What are the key takeaways from this Software Engineering Daily episode?

Key insights include: **Sequence Modeling for Recommendations:** Treat content recommendation as next-item prediction rather than independent item scoring. LinkedIn models user content paths the way LLMs predict tokens — someone reading an intro to PyTorch article can be mapped 15 steps forward toward building a local LLM. This approach captures contextual momentum that flat feature-crossing architectures like DCNV2 cannot replicate.; **Natural Language Quality Policies:** Define content quality as version-controlled text files checked into GitHub, then use LLM-based auto-evaluation to audit whether surfaced content meets those standards. Product managers iterate on these policy prompts using GitHub Copilot CLI, which returns scored examples and suggests wording revisions — creating a closed feedback loop without requiring manual review at scale.; **GPU Cost Reduction via CPU Offloading and Batching:** Reduce inference costs by routing non-model tasks — language filtering, faceted filtering, preprocessing — to CPU before loading items into GPU memory. Batch feature updates on 20-minute windows rather than per-event writes to reduce database load. Custom CUDA attention kernels further reduce per-request compute cost at the GPU level.

What did Tim Yerka discuss on Software Engineering Daily?

LinkedIn VP of Engineering Tim Yerka explains how the platform rebuilt its newsfeed for 1.3 billion members by replacing traditional MLP scoring architectures with sequence-based models and LLM-powered retrieval, covering inference cost management, natural language quality policies, and the engineering team structure required to operate at this scale. Key topics include: **Sequence Modeling for Recommendations:** Treat content recommendation as next-item prediction rather than independent item scoring. LinkedIn models user content paths the way LLMs predict tokens — someone reading an intro to PyTorch article can be mapped 15 steps forward toward building a local LLM. This approach captures contextual momentum that flat feature-crossing architectures like DCNV2 cannot replicate.; **Natural Language Quality Policies:** Define content quality as version-controlled text files checked into GitHub, then use LLM-based auto-evaluation to audit whether surfaced content meets those standards. Product managers iterate on these policy prompts using GitHub Copilot CLI, which returns scored examples and suggests wording revisions — creating a closed feedback loop without requiring manual review at scale..

How long is this episode of Software Engineering Daily?

This episode is 47 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

Software Engineering Daily

How LLMs Are Reshaping Recommendation Systems

August 18, 2026

47 min episode · 2 min read

Tim Yerka

Episode

47 min

Read time

2 min

Topics

Career Growth, Investing, Fundraising & VC

AI-Generated Summary

Published Aug 18, 2026

Key Takeaways

✓Sequence Modeling for Recommendations: Treat content recommendation as next-item prediction rather than independent item scoring. LinkedIn models user content paths the way LLMs predict tokens — someone reading an intro to PyTorch article can be mapped 15 steps forward toward building a local LLM. This approach captures contextual momentum that flat feature-crossing architectures like DCNV2 cannot replicate.
✓Natural Language Quality Policies: Define content quality as version-controlled text files checked into GitHub, then use LLM-based auto-evaluation to audit whether surfaced content meets those standards. Product managers iterate on these policy prompts using GitHub Copilot CLI, which returns scored examples and suggests wording revisions — creating a closed feedback loop without requiring manual review at scale.
✓GPU Cost Reduction via CPU Offloading and Batching: Reduce inference costs by routing non-model tasks — language filtering, faceted filtering, preprocessing — to CPU before loading items into GPU memory. Batch feature updates on 20-minute windows rather than per-event writes to reduce database load. Custom CUDA attention kernels further reduce per-request compute cost at the GPU level.
✓Multi-Objective Engagement Modeling Personalized Per User: Optimize simultaneously for positive signals (comments, reshares, dwell time) and negative signals (skips, viewport passes) using a multi-head model. Critically, the weights of these signals are personalized per individual user — not by role or persona — because a job seeker and a marketer require fundamentally different feed optimization targets.
✓LLM Context Window Discipline: Avoid dumping raw structured data into LLM prompts at inference time. LinkedIn compresses member context (profile, recent sessions) and item context (author, post content, image captions) into compact token representations. Poorly constructed context windows degrade precision and inflate compute costs — a dedicated team works specifically on context construction and sequence sampling strategies.

What It Covers

LinkedIn VP of Engineering Tim Yerka explains how the platform rebuilt its newsfeed for 1.3 billion members by replacing traditional MLP scoring architectures with sequence-based models and LLM-powered retrieval, covering inference cost management, natural language quality policies, and the engineering team structure required to operate at this scale.

