It was a busy day in Sunnyvale with customer presentations, my annual review, and time creating some new material. All good. Meanwhile, there’s a lot of a AI and DeepSeek again into today’s reading list.
[paper] Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. What do agents have to do with RAG patterns? This paper looks at an alternative way to enrich LLM results.
[blog] How to use Gemini over Vertex AI to summarize and categorize job listings with controlled generation. LLM responses are non-deterministic—you might not get the same answer back when asking a question repeatedly—but everything isn’t chaos. This post shows how you can get results in a structured JSON format.
[blog] Over 700 million events/second: How we make sense of too much data. Are you processing THIS much data? I mean, probably not? But it’s useful to see how Cloudflare chews through and makes sense of it all. Also, how they thwarted 47 million cyberthreats against Jewish and Holocaust ed sites.
[article] You’re not a senior engineer until you’ve worked on a legacy project. Throwing down the gauntlet! Probably fair. If you haven’t gotten your hands dirty on some old systems, it’s hard to have the experience to qualify as “senior.”
[blog] A new flexible, simplified, and more secure way to configure GKE cluster connectivity. Everyone won’t care about this, but platform engineers will. Good update for those running Kubernetes.
[article] Nvidia drops $600B off its market cap amid the rise of DeepSeek. You’ve now reached the DeepSeek portion of the reading list. Wow, what a grenade in the stock market yesterday!
[blog] DeepSeek and the Enterprise. Clear-eyed perspective from Stephen, as always. Trust may not yet be there, but showing that you can run open, cheap-to-train, and cheap-to-host models on-prem will be disruptive.
[article] Is DeepSeek really sending data to China? Let’s decode. The managed service and app? Absolutely. If you host the model on your own infrastructure (cloud or on-prem), no.
[article] DeepSeek FAQ. Great content. This is most anything you might be scared to ask at this point.
[blog] Open-R1: a fully open reproduction of DeepSeek-R1. This is the announcement of a PLAN, not a result from Hugging Face. Still, you’ll likely see this repeated elsewhere.
[article] DeepSeek’s new model shows that AI expertise might matter more than compute in 2025. I like this take. Don’t just host AI tech or use SaaS endpoints. Understand it. Upskill your teams. Innovation isn’t done, and make sure you have folks who understand more than just an API!
Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:
Richard Seroter's Architecture Musings 
Leave a comment