Andrew McVeigh

The enthralling part of quantum theory is that there's always "will" in every random way material constructions are built on, that being said it is possible to create matter out of nothing due to quantum probabilities.

2

The concept of neuromodulating agentic memory is helpful for defining "steering behaviours" in generative AI pipelines. Vince Kellen PhD proposed a #neuromodulation approach to agentic AI memory. This idea gives many opportunities to bundle explainable memory operations in rather complex spatial memory architectures. In this post, we discuss an emerging neuromodulation formulation created by recursive generative agents that reason about the memory space using neuroscience knowledge present in the context. The resulting "steering bundles" help add more explainable features to the semantic processes of memory usage. This is an important aspect if we need a human-in-the-loop workflow. We separate memory usage algorithms from vector space methods, which, while powerful, offer only a limited "flavor" of retrieval for agentic reasoning. #Aiagens #agenticmemory #generativeAI #AI #system2 https://lnkd.in/gcbGSr7t

8

see this event HIMSS Event Title: Patients Are Bringing AI Into the Exam Room—Now What? March 5, 2026 12:00 – 1:00pm CDT Presented by Harvey Castro, MD, MBA (DrGPT) and Doug Hohulin, Author and Kansas City AI Club Board Member - Humanities Chair  https://lnkd.in/gUfz6RHa we will discuss the topic outlined in the report. Patient Education is so important to maximize health outcomes. Patients are arriving in exam rooms armed with AI-generated information—some helpful, some misleading, and some potentially harmful. With over 40 million people using AI health tools daily, clinicians must guide how these tools are used—not compete with them. This session explores how physicians can remain in the driver’s seat by building patient literacy, strengthening trust, and turning AI-informed conversations into safer, more effective care. Through real-world examples, it highlights how clinicians translate information into judgment and action to improve adherence and prevent avoidable harm. Key Message: To be useful, AI must be harnessed to enhance the patient-doctor relationship (not replace it), improve preventative care, and address the modifiable risk factors that lead to chronic disease and early mortality. Harvey Castro, MD, MBA (DrGPT) is a board-certified emergency medicine physician, nationally recognized AI thought leader, and one of the earliest voices at the intersection of artificial intelligence, patient safety, and clinical workflow. With more than 20 years on the front lines of emergency care, Dr. Castro brings a uniquely grounded perspective on how AI is reshaping the office visit and the patient-clinician relationship. He is the author of ChatGPT & Healthcare (2022) and a five-time TEDx speaker focused on the practical and ethical adoption of AI in real-world clinical environments. Known as “DrGPT,” Dr. Castro advises health systems, governments, and healthcare organizations on deploying AI to enhance trust, preserve humanity, and improve outcomes without increasing clinician burden. In addition to his clinical work, Dr. Castro serves in executive and advisory roles across healthcare and technology, including as Chief AI Officer at Phantom Space Corporation, where he explores how advanced data infrastructure can support predictive, preventive healthcare models. At HIMSS, Dr. Castro speaks from lived clinical experience, translating emerging AI tools into what matters most at the point of care: how patients perceive them, how clinicians use them, and how trust is built or lost in the exam room. Doug Hohulin - is Working on how AI can save 1 Billion lives - on AI projects “When the AI System Has to Be Right" AI in Healthcare, Regulation/Governance/Policy, Road Safety/AV, Education and Energy. Doug worked for Motorola and Nokia for 33 years in Strategy, Business Development, Account Management and Engineering roles that included Digital Health Projects.  

8

2 Comments

Led by multiple teams from Oxford University, the study identified a significant gap in large language models (LLMs). Although these models are technically advanced in medical knowledge, they fall short in assisting users with personal health issues, researchers discovered. The findings show that their diagnostic results are comparable to traditional self-diagnosis methods, like online searches or personal judgment. Patients should be cautious when asking LLMs about symptoms, as they can lead to incorrect diagnoses or overlook signs of urgent medical needs.

