Arup Nanda

Many #AI projects fail and for different reasons. In our book Generative AI with #langchain we talk about this and how to tackle these issues.

9

1 Comment

New spec to bridge semantic definitions between AI and BI. And a great list of partners in the mix.

1

A very interesting paper to start my weekend reading. Studying variances in LLM behaviours over most common human decision making use cases like Rock-Paper-Scissors is indeed an interesting way to predict and plan for meaningful divergences, when these LLMs are used for problem solving at scale..

8

1 Comment

The landscape of our industry is undergoing a significant transformation due to AI-generated solutions. It's essential for leadership and companies to recognize that change is not just on the horizon; it is already here. By embracing this shift, organizations can create vertically integrated solutions that serve customers more effectively, eliminating the reliance on costly and poorly designed third-party products. This foundational change in how we design solutions will enhance efficiency and improve outcomes. Nate's video on Anthropic's experience with AI being the CMS tool instead of helping with the configuration of a CMS tool. https://lnkd.in/g8dHrACa

1

1 Comment

I’ve been chasing knowledge graphs since 2019, back then we were trying to build a reasoner. They’ve surged in popularity with the rise of LLMs, and now everyone’s talking about them, but very few folks truly grasp the core ideas. Jessica absolutely nailed it, especially the open-world vs. closed-world analogy. Brilliant.

32

5 Comments

Thank you Seth Cohen for sharing Procter & Gamble’s blueprint for #leading_technology, a framework I believe is highly relevant far beyond the consumer goods industry. The three insights that resonate most with me: 1️⃣ Start with business capabilities and processes, not technology. Clearly define the capability or process transformation you want to achieve, then work backward to identify the technology required to enable it. 2️⃣ Adopt a platform approach that democratizes access to data, insights, and AI. This allows algorithms, generative AI models, and agentic AI to be scaled across all lines of business — rapidly accelerating capability development. 3️⃣ Reimagine advanced analytics. Move beyond static dashboards and traditional analytical applications toward intelligent, adaptive, and embedded solutions that let people converse with their data and insights. For me Technology leadership today is about aligning capabilities, platforms, and intelligence to unlock new value, not just for efficiency, but for transformation and an enhanced customer experience.

10

If an agent cannot simulate intervention outcomes, it lacks autonomy and acts as an autocomplete with elevated privileges. In insurance underwriting, this distinction decides if a system can handle a novel pandemic or replay 2019 patterns at the $M scale. In financial services, autonomous systems cannot rely on pattern recall. Fiduciary-grade reliability requires architectures that remain stable under distribution shifts by reasoning about interventions, constraints, and consequences rather than surface correlations. Objective-first learning: optimization over imitation Legacy decision systems often treat historical heuristics as ground truth and overfit yesterday’s decisions in a shifting environment. ⇢ Mechanism: Define reward functions and constraints for solvency, risk-adjusted value, and regulatory compliance. Learn policies that satisfy them. ⇢ Impact: Surface non-obvious, constraint-compliant policies that heuristic replay misses, primarily when historical labels reflect inconsistent or biased human judgment. World models: internal simulators as pre-execution governance Classifiers label cases. Autonomous systems need an internal dynamics model, a world model, to plan sequences and validate actions before execution. ⇢ Mechanism: Roll out candidate actions in an internal simulator, run stress tests and counterfactuals, and block actions that violate safety limits. ⇢ Impact: Move governance from post-hoc audit to pre-execution control. The agent acts only within a verified envelope. The causality boundary: OOD reliability as the benchmark Models can score high on MMLU yet drop double digits on tougher or decontaminated variants—an interpolation–extrapolation gap that sits at the heart of production AI risk under distribution shift. Scaling improves interpolation, but high-stakes systems must stay stable when the environment shifts. ⇢ Mechanism: Make causal assumptions explicit using causal graphs and intervention tests, and validate them with OOD suites that include regime shifts, novel fraud patterns, and tail-risk scenarios. ⇢ Impact: Passing OOD suites signals understanding of why outcomes occur—not memorization of what happened. The architectural pivot Sustainable autonomy pairs internal world models for planning with deterministic guardrails for execution. Metrics must go beyond accuracy to counterfactual safety, constraint satisfaction, and OOD robustness. Production-grade autonomy = reward-centric optimization + simulation-based gates + continuous OOD monitoring. #aiEngineering #systemsArchitecture #financialServices #insuranceTech #causalAI #worldModels #aiGovernance

3

Fifteen years ago, I was onstage with then Microsoft CEO Steve Ballmer receiving a Microsoft Partner of the Year award for our work with MG-ALFA and Azure. I am proud of the innovation as it was exceptional at the time. We were pioneers of HPC.. but we, and the industry, are a different place today. What I took away from this article is that after optimizing with state of the art CPU hardware, a 50% reduction in runtime was achieved for a signature client. And now Milliman are exploring GPU's. It was as much a concession as a success story. Fifteen years ago, Aon began developing #PathWise natively for GPUs - so they followed a different path, and now make a 50% reduction in runtime look meaningless. Recent benchmarks of PathWise have been in excess of 10,000x vs. CPU-based modeling software! So despite the runtime reduction, Milliman, along with every other legacy CPU vendor is pursuing GPUs... now trying to chase down the path PathWise has forged. What does this mean? 1. The competition universally has agreed GPU's are the better hardware for actuarial use cases. 2. The competition is facing massive re-writes of their models and calculation engines to even get started on the GPU journey. 3. #AonLifeSolutions took the long view and invested in GPUs before GPUs were cool, has spent 15 years designing and optimizing for the actuarial use case, and is now ushering in a paradigm-shift for the industry. 4. The competition will never catch up. They can't even enter the race without massive investments themselves, and by their clients... and Aon is not slowing down!

71

8 Comments