Said 28 Feb 2024
Weaker AI models match stronger ones on simple prompts, but strong models are increasingly preferred as prompts get more challenging.
One pattern is that for simple prompts, weak models can do (nearly) as well as strong models. For more challenging prompts, however, users are much more likely to prefer stronger models.
Predictive Human Preference: From Model Ranking to Model Routing Said 28 Feb 2024
The probabilistic nature of AI models is what makes them well suited to creative tasks
This probabilistic nature makes AI great for creative tasks.
Generation configurations: temperature, top-k, top-p, and test time compute Said 16 Jan 2024
When a model behaves unreliably, the better fix is to switch models rather than repeatedly sample it to work around the inconsistency
The optimal thing to do with a fickle model, however, is to swap it out for another.
Generation configurations: temperature, top-k, top-p, and test time compute Said 16 Jan 2024
As models get better at following instructions, techniques like finetuning and constrained sampling for structured outputs will become less necessary
I suspect that in the future, it'll be easier to get models to output exactly what we need with minimal prompting, and these techniques will become less important.
Generation configurations: temperature, top-k, top-p, and test time compute Said 16 Jan 2024
Text is a more powerful output modality for a model than images, because text supports many downstream tasks
Text is a much more powerful mode for model outputs. A model that can generate images can only be used for image generation, whereas a model that can generate text can be used for many tasks: summarization, translation, reasoning, question answering, etc.
Multimodality and Large Multimodal Models (LMMs) Said 10 Oct 2023
Image is the most versatile input modality for a model, since it can represent text, tabular data, and audio
Image is perhaps the most versatile format for model inputs, as it can be used to represent text, tabular data, audio, and to some extent, videos. There's also so much more visual data than text data.
Multimodality and Large Multimodal Models (LMMs) Said 10 Oct 2023
A model that could learn effectively from raw bitstrings or bytestrings would be extremely powerful, able to learn from any data modality
A model that can effectively learn from bitstrings or bytestrings will be very powerful, and it can learn from any data mode.
Multimodality and Large Multimodal Models (LMMs) Said 10 Oct 2023
What is a korrent?
A korrent is a belief a person has stated in their own words: one sentence stating the claim, backed by a quote and a source, kept at korrents.com.
Under a name here, the quoted block is what they actually said. The korrent beneath it is the claim those words support, in korrents' wording — tap it to see the record, its source, and who else holds it.
Nobody here wrote their own korrents. They are compiled from public statements, and a person can change their mind, which is recorded too.
Feed
As its own page →Hiding
24 June
9 June
29 March
From one piece Chip Huyen: Building when it feels like there's nothing left to build - The Pragmatic Summit 14 beliefs · youtube.com
-
Their words
And the question nowadays is less about like how to build because if you can describe the problem in the solution you want, usually AI can do it. Maybe not today, but maybe like two or three years from now on they can do a lot of those. The question is like what to build because we talk about like yes, if something exists, right? AI can replicate it.
-
korrents.com
People no longer read books the way books assume they will — they jump around, whatever they tell the author.Their words
Like I I write books and as much I wish that books as a format will be like will survive. I do think that people don't read books the same way anymore.
-
Their words
And I feel like it's a very human way of doing web search, right? Because we enter things in the Google and we see all the like citations, the quotations part. But like why would we limit AI to that? If AI already visit a page, why don't just pull the entire page out?
+ 11 more
-
Their words
You cannot just reverse that, right? It's just not working. So, I do think that we need to build out the whole guardrails for the reversibility of actions because that actually where things get really really scary.
-
Their words
So, I think that the whole workflow of like reviewing code is very outdated. Like, I don't think the I think that the senior member, instead of like giving feedback on the code, they should be giving feedback on like how you give instruction to AI to produce better.
-
Their words
So, I I think there's a lot of things we still try to imagine. What would be an AI-driven world look like, right? I think people are trying to like retrofit what already exists to fit what they think is a new workflow.
-
Their words
But in Vietnam and also in a lot of other Asian countries, people are on the move all the time. Like people are on the motorbike all the time. So, they actually really don't like typing. So, the voice like a lot of the companies in Vietnam actually deploy voice bots before they do uh do chatbot.
-
Their words
I don't think that human preference is just an equation that people can just like package nicely and like, "Hey, ask people like which which of these two answers people would prefer?" It's very very personal, very culturally dependent, geographically dependent, age dependent.
-
korrents.com
The problems worth a small builder's time are the ones too small for a big AI company to be motivated to solve.Their words
Like if it's a big problem, like everyone can see, then all these big companies will get into it. But whereas it's like there are a lot of problems that's like smaller, then maybe OpenAI won't be motivated to solve it, but maybe I can like all a lot of people can.
-
korrents.com
Problems follow a long tail: AI will keep getting better at the common ones, and the edge cases will never go away.Their words
So, I think of them as the top of the long long tail long tail problem. Like, something very common issues that a lot of people experience. AI will get really good at it. And over time AI will cover more more edge cases, but the edge cases will never go away.
-
Their words
The product you build is just solving a problem. So, is that like is the more you do that, the better you become at problem solving. And one thing I do believe that it will never change that there will always be problems to solve.
-
Their words
And the more you build, right? The more you put things out there, you remove the need for imagination. You can say, "Okay, now I like that website. Do that for me."
