The AI Edition

The AI Edition

Share this post

The AI Edition
The AI Edition
How to win at the NLP game (Part 1)
Copy link
Facebook
Email
Notes
More

How to win at the NLP game (Part 1)

Or how to go from amateur to great!

Jun 16, 2022
∙ Paid
3

Share this post

The AI Edition
The AI Edition
How to win at the NLP game (Part 1)
Copy link
Facebook
Email
Notes
More
Share

Hello everyone, hal here.

Despite another round of wild claims about AI, real progress is being made every day toward building real AI systems used and valued by customers every day.

We are blessed that we live in the digital age, where we can quickly consume, learn and share information with others. The internet is filled with blog posts introducing natural language processing concepts, GitHub repositories with the latest technology, and arXiv papers with bleeding-edge information on language technology.

Make no mistake, there hasn’t been a better time to learn and try out language technology, for free. But no matter how much we consume of that material, it won’t get us to the point of mastery of this technology. Building practical language applications is more than just taking a state-of-the-art approach, blindly applying it to some dataset collected by others, getting 90+% accuracy, and declaring it as a win.

Bronze Medal |  READ TOWARDS DATA SCIENCE POST; GOT 90%+ ACCURACY ON PUBLIC BENCHMARK; DOWNLOADED TRANSFORMERS REPO; CODE RUNNING ON YOUR LOCAL MACHINE; APPLIED NLP | image tagged in bronze medal | made w/ Imgflip meme maker

In fact, once you deploy your solution, you stumble upon the problems that you never expected. Silent failures of your model start to creep in. Customers don’t even get to see the predictions because of the pre-processing error. The calibration of the model is wildly off.

What to fix in this situation?

Debugging and understanding errors in machine learning is much harder compared to traditional software development. The good thing is, as long as you get your core development stack right it would be significantly easier to identify and fix those issues.

Twitter avatar for @gdb
Greg Brockman @gdb
ML systems debugging rewards a willingness to dig across the entire stack, to chase a slightly suspicious signal back to its source, and to derive chains of failures from surprising end results. High cognitive burden but also some of the most exhilarating work upon success.
4:06 PM ∙ May 27, 2022
299Likes27Retweets

Now let us dive into the language application development stack.

This post is for paid subscribers

Already a paid subscriber? Sign in
© 2025 hal
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share

Copy link
Facebook
Email
Notes
More