--- title: "LinkedIn’s Feed and the LLM Tradeoffs" newsletter: "MLOps Community" date: 2025-08-21 source: https://aaif.live/newsletters/mlopscommunity/2025-08-21-linkedin-s-feed-and-the-llm-tradeoffs --- # LinkedIn’s Feed and the LLM Tradeoffs *Plus experiments vs instinct, smarter agents, and multimodal reasoning* *MLOps Community — Agentic AI Foundation, 2025-08-21* Bad news [https://go.mlops.community/Can-eh-da] for anyone who thought ChatGPT was too polite, eh. ## Data vs Instinct Inspired by this short chat [https://go.mlops.community/intro21aug] with Fabricio where he talked about testing 100 ideas and letting experiments win over the CEO’s gut. When experiments and the CEO’s gut disagree, which wins in your org? [short chat](https://go.mlops.community/intro21aug) ## Hidden Gems ## Traditional vs LLM Recommender Systems: Are They Worth It? LinkedIn’s feed is quietly changing as LLMs start replacing the hand-crafted features behind traditional recommenders. Arpita explains what this means for speed, accuracy, and the cold start problem. * Prompts over features - Instead of feeding the model curated signals, LLMs can infer them directly, putting the real skill in how you ask the question. * Speed vs quality - Full-size LLMs can be too slow for live feeds, so teams use distilled models or run them offline for feature generation. * Faster personalisation - LLMs can tailor content with minimal data, beating traditional models in early-user scenarios. A smarter feed is possible, but only if your infra can keep up. Video [https://go.mlops.community/pav21aug] || Spotify [https://go.mlops.community/sav21aug] || Apple [https://go.mlops.community/aav21aug] KNOWLEDGE IS EVENTUALLY CONSISTENT When code is the only reliable source of truth, what happens to everything else engineers need to remember? Devin talked about agents that don’t just scrape knowledge but actually build and maintain it. * Fact-based reasoning - Dosu’s new agent commits verified “facts” into a knowledge base, cutting out duplicated searches and improving with each use. * Audience-aware responses - The agent adjusts its behavior depending on who’s asking - from expert maintainers to first-time users. * Knowledge maintenance at scale - By tying documentation back to code as the ultimate system of record, Dosu detects inconsistencies and prompts updates before docs go stale. The bigger your system, the more brittle knowledge becomes - Dosu’s approach is a bid to turn that fragility into a self-reinforcing loop of truth. Video [https://go.mlops.community/pds21aug] || Spotify [https://go.mlops.community/sds21aug] || Apple [https://go.mlops.community/ads21aug] AI CHANGED STACK OVERFLOW FOR THE BETTER Trust in AI code assistants is falling fast - even as adoption soars. Stack Overflow’s latest developer survey shows usage climbing from 60% to 80% over three years, but trust dropping from 40% to just 29% * Why it matters: AI is great for boilerplate, but often fails on complex, context-heavy tasks. That’s where human-curated Q&A still outperforms. * Stack Overflow’s role: Their data now underpins most major LLMs, with official licensing deals after years of scraping battles. * Inside companies: Enterprises like Uber are plugging private Stack Overflow instances into agents, powering internal copilots with vetted, constantly updated answers. The signal is clear: AI without trusted human knowledge hits a cliff. Video [https://go.mlops.community/ppc21aug] WHAT DOES MULTIMODALITY TRULY MEAN FOR AI? Multimodal AI is moving fast, but the real question is whether today’s systems can truly reason across the same streams of information humans use every day. This piece breaks down where progress is real and where the bottlenecks remain. * Models - From CLIP to Gemini, architectures are pushing toward any-to-any modality reasoning, though alignment and efficiency are still major challenges. * Processing - GPUs and custom accelerators have made real-time multimodal workloads feasible, but cost and optimization shape what’s deployable. * Data foundations - Embeddings alone flatten context; multimodal databases and knowledge graphs are needed to preserve structure and provenance. The takeaway: models and hardware are advancing quickly, but data management will decide how far multimodal AI can really go. Read the blog [https://go.mlops.community/blog21aug] IN PERSON EVENTS Ahead of the AI Agent Builder Summit in San Francisco (Sept 4), we’re adding two practical, hands-on workshops to the mix. Both are designed for people who want to stop talking about agents and actually build them. Here’s what’s on offer: * 4 hours of building (10am–2pm) * Real projects you’ll take away and extend * Food, drinks, and plenty of builder energy * Direct access to experts who’ve solved these problems at scale * A small, focused group of practitioners You can choose between: Agents + Authentication [https://go.mlops.community/AgentAuth]: Learn how to bypass OAuth nightmares, connect LLMs to APIs more intelligently, and set up agents that actually run in production. Voice AI Agents [https://go.mlops.community/VoiceAgents]: Go from STT → LLM → TTS toy examples to production-ready voice systems, complete with deployment strategies, observability templates, and real-world design tips. [Video](https://go.mlops.community/pav21aug) ## Job of the Week [https://go.mlops.community/jobsnl](https://go.mlops.community/jobsnl) ## Found In The Wild - Too Fast to Be Real [https://go.mlops.community/MLConfess](https://go.mlops.community/MLConfess) ## Making the hard stuff simpler Working on something tricky or planning ahead? Here’s how we can help - just hit reply: * Custom workshops tailored to your company’s needs * Hiring? I know some quality folks looking for a new adventure * Want to connect with someone tackling similar problems? I can introduce you Thanks for reading, catch you next time! --- Source: https://aaif.live/newsletters/mlopscommunity/2025-08-21-linkedin-s-feed-and-the-llm-tradeoffs