6

2 Comments

There was an excellent post this morning shared by Jan Beger on multimodal AI. It was a commentary on a scoping review by Daan Schouten, Geert Litjens et al https://lnkd.in/gBgrP-_i that catalogs technical challenges and clinical applications in a well-written paper. That got me thinking about the definition of multimodal AI in the context of healthcare. While most of emerging literature and conceptual thinking still gravitates towards clinical data and imaging, there is a much broader view of multimodal AI that is imperative here and does not get enough attention yet. My definition of applied multimodal AI in healthcare refers to AI systems that can process, integrate, and reason across diverse, heterogeneous data types—including clinical, behavioral, environmental, socioeconomic, genetic, and self-reported information—to generate a comprehensive, context-aware understanding of an individual's health status and risk trajectory. It is not limited to clinical markers and signals as current agentic approaches allow a much wider address of risk markers across a patient's life journey - something that was hitherto impossible given the width of data this generates. Again, I speak from the perspective of a practitioner and at #CareCentra we have had some early success in synthesizing such multimodal signal data working with partners Intermountain Health and U.S. Department of Veterans Affairs in building such multimodal signal streams using constantly connected, remote care models to both predict and pre-empt exacerbations of risk in patients with a variety of chronic conditions. Thus for a congestive heart failure patient, multimodal signals in the applied world would include the following: Clinical: Ejection fraction, BNP levels, medications Behavioral: Step count trending down, sodium intake increasing Patient-Reported: Mild ankle swelling, reduced appetite Environmental: High heat + humidity index (risk of dehydration) Socioeconomic: Lives alone, limited transportation access Multimodal AI can integrate all this to predict risk of decompensation, trigger early intervention (e.g., virtual visit, medication titration), and recommend lifestyle changes personalized to the patient's motivation and ability. The implications of this broader definition are that: 1. It yields a richer digital phenotype: Moves us toward a more accurate and dynamic understanding of health and risk beyond traditional diagnoses. 2. It addresses continuous, not episodic care: Enables proactive care models that adapt based on real-time or near-real-time patient context. 3. It allows personalization at scale: By incorporating lifestyle, environment, and preferences, AI can tailor interventions and nudges; and 4. It builds towards equity & inclusion: Socioeconomic and environmental data help flag structural barriers to health—important for addressing disparities.

56

7 Comments

The paper introduces MedErrBench, a new benchmark specifically designed to evaluate medical error detection and correction in clinical text, addressing a critical gap in natural language processing (NLP) for healthcare. The authors highlight that inaccuracies in clinical content—such as misdiagnoses or incorrect treatment suggestions—can have serious and potentially harmful consequences. To support safer and more reliable applications of large language models (LLMs) and related NLP systems in medicine, MedErrBench provides a fine‑grained taxonomy of ten common medical error types and includes expert‑annotated clinical cases in three languages: English, Arabic, and Chinese. This multilingual approach ensures broader applicability and addresses underrepresentation of non‑English contexts in existing resources. The benchmark is used to assess a range of models, including general‑purpose, language‑specific, and medical‑domain language models, on three tasks: error detection, localization, and correction. Evaluation results reveal substantial performance gaps, particularly for non‑English languages and specialized medical content, underscoring the limitations of current models in accurately understanding and correcting clinically relevant errors. By releasing both MedErrBench and detailed evaluation protocols, the authors aim to foster research toward more clinically grounded, language‑aware systems that can improve patient safety and equity in AI‑based healthcare. https://lnkd.in/ge6iYm8Z

13

Regression to the mean is the quiet flaw of large LLMs. When we drown a model in broad context, meanings get averaged. The sharp edges, the domain-specific sense of a word, get blurred into something safe and generic. A proposal: treat skills as attachable parts. Use LoRA (low-rank adapters) as compact, auditable skill modules (SQL-LoRA, React-LoRA, etc.) and let the system snap them on only when the task calls for it, then pop them off when it doesn’t. RAG-aware dynamic adapters Imagine a router that looks at your prompt and what retrieval pulls in, then selects a small set of adapters to tighten the model’s semantics for this specific exchange. Next turn, new context, new adapters. Specialization becomes momentary and precise, not a permanent fine-tune. How this differs from MoE (Mixture-of-Experts) • Explicit control vs. hidden gates: MoE routes tokens to learned experts; dynamic LoRA is an explicit, human-governable skill switchboard. • Add skills without retraining: New adapter, new capability—no heavyweight re-gating. • Lean infra: Works on a single base model with lightweight adapter loading, not sparse-expert plumbing. • Predictable ops: Small, swappable artifacts, easier to version, permission, and reason about. Why this matters • Sharper semantics on domain tasks (less averaging, more intent). • Composable skills for cross-domain problems. • Practical path to breadth without a zoo of fragile full fine-tunes. I’m exploring this pattern, RAG-aware LoRA routing, as a way to shift LLMs from “generally smart” to “situationally precise” without ballooning cost or complexity. If you’re working on adapter routing, retrieval, or serving LoRAs in production, I’d love to compare notes. #LLM #LoRA #RAG #GenAI #MLOps #AIEngineering #NLP

