LLM in Material Production

Thanksgiving has been and gone, and hopefully you enjoyed yours if you celebrate.

But it can be a stressful time with lots of organizing and planning, and we’re yet to reach peak holiday season. If you find yourself starting to get stressed, remember to look after yourself. Part of that might just be to go with the flow.

And speaking of going with the flow, we’d like to welcome our newest sponsors, LatticeFlow!

So, while organizing festivities might stress you out, at least you’ll be relaxed finding and fixing data and model issues going with the flow!

LLM in Material Production // Pierre Salvy // MLOps Podcast #193

During my mis-spent youth there may have been times I’d go for the cheapest alcohol, which often felt like I was drinking jet fuel.

According to Pierre, it wasn’t that far off.

This may not be one of the most technical episodes, but man, it was one of the most interesting.

We still dig into the technical side of Pierre’s work, like servicing researchers rather than data scientists, and working with different constraints like the huge expense to do an experiment and why no data leakage is so important.

But it’s when Pierre talks about what they’re doing with the tech that blew my mind. They’ve commercialized a biotech product which is basically vegan human collagen, and talks about the advances in other areas too, like reproducing cheese molecule by molecule, and how almost everything seems to be one protein away from being rocket fuel! There’s also a great explanation about how LLMs and proteins are similar. Oh, and having an API to order DNA!

He also shares how they want to continue to make a positive real-world impact with what they do. Yeah, the initial product it cosmetic, but it will have implications for medical use and how regulations make it a better initial step than other areas. Through it all the respect for his team and the passion for what they do shines through.

Be sure to relax and listen to it with a drink of your choice. Just make sure it isn’t jet fuel.

Video || Spotify || Apple

With Weights & Biases’ newest course on Training and Fine-tuning LLMs you’ll learn how to customise your own LLM for your particular use case. In this free, on-demand course you’ll also learn advanced fine-tuning techniques such as LoRA and RLHF, understand LLM architectures, and gain practical skills in text data selection and processing.

This course, led by world-class instructors; Jonathan Frankle (MosaicML), Weiwei Yang (Microsoft), and Mark Saroufim (Meta) as well as LLM experts from Weights & Biases, offers in-depth knowledge and hands-on experience to maximise performance of your LLM. In addition, upon completion you'll receive a certificate showcasing your new abilities in this cutting-edge domain.

Whether you're refining existing skills or exploring new horizons with fine-tuned LLMs, this course will help you to stay ahead in the rapidly evolving AI landscape.

Sign up now to be part of this transformative learning experience!

Charles Frye // Sahar Mor // Sarah Guo // Shyamala Prayaga // Willem Pienaar

As we approach December it’s often a time for reflecting on the year.
A huge part for us was the LLMs in Production conference in October, with 2.8k attendees watching 60 speakers over 44 sessions. And with all that going on, it’s easy to miss some of it, so I’m throwing a spotlight on one of the panel sessions here.

This panel was all about LLMs and commercially viable products. Challenges like latency, user experience, evaluation, prompt engineering and AI integration difficulties are all discussed.

Plus, the game-changing potential of multimodality as a way to improve accessibility for people with disabilities. This panel on production produced the goods!

Watch it here

Mini-summit meetup on 29 Nov brought to you by Union.

Ben’s back with an other bumper haul of speakers!

This time he’s joined by David Espejo, Open Source Developer Advocate at Union, Fabio Grätz, Senior Software Engineer at Recogni, and Arno Hollosi, CTO of Blackshark.ai.

So, we know big data can come with big headaches. Scaling, monitoring, managing compute clusters and more.

Well, this mini-summit should help as it includes talks about open-source platforms changing AI product development, accelerating ML experiments and insights on running AI models at a global scale.

Be sure to register here.

LLM Finetuning // Jonathan Whitaker AI Researcher at Data Science Castnet, Boris Dayma CEO of Craiyon, Thomas Capelle ML Engineer at Weights & Biases, and Robbie McCorkell, Founding Engineer of Leap Labs // MLOps Mini Summit #3

It was a good job we extended this mini summit over the usual one hour!

Such a packed session on fine-tuning, including:

Thomas from Weights & Biases sharing some tips and tricks, including an experiment to build a tool for better AI assistants using LLMs.
Boris from Craiyon giving his insights and techniques to improve model training and prevent overfitting.
Robbie from Leap Labs talking about the importance of inspecting tokens in data for debugging.
Jonathan from Castnet presenting a generative process for identifying the key features learned by a model.

All presided over by the fantastic Ben Epstein.

Watch it here

💡Job of the week

Lead MLOps Engineer // Klaviyo (Boston, US)

Klaviyo operates a real-time data analytics platform coded primarily in Python that is built for massive scale and hosted on AWS. We love tackling tough engineering problems and look for employees who specialize in certain areas but are passionate about building, owning & scaling features end to end from scratch and breaking through any obstacle or technical challenge in their way.

This Lead MLE is responsible for technical leadership on the team of Machine Learning Engineers who build and maintain the services that enable and accelerate Data Science and Machine Learning at Klaviyo, including tooling for training, testing, serving, and monitoring models.

What we're looking for:

Minimum 5 years experience in ML Ops
Experience putting ML models with a variety of applications and usage into production
Proficient with Python, AWS, Terraform, Kubernetes
Excited to read the code, understand the product area, and give applicable technical mentorship.

Competitive Differentiation for Foundation Models in the LLM Space

It's getting pretty packed in the ML world, and if you want to stand out, you often have to be different.

As the field of machine learning foundation models grows, companies are finding new ways to make their models unique . This hidden gem identifies three main ways in which these models are becoming different from each other:

Compute Performance
Safety and Alignment
Accuracy and Retrieval Augmented Generation (RAG)

It also shares other characteristics that can set models apart, such as customization and flexibility for specific types of applications and ease of development.

Be sure to read it so you can stand out from the crowd!

Many thanks to Alex Irina Sandu for the contribution.