|
|
|
|
Some suggestions for an MLOps playlist...
The BlockChain by Fleetwood Mac
Back in Batch by AC/DC
Walk This Array by Aerosmith
Don't Stop Retrieving by Journey
I Want to Code Your Brand by The Beatles
We Will Dock(er) You by Queen
|
|
|
|
|
|
|
The Myth of AI Breakthroughs // Jonathan Frankle // MLOps Podcast #205
Move over Bill Nye, we’ve got ourselves a new science guy!
Actually, I think Jonathan would be far too civil to usurp Bill.
Well, I say civil, but he has managed to get himself ‘blacklisted’ from some panels, but that’s just for the tenacious way he tries to get behind the hype and over-used ‘lines’ to the nuts and bolts. This ties into how he made it clear in our chat the need to “do the science” – that is, to cut through the hype and let the scientific process run its course.
He shares insights about his role leading a
research and science team, the creation of Mosaic ML, transparency thorough research, and the importance of practical application. Jonathan also highlights the balance between policy, technological advancements, and ethics, particularly when it comes to face recognition systems.
The work he’s most proud of, The Perpetual Line-Up, relates to this. Rather than the work he gets asked about the most, the lottery ticket hypothesis, The Perpetual Line-Up is all about improving the efficiency of training models and being empirical about how we study AI.
It was a great chat, and he’s not blacklisted from the MLOps Community!
|
|
|
|
|
RAG Has Been Oversimplified // Yujian Tang // MLOps Podcast #206
I’m not exaggerating when I say this chat was on fire!
Literally, we chat about Yujian’s
fire-making capabilities.
But of course, it wasn’t just on fire, it’s also on RAGs. What is a RAG, optimizing RAGs, multi-modal RAGs, and the possibilities, things that you need for the RAG stack like the embedding model, the LLM, and the vector database, and of course, how RAGs are not for everything.
A really interesting aspect was also the functional difference between context and relevance, and how rags use both for accurate predictions.
We also get into application design, running apps involving ingestion processes and setup for a GUI, and the suggested architecture for creating conversational bots.
He also shares how he’s built a fashion AI tool for fun. But, as an example of how RAGs aren’t needed for everything, it uses a vector database similarity search, not a RAG app.
So, ironically not using RAGs to help with your glad rags.
|
|
|
Gen AI Roundtable in collaboration with QuantumBlack,
25 January
|
|
|
|
|
|
|
Join industry experts Nayur Khan, Ilona Logvinova, and Mo Abusaid from QuantumBlack for a session filled with insightful discussions on:- The trade-offs in the GenAI space.
- The challenges of black box solutions and data transparency.
- Essential insights for both business leaders and developers.
Whether you are steering a business through these technological changes or a developer at the forefront of GenAI innovation, this roundtable promises to equip you
with the knowledge to make more informed decisions. Click here to register now
|
|
|
💡Job of the weekSenior Product Manager - Generative AI & ML // Fiddler (US, remote) Fiddler partners with AI-first organizations to help build a long-term framework for responsible AI practices, which, in turn, builds trust with their user base. Joining us means you get to make an impact by helping reduce algorithmic bias and ensure that models in production across many different industries are transparent and ethical.
What You'll Do- Drive an E2E product within the LLMOps and MLOps workflow, including research, design, execution, and GTM
- Iterate the product over time to incorporate user and market feedback
- Inform our product roadmap through industry and customer research and champion the product to both internal and external stakeholders
What They're Looking For- 5+ years of software product management experience with at least two years in ML
- Enterprise or early-stage experience preferred
- Technically fluent: You have experience building technical systems for technical users
- Strategically minded: You can quickly grasp both business and customer
motivations and goals and craft a plan to drive toward them
|
|
|
MLOps Community IRL Meetup
|
|
|
|
|
Lessons from Hacking Together a Customer Research Tool // Habeeb Shopeju // IRL Meetup #61 Lagos
Finding and analyzing pain points can be a pain. Helpfully, this talk walks through the process of building a pain-point research tool. Habeeb takes us through his journey of hacking together a cutting-edge customer research tool. The tool pulls relevant posts from Reddit, processes them with a potent combination of complaints classification, keyword extraction, and AI annotation, and stores them in a simplified vector database. He also sheds light on the significant role of embeddings, LLMs, and abstraction in ML systems. Some good research for your research tool!
|
|
|
|
|
|
|
Survey of Vector Database Management SystemsWhat’s our vector, Victor? It’s a fair question these days, with over 20 commercial vector database management systems available. The paper provides an in-depth exploration of VDBMS, focusing on their role in handling large, unstructured data. It outlines the challenges such as semantic similarity ambiguity, managing large vector sizes, and the high cost of similarity comparisons. It also discusses innovative techniques in query processing, storage, indexing, and optimization, highlighting advancements in similarity scoring and vector compression. It differentiates between 'native' VDBMS, built specifically for vector data, and 'extended' systems, which add vector capabilities to existing databases. Finally, it
touches on future research directions, emphasizing the need for benchmarks and solutions to emerging challenges in VDBMS. We have clearance, Clarence.
|
|
|
|
|
Add your profile to our jobs board here
|
|
|
|
|
|
|
|
|
|
|
|
|