10

2 Comments

I’ve been exploring a simple question: How far can a purely classical, geometry-based search heuristic go on commodity hardware? The screenshot below is from an experiment I’ve been running over the last few days. It’s a 1-D manifold search kernel that adapts using a k-means-style update rule. Nothing exotic—just gradients, curvature, and a lot of trial-and-error tuning. For this run: Search space: 1.6 billion Cells: 80 million Batch size: 100 random targets Mean absolute error: ~5 Ideal grid spacing: ~5 Hardware: standard laptop Most of the results land within a small integer window around the target. The surprising part, at least to me, is how stable the behavior gets when you scale the number of “cells” into the tens of millions and keep the learning rate small. It settles into something close to an evenly spaced manifold on its own. I’m not making any big claims about this work. I’m still trying to understand why the convergence is so consistent and what the limits are. But the experiments have been fun, and I’m sharing them in case others enjoy this kind of exploratory tinkering. If anyone has background in numerical optimization, vector quantization, or geometric search heuristics and wants to offer critique or suggestions, I’d appreciate it. Still early days, still learning. #GLHF #YOLO #PicsOrNotReal

10

3 Comments

Here are some interesting examples of how to blend rule-based expert systems with LLMs. Here, the rule-based system guides the user through collecting the information to make a well-informed prompt for the LLM, which then gives advice. (Thus avoiding the problem of the LLM being given too little to be helpful, and also confining the LLM to the intended subject.) Note that these are demos, and you will see some technical details of how they work. When you get to the end, be sure to SCROLL DOWN on the last screen in order to see conclusions. Rule-based AI is a time-honored approach, and I don't think we should throw it away now that we have LLMs -- instead, blend them, and let each do what it's best at!

7

3 Comments

Where's the Shovelware? Why AI Coding Claims Don't Add Up, https://lnkd.in/e62WuSi5 IA Summary: Are claims of AI coding tools boosting developer productivity by 10x or more a massive exaggeration? This article presents compelling evidence—including flatlining software release data—that challenges these claims and exposes the hype surrounding AI-assisted coding.

1

We built a high-compute solver engine for constraint-heavy problems (logistics, scheduling, security verification, biotech design constraints). When connected to KARIOS, it returns a concrete answer (SAT/UNSAT + witness when applicable) — and it also returns a verifiable audit bundle (proof/witness + replay artifacts). On many workloads it runs in sub-second to seconds on laptop-class hardware. #Optimization #Logistics #Cybersecurity #Verification #AI #Algorithms #SAT #ConstraintSolving #OperationsResearch #EnterpriseSoftware #Biotech #Scalability

2

UC Berkeley published a failure analysis for multi-agent systems.                                                                            They ran 1,642 execution traces across 7 multi-agent frameworks, and it turns out that 79% of failures come from agents misunderstanding each other. Not from the models being incapable. Context gets lost during handoffs, and outputs are misinterpreted between different agents. Almost 4 out of 5 failures are coordination problems. This matches something we learned building Parlant.. Parlant is designed for conversation, where users jump topics, ask compound questions, and change their minds regarding their intent. What makes dialogue problematic for multi-agent frameworks is that dialogue has this property where everything connects to everything else, and splitting that across agents creates more problems than it solves, architecturally.            What worked for us was having just one agent - but it dynamically loads all the relevant rules for each conversation turn. This way you get the topic-specific focus of a sub-agent in a multi-agent system, while side-stepping the coordination problem. That's not to say single-agent is always better. For linear task pipelines, multi-agent can work great. But customer conversations aren't a pipeline like that. Incidentally, the paper concludes that multi-agent systems probably need fundamental structural redesigns. For conversations, that's Parlant. https://lnkd.in/dmfEkDxS

