At the intersection of Technology, Data, AI and Humanity.
The gap between the builder's harness and the user's 'AI' is not a UX problem awaiting a better dashboard. It is structural. The builder side pursues fluency, autonomy, and minimal friction because those are what make the harness work: the closed-loop control, the automatic...
Read on LinkedIn →This is becoming a familiar scene. A doctor opens an AI structured note on her screen. The recommendation is precise, well-organized, supported by three references. She reads it twice, accepts it, signs off. Months later, when the case is reviewed, what will be examined is not...
Read on LinkedIn →Taken together, these studies share a crucial structural feature that is rarely made explicit in the coverage they receive. They all begin by decomposing a profession into a list of discrete tasks. They then score each task for AI overlap. They then aggregate those scores back...
Read on LinkedIn →John Ellis's classic book "Social History of the Machine Gun", published in 1975, is ostensibly a study of how one nineteenth-century weapon was invented, adopted, resisted, and ultimately used. Reading through it today can give us parallels and insights into how institutions...
Read on LinkedIn →The challenges of deploying AI, specifically LLMs, into air-gapped environments are typically framed as technical: How do we get the model onto isolated infrastructure? How do we maintain it without cloud connectivity? How do we ensure it doesn't hallucinate when cut off from...
Read on LinkedIn →Consider a model being trained to give health advice. Should it prioritize medical accuracy, deferral to experts, or reducing patient anxiety? All three are legitimate principles, but different annotators, shaped by different cultural norms around authority, directness, and...
Read on LinkedIn →The question for organizations operating in high-stakes environments is not whether LLMs are powerful, they plainly are. It is whether power alone is sufficient where decisions carry legal, medical, financial or life-and-death consequences. We have previously argued that...
Read on LinkedIn →The answer is not a simple story of decline. It is a story of inversion. The same transparency that empowers collection now exposes operations. The infrastructure of secrecy, once the source of advantage, increasingly becomes a constraint on the speed and scope of analysis. And...
Read on LinkedIn →In December 2014, the computer scientist Hannah Wallach stood before an audience of technologists and posed a question: "would you build algorithms on astronomers' data without involving astronomers? If not, then why are you building systems on social data without involving...
Read on LinkedIn →This is not a story about organizations that are too slow to modernize. It is a story about organizations whose core operating principles, the very things that make them effective, are in fundamental tension with the conditions AI requires to function. Understanding that tension...
Read on LinkedIn →While the military environment stands to benefit most from the latest AI technology, these are precisely the contexts where reliability and trust are hardest to achieve: adversarial settings where data patterns deliberately shift, where the cost of drift is measured in lives, and...
Read on LinkedIn →The most consequential change AI introduces is not to the analyst’s toolkit, but to the analyst’s role in the cognitive process. Across recent studies, a consistent pattern emerges: AI is pushing analysts “up the cognitive stack”: Routine data compilation gives way to judgment...
Read on LinkedIn →The pattern that emerges from the research is not 'chat versus GUI', but rather task-dependent preferences. Users favor conversational interfaces for simple factual queries and open-ended exploratory tasks. They prefer traditional visual interfaces for complex multi-step...
Read on LinkedIn →While productivity gains from AI are evident to most observers, societal acceptance has lagged behind technical progress. AI use can carry social and economic costs when outcomes are judged by humans. These dynamics introduce what researchers call the augmentation-approval...
Read on LinkedIn →The past year has seen a flood of 'reasoning models' promising everything from solving PhD-level math to generating production code. But a careful look at the research reveals a more nuanced picture: these models reason well within narrow, well-defined domains but struggle with...
Read on LinkedIn →The tale of the transformation of the archival institution in the AI age has lessons for us all in the way we should embrace the opportunities the new technology provides, yet still be vigilant and present when decisions of importance as to its implementation are made and chosen....
Read on LinkedIn →The migration of AI from centralized cloud infrastructure to edge devices brings new capabilities and new risks. When AI models operate on smartphones, sensors, and autonomous systems, the challenge of human oversight becomes dramatically more complex. How do we maintain...
Read on LinkedIn →Continual learning, the ability of AI systems to learn from new data without forgetting previous knowledge, seems like an unequivocal good. But recent research reveals an uncomfortable trade-off: the mechanisms that help AI retain old knowledge can actively interfere with its...
Read on LinkedIn →Within the enterprise setting, that complexity is magnified: dispersed silos, inconsistent standards, and incompatible systems, often managed under different governance rules, networks, or even jurisdictions, pose formidable obstacles to any unified data fabric. The absence of...
Read on LinkedIn →The rapid integration of AI into our daily workflows has outpaced our ability to establish clear norms and expectations. While regulators worldwide scramble to codify transparency requirements, recent research reveals a troubling dilemma: the very act of disclosing AI involvement...
Read on LinkedIn →An index appears at the end of many books: a list of words or topics with page references that help readers locate information quickly. Most modern indexes are word indexes (or concordances), which simply show word frequency and location. A subject index, by contrast, groups...
Read on LinkedIn →The standard prompting techniques that help AI with logical tasks, those step-by-step reasoning chains, actually work against creativity. Human creativity doesn't follow straight lines. It jumps, associates, surprises. It makes unexpected connections that current AI models, bound...
Read on LinkedIn →While model explainability has dominated AI discussions, contestability receives far less attention. Explainability helps us understand why an AI made a particular decision, while contestability empowers users to actively challenge, scrutinize, and influence AI outcomes.
Read on LinkedIn →As large language models reshape industries from healthcare to hiring, they carry a fundamental challenge: the biases embedded in their training data. These biases manifest in various ways but ultimately stem from the data these models were trained and fine-tuned on. Since models...
Read on LinkedIn →One of the trademarks of Large Language Models, and perhaps one of their endearing features, is their ability to mimic the nuances of human conversational interaction in all its complexity. One key aspect of this is the addition of controlled randomness to responses. Even when...
Read on LinkedIn →The promise of multi-agent AI systems is compelling. Multiple specialized agents collaborating seamlessly to tackle complex challenges. Yet despite the hype, these sophisticated frameworks often deliver surprisingly modest gains compared to simpler single-agent approaches. Why do...
Read on LinkedIn →The AI revolution has given us powerful generalist models that excel at many tasks. But what happens when you need deep expertise in a specific domain?
Read on LinkedIn →The use of AI Agent frameworks and workflows for automation and task accomplishment is well known. Perhaps less known is the ability of AI Agents to simulate and experiment with agents assuming various personas, allowing us to gain data and insights which would be either costly,...
Read on LinkedIn →