-
Their words
if you can describe a style, AI can generate it. And the same with software. If you can describe a software, then AI can build it for you.
-
Their words
I feel like now I can build anything I want. But at the same time, anyone can build anything I want. So, what is the incentive structure for me to do anything?
8 August 2025
21 March 2025
16 January 2025
From one piece Common pitfalls when building generative AI applications 3 beliefs · huyenchip.com
-
korrents.com
AI judges used to evaluate AI outputs need ongoing validation and iteration, just like any other AI system.Their words
AI judges must be evaluated and iterated over time, just like all other AI applications.
-
korrents.com
Building a working AI demo is far easier than turning it into a reliable production product.Their words
This is perhaps one of the first painful lessons anyone who has built an AI product quickly learns. It's easy to build a demo, but hard to build a product.
-
Their words
Because everyone uses the same models nowadays, the AI components of AI products are similar, and the differentiation is product.
7 January 2025
From one piece Agents 3 beliefs · huyenchip.com
-
Their words
Personally, I would trust a self-driving car more than the average stranger to give me a lift.
-
korrents.com
Most real-world AI agent systems are actually multi-agent systems made of multiple components.Their words
Because most agentic workflows are sufficiently complex to involve multiple components, most agents are multi-agent.
-
korrents.com
Planning is fundamentally a search problem: explore paths, predict outcomes, and pick the best one.Their words
Planning, at its core, is a search problem. You search among different paths towards the goal, predict the outcome (reward) of each path, and pick the path with the most promising outcome.
25 July 2024
From one piece Building A Generative AI Platform 3 beliefs · huyenchip.com
-
Their words
AI models are probabilistic, making their outputs unreliable.
-
Their words
In-context learning, learning from the context, is a form of continual learning.
-
Their words
Context construction for foundation models is equivalent to feature engineering for classical ML models.
17 April 2024
From one piece Measuring personal growth 3 beliefs · huyenchip.com
-
korrents.com
Under uncertainty, the best decision is the one that preserves the most future options.Their words
In the face of uncertainty, I lean towards the decision that would give me the most future options.
-
korrents.com
Growing older tends to expand rather than shrink the number of achievable dreams and options.Their words
As I grow older, I have more dreams. I now know many things that I didn't know before, and I have access to more resources than I ever did.
-
Their words
I have this theory that life has a circadian rhythm. Every 3-6 years, you become a different person.
14 March 2024
From one piece What I learned from looking at 900 most popular open source AI tools 2 beliefs · huyenchip.com
-
Their words
Several people have speculated that we'll see many very valuable one-person companies (see Sam Altman's interview and Reddit discussion). I think they might be right.
-
korrents.com
The lower a layer sits in the software stack, the harder it is for an individual developer to build in it.Their words
the lower we go in the stack, the harder it is for individuals to build.
28 February 2024
From one piece Predictive Human Preference: From Model Ranking to Model Routing 3 beliefs · huyenchip.com
-
Their words
One pattern is that for simple prompts, weak models can do (nearly) as well as strong models. For more challenging prompts, however, users are much more likely to prefer stronger models.
-
Their words
With more and more models being developed, each with different capabilities and a cost structure, model routing has clear economic values.
-
korrents.com
Routing prompts to different AI models can improve response quality while lowering cost and latency.Their words
Model routing has the potential to increase response quality while reducing costs and latency.
16 January 2024
From one piece Generation configurations: temperature, top-k, top-p, and test time compute 3 beliefs · huyenchip.com
-
Their words
I suspect that in the future, it'll be easier to get models to output exactly what we need with minimal prompting, and these techniques will become less important.
-
Their words
The optimal thing to do with a fickle model, however, is to swap it out for another.
-
Their words
This probabilistic nature makes AI great for creative tasks.
10 October 2023
From one piece Multimodality and Large Multimodal Models (LMMs) 3 beliefs · huyenchip.com
-
Their words
A model that can effectively learn from bitstrings or bytestrings will be very powerful, and it can learn from any data mode.
Multimodality and Large Multimodal Models (LMMs)huyenchip.com
-
korrents.com
Image is the most versatile input modality for a model, since it can represent text, tabular data, and audioTheir words
Image is perhaps the most versatile format for model inputs, as it can be used to represent text, tabular data, audio, and to some extent, videos. There's also so much more visual data than text data.
Multimodality and Large Multimodal Models (LMMs)huyenchip.com
-
korrents.com
Text is a more powerful output modality for a model than images, because text supports many downstream tasksTheir words
Text is a much more powerful mode for model outputs. A model that can generate images can only be used for image generation, whereas a model that can generate text can be used for many tasks: summarization, translation, reasoning, question answering, etc.
Multimodality and Large Multimodal Models (LMMs)huyenchip.com
16 August 2023
7 June 2023
16 November 2021
27 December 2019
Nothing matches.
About the English under a post
Some people here publish in a language other than English. Where they do, this site shows a machine translation beneath the post, in this typeface — the site's own, not theirs.
The post itself is never changed, moved or hidden: what is set in the serif above is exactly what the person published, and it is what to quote them on. A translation can be wrong in ways that matter, especially about tone.
Only the post's own words are translated. A quoted post, a linked article and a belief on korrents.com are left in their original language.