Key Questions Answered

•Sequence Modeling for Recommendations: Treat content recommendation as next-item prediction rather than independent item scoring. LinkedIn models user content paths the way LLMs predict tokens — someone reading an intro to PyTorch article can be mapped 15 steps forward toward building a local LLM. This approach captures contextual momentum that flat feature-crossing architectures like DCNV2 cannot replicate.
•Natural Language Quality Policies: Define content quality as version-controlled text files checked into GitHub, then use LLM-based auto-evaluation to audit whether surfaced content meets those standards. Product managers iterate on these policy prompts using GitHub Copilot CLI, which returns scored examples and suggests wording revisions — creating a closed feedback loop without requiring manual review at scale.
•GPU Cost Reduction via CPU Offloading and Batching: Reduce inference costs by routing non-model tasks — language filtering, faceted filtering, preprocessing — to CPU before loading items into GPU memory. Batch feature updates on 20-minute windows rather than per-event writes to reduce database load. Custom CUDA attention kernels further reduce per-request compute cost at the GPU level.
•Multi-Objective Engagement Modeling Personalized Per User: Optimize simultaneously for positive signals (comments, reshares, dwell time) and negative signals (skips, viewport passes) using a multi-head model. Critically, the weights of these signals are personalized per individual user — not by role or persona — because a job seeker and a marketer require fundamentally different feed optimization targets.
•LLM Context Window Discipline: Avoid dumping raw structured data into LLM prompts at inference time. LinkedIn compresses member context (profile, recent sessions) and item context (author, post content, image captions) into compact token representations. Poorly constructed context windows degrade precision and inflate compute costs — a dedicated team works specifically on context construction and sequence sampling strategies.

Notable Moment

Yerka describes a compounding reliability problem with multi-stage agentic systems: if each of 20 sequential steps succeeds 99.5% of the time, the end-to-end workflow becomes effectively unusable. This framing reframes agentic AI reliability as a systems engineering problem requiring grounding signals, not just model accuracy improvements.

Know someone who'd find this useful?

Episode Transcript

News feeds and recommendation systems have long relied on deep learning architectures that score each candidate item independently. As LLMs have matured, they have opened up a fundamentally different approach where a system can reason about content the way it reasons about language. However, that power comes with a fresh set of engineering challenges around cost, scale, and evaluation. LinkedIn recently rebuilt its newsfeed to treat content recommendation as a sequence modeling problem. The general approach is to predict what a user will want next, much like an LLM predicts the next token in a sentence. Tim Yerka has worked at LinkedIn for thirteen years and is currently a VP of engineering. In this episode, Tim joins Matt Merrill to discuss how LinkedIn reengineered its feed, how the team combines LLMs with traditional signals, managing inference costs at massive scale, steering content quality using natural language policies, and more. Matt Merrill is a software engineering leader with over twenty years of experience building and scaling software teams across enterprise and product focused organizations. His background is in back end development, cloud architecture, and distributed systems design. He currently architects and delivers software products and leads a team of engineers at Dept Agency. You can learn more about his work at co.theothermattm.com. Hey, everybody. I'm Matt Merrill, and I am here with Tim Jurka, VP of engineering from LinkedIn. So today, we're gonna talk a little bit about what LinkedIn's doing with revamping their social feed with AI. But before we start, Tim, can you tell us a little bit about yourself and your current role at LinkedIn? Thanks for having me, Matt. Like, as you mentioned, my name is Tim Yerka. I am VP of engineering at LinkedIn, and I work on our consumer products. So what that would be is things like the LinkedIn feed, search experience, profile, pretty much anything you will engage with when you open up the LinkedIn app. And I actually started here thirteen years ago as an engineer on the feed. I was an AI engineer, one of the first AI engineers building the LinkedIn feed. So it's been pretty amazing to go from actually, like, building the thing to now see the thing operating out just massive scale with, you know, 1,300,000,000 members. It's been quite a journey. So recently, you guys have rolled out some pretty major updates to how you're making the LinkedIn feed relevant to users. And I know that's welcome news to me, and it's probably welcome news to a lot of people who may not have otherwise engaged with LinkedIn because of content that was irrelevant or they thought was annoying. So I'm really excited to hear about what you did and also how it intersects with AI tooling, which I think this audience will really like. So let's start with just what was the problem. Right? Like, what was the overall problem, and why was it challenging? I mean, I can maybe start by rooting it in, …

Get the full transcript (9,941 words) + summary by email — free

One-time email with the complete transcript and AI summary of this episode. No account needed.

One email, no spam. We’ll also show you what SignalCast does.

Browse all Software Engineering Daily transcripts →