Owning our technology stack end-to-end enables us to optimize every layer - something that becomes increasingly important as we deploy ever-hungrier AI workloads. When we decided to double down on our own data centers and infrastructure rather than continue… more
Owning our technology stack end-to-end enables us to optimize every layer - something that becomes increasingly important as we deploy ever-hungrier AI workloads. When we decided to double down on our own data centers and infrastructure rather than continue migrating to Azure, it really came down to timing and scale. Azure was growing at an incredible rate (and still is!) because customer demand was going through the roof, and LinkedIn was seeing its own skyrocketing growth at the same time. Building fresh for the platform we needed was the right strategic bet for LinkedIn's future. While the internet's ChatGPT moment hadn't happened yet, we already saw companies like Meta reaping the benefits of developing larger and larger AI models, and the leverage from the systems they were building to support them. We bet on building that same kind of foundation ourselves and years later, the payoff is that our engineers can design entire systems end-to-end: intelligently splitting inference workloads across GPUs and CPUs, distilling large models into smaller ones that are very good at solving a specific optimization problem, and going as deep as writing our own GPU kernels with Liger Kernel to squeeze more out of every training run. Owning all of this made our teams more disciplined and more creative at the same time.
Thanks to Frederic Lardinois for a great conversation at the WeAreDevelopers Conference - video to come!like 73celebrate 7love 6insightful 2
Through 17 years of working at LinkedIn, I think I’ve interviewed thousands of candidates and there are a few things that consistently surprise people about us. From operating our own datacenters to our traffic scale to the innovation we're doing with AI… more
Through 17 years of working at LinkedIn, I think I’ve interviewed thousands of candidates and there are a few things that consistently surprise people about us. From operating our own datacenters to our traffic scale to the innovation we're doing with AI across both consumer and enterprise, it's a pretty remarkable range.
I recently sat down with Tim Jurka and Wenjing Zhang to discuss some of the technical challenges we're tackling that aren’t always obvious from the outside.
Check it out!like 126love 27celebrate 9insightful 5support 1
Must read alert!
Lea Kissner
I can't believe that our new book, "Building Safer Technology: A Field Guide to Failing Well" is here and in print. This book teaches how to think about… more
I can't believe that our new book, "Building Safer Technology: A Field Guide to Failing Well" is here and in print. This book teaches how to think about building technology in a world where things fail -- security, privacy, trust&safety, AI safety, and more. Because we concentrate on the underlying thinking and skills, this is durable knowledge which you can use even as the details of the technology changes every single day.
I don't have a copy yet, so you might be able to get one before I do! Get it from the publisher: https://lnkd.in/gG3G7Bzn
Or Amazon:
https://lnkd.in/g4r2QH8Ylike 42love 4celebrate 3insightful 2
Next “season” of #InTheTech filming today with these two characters. Thanks for being good sports, Wenjing & Tim!
Can’t wait to see the final cuts!
like 211love 48celebrate 20
Spot on Alborz.
Alborz Geramifard
AI isn't eliminating engineering. It's changing what great engineers do.
A conversation with Ashwin Machanavajjhala and Erran Berger inspired me to think… more
AI isn't eliminating engineering. It's changing what great engineers do.
A conversation with Ashwin Machanavajjhala and Erran Berger inspired me to think about this through the lens of one of computing's biggest transitions: from assembly to Python. I believe we're living through an equally important shift today. Here's my perspective.like 23insightful 3celebrate 1support 1
As AI moves from experimentation to production, discipline is going to matter just as much as ambition.
At LinkedIn we saw a while ago that the cost trends were clear, and that the future of our products would be powered by increasingly compute-expensive AI… more
As AI moves from experimentation to production, discipline is going to matter just as much as ambition.
At LinkedIn we saw a while ago that the cost trends were clear, and that the future of our products would be powered by increasingly compute-expensive AI workloads that are more intelligent, more capable, and more useful. But those workloads have to be ROI positive, and that’s not something you can just wake up and do overnight. So we set ourselves a challenge that sounds counterintuitive in today’s environment: how do we deliver increasingly sophisticated AI experiences while keeping our compute footprint as close to flat as possible?
This is where owning our own data centers and infrastructure as an applied ML company really pays off, because we can put in the instrumentation and make efficiency a long term investment up and down the stack rather than a one-time push. None of it came from a single breakthrough - it’s the outcome of hundreds of improvements, from optimizing GPU utilization and workload allocation to distilling larger models into smaller and more efficient ones to rethinking how work is distributed across training, inference, storage, and systems design. The results of that intentional push are starting to play out in terms of better experiences for our members and customers while giving our engineers the flexibility to keep innovating.
Raghu and I recently sat down with Paresh Dave at WIRED to talk through all of this, check it out!
https://lnkd.in/guYj674dlike 926insightful 31celebrate 29love 15support 6
The most interesting AI stories aren’t about the model, they’re about constraints.
At LinkedIn, we had a hypotheses that LLMs can be applied to traditional ranking and recommendations problems by better understanding what members find relevant. But using… more
The most interesting AI stories aren’t about the model, they’re about constraints.
At LinkedIn, we had a hypotheses that LLMs can be applied to traditional ranking and recommendations problems by better understanding what members find relevant. But using them for a domain like ads is far from straightforward. Decisions need to happen at enormous scale and in a fraction of a second, which means the usual approach quickly becomes too slow and too expensive.
That’s the engineering challenge Xīan Xíng Zhāng, Distinguished Software Engineer for Ads AI, and his team took on. They designed a way to use LLMs to better match what members are looking for with what advertisers are offering, while maintaining the efficiency and performance needed to operate at the scale of LinkedIn.like 116celebrate 7love 4insightful 4