54

7 Comments

📚 AI-Slop in science - the meaningless phrase "vegetative electron microscopy" appearing in nearly two dozen published papers is likely caused by a strange AI-misinterpretation. I highly advice reading this article form Retraction Watch: https://lnkd.in/eNc4ad7A This is the dark side of AI - when automation meets academic fraud, nonsense can spread faster than ever. We need both better detection tools and human vigilance. #ResearchIntegrity #Publishing #PeerReview #AISlop

2

The following essay is about Aryanology: Rediscovering the Wisdom of the Aryan Civilization!

1

“Everyone says their system is people-centric. But where is that actually embedded? Who governs the governance? Who defines the boundaries, and at what layer? If the guardrails for superintelligence exist only at the policy layer or application layer, they’re conceptual — not structural. Any sufficiently advanced optimizer will maneuver around them. The only real protection is to anchor sovereign governance at the substrate level — the silicon, or whatever material platform the end-state intelligence runs on. If human sovereignty isn’t a primitive at the lowest layer, then ‘people-centric’ becomes a slogan instead of a rule the system must obey.”

1

Synthetic data might just be the game-changer we need! 🌟 A team of Dutch researchers has uncovered how LLM-generated synthetic data could address the looming data crunch—especially in evaluating RAG systems. This could mean smarter workflows, faster innovation, and a whole new way to tackle enterprise challenges. 🚀 At Pega, we know the power of innovation and AI-driven solutions to solve complex problems. This breakthrough resonates with our mission to "Build for Change®" and transform how businesses operate. Curious? Dive into the full article here: https://lnkd.in/gaHgdZvG Let’s keep pushing boundaries together! 💡

2

Quantum Coherence Theory is now publicly available. Key Ideas in QCT 1. Coherence Is Structural, Not Merely Dynamical: Quantum coherence is not just superposition or phase correlation. It is the maintenance of internal consistency relative to shared constraints, even under interaction. Coherence fails when alternative configurations can no longer remain jointly admissible under accumulated constraint. 2. Decoherence Is Admissibility Reduction: Decoherence is not collapse, observation, or epistemic update. It is a structural narrowing of admissible configurations as environmental coupling increases. Classical behavior emerges when the admissible set becomes so restricted that alternative trajectories are no longer viable. 3. Energy Is Distributed and Constrained Potential: In QCT, energy is treated not as a driver of outcomes but as distributed potential shaped by constraint density. Stability arises where energy distributions are constrained into configurations capable of persisting under interaction. Classical determinism corresponds to high-constraint regimes. 4. Nonlocal Does Not Mean Unconstrained: Quantum nonlocality is treated as non-spatial correlation, not unconstrained influence. Entangled systems share constraint structure across spatial separation. Nonlocal correlation reflects shared admissibility, not violation of causality or unrestricted signaling. 5. Shared Global Constraint Is Required: For interaction to remain coherent, all physical processes must be constrained under a shared global invariant structure. Local freedom exists, but only within bounds compatible with that invariant. Without shared constraint, interaction dissolves into incoherence rather than branching worlds. 6. Curl Equilibrium and Stability: QCT interprets stable quantum and classical configurations as curl-balanced constraint flows. Persistence requires that constraint circulation does not accumulate unresolved tension. Where curl equilibrium fails, coherence degrades and admissible structure collapses. 7. Classicality Is a High-Constraint Limit: The quantum–classical transition is continuous. Classical behavior is not introduced; it emerges when constraint density restricts admissibility to a narrow trajectory. No new laws appear. The space of viable alternatives simply vanishes. What QCT Is Not - ❌ Not the Many-Worlds Interpretation - ❌ Not wavefunction collapse - ❌ Not epistemic Bayesian updating - ❌ Not observer-dependent realism - ❌ Not a new quantum dynamics QCT is an explanatory framework, not a replacement formalism. More Information To check out the full publication regarding Quantum Coherence Theory, check out the following: Viable Worlds Theory: A Coherence Science Account of Quantum Coherence and World Viability https://lnkd.in/es3NGSV9

1