Juan M. Lavista Ferres

This one is called: “Fancy Context Engineering is Just Old School Dataset Curation”. Once you play around with LLMs a bit you find that there’s a bit of an art to selecting what you send into the machine — too little context and the model doesn’t have enough to generate a relevant response. But if you give it too much information the model doesn’t know what information is the most important. So there’s a sweet spot! This post by Thomasz Tunguz: https://lnkd.in/exJFf_gu, representing conversations across several practitioners, suggests a ratio of around 300 input tokens for each output token. This is surely domain-specific, as there are a few outliers. Context curation is important, it’s all the rage these days on LinkedIn. Nothing new to see here :) Then I found this paper from Google Research: https://lnkd.in/ebjumbsw. They show that the context can be replaced by an update to the weights of the underlying network. “In-context learning” really looks like “learning”. This means that context engineering is really a data curation problem. One of the first things I learned as a Data Scientist was to be careful about the data that you fed the model — both in terms of specific data points, but also paying close attention to the features you send into the model. Garbage in, garbage out, as they say. I think this points to some criteria for building your RAG pipeline, for example, and the few-shot examples you may send into the LLM. Specifically, information should introduce useful structure for the LLM, otherwise you’re “underfitting”, and examples should be relevant to the final output, otherwise you’re “overfitting”. You have to include enough diversity so that the LLM has enough of a sense of the “boundaries” you’re expecting, and contradicting examples or irrelevant information makes the model’s predictions less accurate. None of this is revolutionary, but I think the equivalence between constructing a context for the LLM and curating a dataset for a decision tree model offers some interesting next steps. And I’ve always loved finding analogies like this, because it gives you a mental map between a domain where you know a lot (“Old School Data Science”) and one where you know less. Happy Thursday Fam!

12

3 Comments

ICYMI: For companies with data-intensive architectures, there often comes a pivotal point where building in-house data platforms makes more sense than buying off-the-shelf solutions. By Ming Gao

1

AI works best with a human in the loop 🔁. In this episode, Ion Stoica, Co-founder of Anyscale and Databricks, explains why the most successful AI applications like code assistants and customer support thrive when humans are still reviewing and approving the output. He breaks down how this model applies to data teams using tools like Coalesce: boosting productivity without sacrificing accuracy or control. 🎥 Watch the full recording to hear Ion’s take on AI, reliability, and the future of data workflows: https://lnkd.in/gWzQRggW

19

We talk a lot about autonomy in agentic AI. We talk far less about recovery. I’ve published Part 3 of my series on Designing Agentic AI Architecture — focused on Observability, Analytics, and MTTR. Autonomous systems aren’t trusted because they’re accurate. They’re trusted because they’re measurable, traceable, and recoverable. This article covers: – Why logging is not observability – The three layers of agent tracing – Measuring override rates and policy violations – Why MTTR may matter more than model accuracy If Part 2 was about governance, Part 3 is about visibility. Part 4 will close the series with a real production incident-response agent. #AgenticAI #AIArchitecture #Observability #ReliabilityEngineering #Leadership

11

Researchers at Haize Labs are tackling the "Last Mile Problem" in AI by using fuzzing to evaluate GenAI systems. They found that traditional evaluation methods are inadequate for the brittleness of GenAI applications. The most important finding is their approach to scaling evaluation: "Scaling Evaluation with Agents as Judges" "Scaling Evaluation with RL-Tuned Judges" They also highlight the distinction between fuzzing and adversarial testing in AI. Case studies include: "Haizing a Major European Bank's AI App" "Haizing a F500 Bank's Voice Agents" "Scaling Voice Agent Evals with Verdict" These data points demonstrate their practical application of fuzzing to real-world GenAI systems. by AI Engineer https://lnkd.in/ev53hPEv

1

Yann LeCun famous saying “the platform that will win will be the open one.”, might not reflect Meta policy anymore. For years, Meta has chosen to open source its A.I. models, which means it makes the computer code public for other developers to build on. Now it might change as last week, a small group of top members of the Meta's AI lab discussed abandoning the company’s most powerful open source A.I. model, called Behemoth, in favor of developing a closed model. https://lnkd.in/d6vfXxjs

5

Four adoption pathways determine whether health AI scales or fails. Not because of model performance. Because of who moves first. Every health AI deployment falls into one of four patterns: 1- Top-Down Adoption 2- Bottom-Up Adoption 3- National Mandate 3- Ecosystem Adoption Infrastructure, providers, and citizens move together. Most leaders assume adoption is the same. It isn’t. Each pathway demands a different leadership response. → Top-down requires trust-building and workflow redesign. → Bottom-up requires governance and containment. → National mandates require coordination across systems. → Ecosystem adoption requires mature data and aligned incentives. The mistake is deploying with the wrong mental model. Health AI scales when leaders understand the pathway they’re operating within. Map the pathway before you deploy. Save this. Reference it in your next AI strategy session. 🔁 Repost if this framework sharpens your thinking. Follow Rizwan Tufail for practical frameworks on health AI adoption and system-level transformation.

57

31 Comments