Chip Huyen

Thanks! I saw your earlier question related to working with AI at FAANG, from the context of being 30ish, and self-taught. I was around 27 when I started learning software engineering and AI. I made a free e-book here (https://emilwallner.gumroad.com/l/no-ml-degree), and I believe most of the key points are still valid, however, when I learned software engineering and ML, ML was a rather small field and tools like chatGPT and claude didn’t exist. Imo, asking about a curriculum is the wrong framing, for me, it was more about how to find resources to focus full-time and being in an environment that increases my motivation. I started learning software engineering at home taking courses, but I procrastinated too much to be effective, maybe I did around 10 hours of effective learning per week. For me, studying C at 42 (https://www.42network.org/42-schools/), a free peer-to-peer school was crucial, and I recommend something similar. It enabled me to focus 70-90 hours a week, and after 6 months I was good enough to get competitive startup job offers. During my time, the FastAI course (https://www.fast.ai/) was the best practical AI course. I'd probably spend a week looking for ambitious projects made by recent autodidacts, and ask them which course they think is best now. And spend max 1-2 months taking the course. As for picking projects and building a portfolio, the advice in my e-book is still valid. An ambitious but realistic timeframe for landing a FAANG job is 3-5 years.…

For SWEs interested in "AI Engineering" (either getting involved in how models work, or building applications on them), there's a critical paradigm shift in that using "AI" requires more of an experimental mindset than software engineering typically does. - I strongly recommend Chip Huyen's books ("Designing Machine Learning Systems" and "AI Engineering") and blog (https://huyenchip.com/blog/). - Andreessen Horowitz' "AI Cannon" is a good reference listicle (https://a16z.com/ai-canon/) - "12 factor agents" (https://github.com/humanlayer/12-factor-agents)

I haven't read these all the way through myself but I've seen enough of them that I'm confident suggesting them: - Prompt Engineering for LLMs by John Berryman and Albert Ziegler: https://www.amazon.com/Prompt-Engineering-LLMs-Model-Based-Application… - AI Engineering by Chip Huyen, which I recommend based on the strength of this extract about "agents": https://huyenchip.com/2025/01/07/agents.html

It's just a fancy word for clamping the new reward value to within some delta of the original value. Otherwise the model ends up "exploiting" outliers that make sense to machines but not to humans. They do the same thing with PPO in RLHF. Great article, if you're interested: https://huyenchip.com/2023/05/02/rlhf.html#3_2_finetuning_using_the_re…

I was a MLE Tech Lead at Snap and laid some of the foundations of the generative AI infra at Snap. I would highly recommend MLE route as a very rewarding career path. This book is a very good introduction to designing Machine Learning Systems for production: https://www.amazon.com/Designing-Machine-Learning-Systems-Production-R… This blog by the same author is highly recommended as an intro into building production grade AI and ML systems: https://huyenchip.com/2023/04/11/llm-engineering.html To summarize answers to your questions: (1) Yes it is wise to do this transition especially at an inflection point in the zeitgeist of the times as now (2) Yes (3) See above for resources on how to get started and reach mastery in the craft of ML Engineering.

You're absolutely right. And I am in academia so I'm not aware of what's being done in industry. I've heard people prefer conformal prediction since it is simple and distribution-free. Also, I recommend this fantastic article by Chip Huyen: https://huyenchip.com/2022/02/07/data-distribution-shifts-and-monitori…

I think the another great paper is RLHF article from Chip Huyen https://huyenchip.com/2023/05/02/rlhf.html

Try following the links in the article. They give much more detailed information. For example, your temperture explination can be found here[1] (Ctrl+F), which is also linked in the article. [1]: https://huyenchip.com/2023/04/11/llm-engineering.html#prompting_vs_fin…

s/thinking about what he's describing/thinking about what she's describing https://huyenchip.com/

Not sure it is true LLMs don't see code or cli commands directly in their training. They go through reinforcement learning and they could easily be trained on a command line. People are paid to give human feedback. See https://huyenchip.com/2023/05/02/rlhf.html

Yes. The learning comes from running tests on the program and ensuring they pass. So running as an agent. Tests and compiler give hard feedback- thats the data outside the model that it learns from. I think modern RLHF schemes have models that train LLMs. LLMs teaching each other isn't new. My knowledge is limited, just based on a read of https://huyenchip.com/2023/05/02/rlhf.html though.

