lquera

Things I believe: domain experts+ML experts = game-changers; AI as Augmentation Intelligence empowers & elevates; BPR before RPA; kindness matters.

[lquera @Surreabral]

1/2: Something deeply internalized as an economist that is useful transfer learning for being an ML engineer: models are useful but never capture complex reality so rejoice in what can be learned from them but invest in identifying what might be missing/wrong.

[lquera @Surreabral]

If you get annoyed trying to debug a blank-space issue in python, just remember the difference between “It’s worth less” and “It’s worthless”. A blank space can have tremendous value 😃

[Bindu Reddy @bindureddy]

RAG Or Long Context LLMs - A Hamlet-Level Question!

There has been much debate and argument about RAG vs. long-context LLMs. In this study, they find that the long context outperforms RAG.

Our in-house experiments reveal that long context is ineffective when the LLM has to use complex reasoning based on the content in the context to provide the correct response. RAG is the only approach for those use cases.

That being said, even the paper admits that RAG is much more cost-efficient, given that you only need to pass the most relevant parts of the document to the LLM

This paper proposes SELF-ROUTE, a simple yet effective method that routes queries to RAG or LC based on model self-reflection.

SELFROUTE significantly reduces the computation cost while maintaining comparable performance to LC. Their findings provide a guideline for long-context applications of LLMs using RAG and LC.

So, the answer is BOTH for non-complex use cases and RAG for complex ones.

TLDR: RAG isn't going anywhere for a very long time.

[lquera @Surreabral]

When I think of privacy-preserving ML, I think of @openminedorg but today @DeepLearningAI introduced me to Flower Labs https://t.co/NaWhWB6Oxy with https://t.co/jTStrtJ2oT

[lquera @Surreabral]

Say what you want about X/Twitter...the ML Community on this platform is still the most knowledgeable, welcoming, and generous anywhere. Here's yet another example 👇Awesome experts freely sharing their knowledge 🤩👏

[Kyunghyun Cho @kchonyc]

enjoying #ICML2024 ? already finished with llama-3.1 tech report? if so, you must be concerned about the emptiness you'll feel on your flight back home in a couple of days.

do not worry! Wanmo and i have a new textbook on linear algebra for you to read, enjoy and cry on your long flight.

(1/5)

[lquera @Surreabral]

There is a wildly underpriced tool where 81% of the users agreed that it increased their productivity and 62% said that it sped up their learning.
https://t.co/v9oXDRrpbC
If you are NOT implementing AI tools in your development process, why on earth not?

[lquera @Surreabral]

*1, you generally have it as part of your IDE, but you know what I mean
*2, of course, it’s not perfect but it’s a magnitude of improvement over manual searches

[lquera @Surreabral]

Coding-tuned LLMs have made coding so much more fun! Just toss the error code and suspect code into the prompt* and boom, an hour of frustration avoided. Plus the learning benefit of the associated explanation, wow, the whole experience is so productive and empowering*. 😃

[lquera @Surreabral]

It takes a shockingly small amount of code to make carbon sensitivity a part of the process. Business justification for that small effort might be 1) inclusion in ESG funds can raise stock price, 2) entice high quality employees, 3) possible carbon trading options...

[lquera @Surreabral]

Learn how to be an environmentally-aware developer, especially as model builders and deployers, with this new course from Google Cloud and DeepLearningAI: https://t.co/YgQ4nBcma5

Thank you to Google for making Carbon Awareness part of the console and their advocacy 👏

[Harrison Chase @hwchase17]

🤖"In the loop"

Everyone is trying to figure out how to best build agentic applications

I've decided to start sharing some of my thoughts and learnings on how to do so in a new blog series (titled "In the loop")

First one: what is an agent?

https://t.co/KCwYmzSd5Z

[MIT CSAIL @MIT_CSAIL]

This repository on ML operations has free talks, books, papers and more: https://t.co/wxiXkGWyq3

Developed by @visenger

[lquera @Surreabral]

I've noticed new framework, API, etc. docs getting much better. Have we finally learned to value them and reward their creators better, or have LLMs made the process 10x faster/easier/better? Whichever, the end result will massively increase productivity and creativity.

[lquera @Surreabral]

Check out AutoGen from @Microsoft to create your next Agentic AI application.

https://t.co/r0lxu7hjgy

Intuitive, powerful, and very cool 🤩
(Also an outstanding example of effective documentation 👏)

[lquera @Surreabral]

“Be regular and orderly in your life like a bourgeois, so that you may be violent and original in your work.”
— ​Flaubert​
via 5BF newsletter from @tferriss

Maybe replace “violent” w “shocking” but you know what he means 😃
Similar to my fav “From discipline comes freedom”

[Ansong Ni @AnsongNi]

Excited to share our work at @GoogleDeepMind!

We propose Naturalized Execution Tuning (NExT), a self-training method that drastically improves the LLM's ability to reason about code execution, by learning to inspect execution traces and generate chain-of-thought rationales 🧵👇

[Google AI @GoogleAI]

Introducing Google AI Essentials: Learn how to apply #generativeAI to your work from experts at Google, zero experience required → https://t.co/G15LgpTKYV #GrowWithGoogle

[lquera @Surreabral]

Paul Hudson @twostraws taught me everything I know about iOS dev and he has long been such a generous educator for the Swift* community. Now he's created a free Swift website builder at https://t.co/gNbpOAtu95

*language not artist. Although, maybe artist too but IDK 😀

[lquera @Surreabral]

Note, however, that my code for calling the same LLM via API does not include the same markers of civility. So clearly reason two is more important to me. Apparently I'm a polite speaker and rude coder 😉

[lquera @Surreabral]

A friend saw my interaction with a popular LLM UI and teased me for my politeness. My reasoning: 1) using civil discourse may have positive influence on model training datasets, & 2) it's to my own benefit that I don't get into bad conversational habits. We become what we do.

[Sebastian Raschka @rasbt]

A good start to the new week: The highly requested Chapter 5 of my "Build an LLM from Scratch" book is finally available, "Pretraining an LLM on Unlabeled Data".

In the previous chapters, we implemented the data sampling and attention mechanism and coded the LLM architecture from scratch. The core focus of this chapter is to finally implement a training function and pretrain the LLM!

In particular, this chapter covers:
- Computing the training and validation set losses to assess the quality of LLM-generated text during training
- Implementing a training function and pretraining the LLM
- Saving and loading model weights to continue training an LLM
- Loading pretrained weights from OpenAI

Happy reading!