Infrastructure, intuition, challenges and startups!

Happy New Year! (I’m never sure when the cut-off for that is…February?)

I hope you had a great break and are ready and raring to take on 2024!
So, what's on the horizon for the world of machine learning this year? Are we on the brink of a computer vision craze to overshadow the LLMs hype of 2023? With the shift from large to smaller language models, will we end up with just LMs, or perhaps SLMs? And is AI on the verge of becoming self-aware, ready to seek out the last business yet to integrate an LLM in its strategy?

Whatever’s in store, we’ll be here helping you make sense of it all. I mean, it’s only the first week in January, and already there are loads to catch up on, so I’ll quit my yacking and let you get cracking!

The Role of Infrastructure in ML Leveraging Open Source // Niels Bantilan // MLOps Podcast #197

Infrastructure is like punctuation.
you only notice it, when its Missing or. poorly Placed

Man, that was painful to type, but hopefully illustrated the point!

And it’s one that Niels makes in this chat, discussing the critical role of infrastructure in ML and harnessing AI and open-source tools.

He talks about the challenges of scaling models and data sets, the need for structured resources, the use of reinforcement learning signals, and the necessity of a metadata layer. He also shares how some of the problems he faced with data led him to create Pandera, a data quality tool, and its merger with Union, and why he decided against creating a company out of it.

be sure to, listen to it by Hitting the link; below {Sorry!]

Video || Spotify || Apple

Gotta throw in a GIF if I’m talking to the VP of Product Machine Learning at Giphy!

Even if it does show my age; Nick told me the hip young trendsetters are all about the stickers these days.

He does provide more relevant and less personally upsetting insights in this chat too. We talk about Giphy's tech journey, and how machine learning upped their game in search, vector embeddings, and creating GIFs.

We also get into his take on the mix between intuition and data in developing AI products, finding the right balance, and keeping your goals in sight. He also does great work supporting startups, and he shares his thoughts on the current scene of venture capital and startups.

A great chat to listen to while you sit and wait for GIFs to become cool again.

Video || Spotify || Apple

Challenges Operationalizing ML (And Some Solutions) // Nathan Frank // MLOps Podcast #199

They say aim for the stars. But do they mean celestial bodies, exceptional sports players, or just the Starbursts candy?

I don’t think anyone told Nathan, so he went for all three. We start off talking about how he started out as an astrophysicist working on the Skynet Robotic Telescope Network, and then running the numbers on athletes at Stats Perform. And I’m sure he’s probably had Starburst too at some time. There wasn’t time to ask in this chat though.

Instead, we got into Nathan's thoughts on tool development and how he balances using existing tools in his toolkit. We also dug into the similarities and differences between ML and software engineering, and Nathan shared his insights on MLOps from a DevOps perspective, as well as his thoughts on team communication and language barriers.

A great chat with a rising ML star - as long as he doesn’t like the yellow Starburst.

Video || Spotify || Apple

Founding, Funding, and the Future of MLOps // Mihail Eric // MLOps Podcast #200

This it! I can’t believe it’s happened, but here we are.

It’s quite an emotional moment, and one I wasn’t even sure would happen.

Yep, I’ve finally got to have another chat with Mihail! Oh, and this is also our 200th episode! Amazing! Thanks so much to all who listen. ❤️

And I want to repay your commitment with another great episode. We get into how AI can jazz up everything from video games to architecture, and even niche areas like print-on-demand.

He’s really open about fundraising attempts, including failures, and how he and Storia AI co-founder Julia Turc developed their project, "Rick and Mortify” into an AI storyboarding tool. He also shared his concern about overinflated expectations and the spread of inaccurate information. He has experience with the recent strikes in Hollywood, but I’m sure it’s something we can all relate to.

A great chat, just let’s not leave it until the 300th episode to catch up again though, Mihail!

Video || Spotify || Apple

💡Job of the week

Sr. Applied Scientist, Computer Vision // SafelyYou (California/remote, US)
As a Senior Applied Scientist in Computer Vision, you will develop and productize computer vision and machine learning models to support individuals living with dementia in assisted living and memory care settings. Responsibilities:

Develop and productize machine learning models that analyze images and videos to assist individuals living with dementia as well as their families and caretakers.
Work on core computer vision tasks such as object detection, human pose estimation, activity recognition, and video understanding in the context of assisted living and memory care.
Leverage our vast video dataset to enhance the precision and effectiveness of our services, both on embedded systems and in the cloud.

Requirements:

Postgraduate degree in Computer Science, Electrical Engineering, or related field
4-6 years of demonstrated experience in the field
Research experience in machine learning, computer vision, or image processing
Proficiency with Python and deep learning frameworks such as TensorFlow or PyTorch

How to Build a Knowledge Assistant at Scale

Knowledge assistants are great. Building them, not so great.

What would be handy would be a knowledge assistant for building a knowledge assistant.

I don’t have that for you, but this blog comes pretty close.

The diagrams alone are worth their weight in data gold. But beyond them, this comprehensive guide outlines the architecture and key considerations for creating a scalable, secure, and efficient knowledge assistant. Emphasizing scalability, security, transparency, modularity, and reusability, it presents an in-depth look at the multi-layered architecture of a knowledge assistant, including the data, LLM, reporting, and application layers.

Have a read and scale up your knowledge about knowledge assistants at scale!

With thanks to the QuantumBlack Team, Jannik Wiedenhaupt, Roman Drapeko, Mohamed Abusaid, and Nayur Khan for their contribution.

Unleashing the Power of LLMs in Healthcare and Wellness: Practical Context Providing in Healthcare and Wellness with Mistral-7B

It’s that time of year again to give up on your new year’s resolutions.
Honestly, it seems like it comes around earlier and earlier every year.

If only there was some ML way to support kick-starting a healthy new year… well, this blog looks at the integration of LLMs in the healthcare and wellness sector.

This approach could change how we use health data, offering deeper insights and personalized advice through the interpretation of outputs from eating and human activity recognition algorithms. While the potential for enhancing healthcare services is significant, the article also addresses the challenges, including privacy concerns and the need for ethical considerations.

A great read while you’re clocking the miles up on the treadmill. Or while you’ve got your feet up, what ever works for you.
With thanks to Bojan Jakimovski for the contribution.

AI Tidbits 2023 SOTA Report

There was a lot said about AI in 2023, much from those in the know, and even more from those not in the know!

Depending on who you listened to, AI was going to enslave us all or implode faster than FTX.

But, some people definitely in the know are the folks at AI Tidbits, and you will be too with this run-down of the state-of-the-art models in different areas.

They look at seven key areas, including language models, multimodal AI, and autonomous agents giving an overview and the SOTA model.

A great round-up that shows just how far things have come in a year. It will be great to see the progress we’ll have made this time next year too.

With thanks to Sahar Mor for the contribution.