I think this is definitely not true of catastrophic forgetting from finetuning. And with other related types of forgetting from model abliteration there are often extreme increases hallucination. The InstructGPT paper also showed that RLHF made hallucination worse (with more user data rejecting common hallucinations instruction tuning and RLHF may lower specific hallucinations rejected by users though). Some mention of that here: https://huyenchip.com/2023/05/02/rlhf.html#rlhf_and_hallucination

Some youtube channels are good too. https://www.youtube.com/@umarjamilai https://huyenchip.com/blog/

I think this is how RLHF actually works. https://huyenchip.com/2023/05/02/rlhf.html#3_1_reward_model

"RTFM"[0] as they say [0]https://huyenchip.com/ml-interviews-book/contents/3.1.1-base-salary.ht…

A complementary resource is Chip Huyen’s interview book, https://huyenchip.com/ml-interviews-book/ I find her book to be more practical and humble, though perhaps less feature-complete than the book from this post. IMO Kashani and Ivry’s writing feels a little uppity and needlessly offputting — for example, logistic regression is lumped under Kindergarten in the table of contents. Sure it’s fundamental, but implying it’s not useful anymore in our age of deep learning (listed under “Bachelor’s,” of course) is a little myopic/insulting, no? Students who feel pandered to by know-it-alls probably learn worse than students whose skills are being collaboratively built up by enthusiastic mentors. This feeling was particularly loud from the foreward, ToC, and intro; maybe it gets better after…?

This section invites some discussion (says British bloke): "https://huyenchip.com/2024/03/14/ai-oss.html#the_growing_china_open_so…

Hi, I'm Chip, the author of the post. I spent waaaay too much time doing research for this. The AI engineering layer was especially hard because so many tools have similar and/or overlapping features. It was also a lot of pain trying to understand the repos that only have Chinese in the README do. The full list of the repos is published here: https://huyenchip.com/llama-police

> Among those who labeled demonstration data for InstructGPT, ~90% have at least a college degree and more than one-third have a master’s degree. Source: https://huyenchip.com/2023/05/02/rlhf.html#demonstration_data

...because the technology cannot yet live up to the insane hype about and around it... fueled by over-enthusiastic promoters selling snake-oil. There are still quite a few challenges to making LLMs work reliably. For example, see https://huyenchip.com/2023/08/16/llm-research-open-challenges.html. Until AI researchers and practitioners find ways to overcome those challenges -- possibly requiring new breakthroughs -- the technology won't live up to the hype. Eventually, I believe AI will live up to the hype. But no one knows with certainty how long it will take to get there. Shame on anyone who pretends otherwise to sell snake-oil!

Reasonable argument, I like it, however would add a bit of salt: I've seen several times so far that people maintaining things are actually doing a much better job, while those who were granted honor of doing a new thing because hey they are "pros". And once they did the prototype, they are _considered_ even more "pros", because hey, they wrote it. The people who rethought and rewrote the thing later were real makers, but didn't get the appreciation of "pros". The worst thing is that "pros" status keeps working for itself, even while delivering worse job. I wouldn't have a problem if the system could fix itself over time. Kinda related: "Confession of a so-called AI expert " https://huyenchip.com/2017/07/28/confession.html

There's also this exhaustive post from one and only Chip Huyen: https://huyenchip.com/2023/05/02/rlhf.html

There is also the much larger confusion arising of whether it's an ML compiler in the sense of being a specialized compiler for ML code (compiles a description of a network to efficient ML-specific low-level IR to enable further optimization and code generation), or whether it's an ML compiler, a compiler (or part of it) that uses ML. The first meaning is extremly common, the top 2 results for "ML Compiler" on google returns[1][2], both using it in the first sense. It's not that ML or ML-like techniques in compiler writing is new but the first sense is definitely at least as popular an interpretation. [1] https://huyenchip.com/2021/09/07/a-friendly-introduction-to-machine-le… [2] https://petewarden.com/2021/12/24/why-are-ml-compilers-so-hard/

Read these three articles: - https://erikbern.com/2018/05/02/interviewing-is-a-noisy-prediction-pro… - https://erikbern.com/2021/07/07/the-data-team-a-short-story.html - https://huyenchip.com/ml-interviews-book/

Plenty of women in tech did read it, and thought it was pretty sexist: https://huyenchip.com/2019/03/11/silicon-valley-misogyny.html https://openheroines.org/tell-me-about-your-books-and-i-will-tell-you-…

Here’s a fun quote from his book, via Chip Huyen’s blog (https://huyenchip.com/2019/03/11/silicon-valley-misogyny.html) > “Most women in the Bay Area are soft and weak, cosseted and naive despite their claims of worldliness, and generally full of shit. They have their self-regarding entitlement feminism, and ceaselessly vaunt their independence, but the reality is, come the epidemic plague or foreign invasion, they’d become precisely the sort of useless baggage you’d trade for a box of shotgun shells or a jerry can of diesel.” I’m also not sure if his stance on privacy will fit in well at Apple: > “Most people don’t care about privacy. Media elites care about it, underemployed Eurocrats care about it. And the entire privacy-industrial complex — there’s an entire set of very loud voices who are constantly beating the drum and building media careers around this.”

This guy sounds like the worst culture fit at Apple. https://huyenchip.com/2019/03/11/silicon-valley-misogyny.html

You can give this article by Chip Huyen a read. Mayhaps you will find a niche for a solo or small dev team. Though it is focused on MLOps if that makes a different for the type of niche you're looking for. https://huyenchip.com/2020/06/22/mlops.html

Is she having doubts brought on by the recent Google Manifesto perhaps: http://huyenchip.com/2017/08/09/sexism-in-silicon-valley.html posted earlier.

This is good also https://huyenchip.com/2023/05/02/rlhf.html