Alan Cowen

🧬 The concept of Neural DNA (nDNA), initially proposed by Dr. Amitava Das, opens up a new avenue for interpreting LLM behavior. nDNA treats an LLM as something we can sequence (just like DNA sequencing) by reading its internal signals -- its lineage and layer-level changes. It captures a model’s lineage, tuning-induced mutations, and functional expression so we can compare LLM families, spot drift, and explain why similar benchmarks behave differently across a range of LLMs. In practice, nDNA looks for signals such as: (i) layer shifts, (ii) MoE routing patterns, and (iii) controlled probes. It lets us audit change over time and pick models for tasks with evidence. Collaborating with Dr. Amitava Das, and his Pragya group (https://pragyaai.github.io) at the Department of CSIS BITS Pilani Goa Campus, we've been building and testing the concept of nDNA to understand LLM behavior across major job families: - Llama tends to refine mid–to-late layers while early layers remain stable; - Mistral localizes expertise while the early layers remain steady; - Gemma aligns with light-touch parameter changes; - Qwen rewrites more in its middle layers, especially in multilingual scenarios; - DeepSeek adds dialogue capability in mid layers, while preserving base geometry. One of the other distinct advantages of nDNA is the ability to identify model collapse. On the surface you might see a slow loss of creativity and nuance in a model’s behavior. Inside, through the nDNA lens, the model’s internal landscape flattens -- fewer distinct patterns light up -- much like reduced neural diversity in biology. With nDNA, that shift becomes measurable over time, which makes collapse diagnosable and monitorable. Dr. Amitava Das's novel thought process and insistence on clear hypotheses has led to the inception of nDNA as a practical lens for AI -- one that detects drift before deployment, keeps model lineages auditable, and matches models to tasks based on evidence rather than guesswork. Taken together, these nDNA lenses reframe how we interpret LLM behavior. Our other recent works, DPO-Kernels (ACL 2025; https://lnkd.in/gXGQvXwX), Yin-Yang Align (ACL 2025; https://lnkd.in/gWVCxG8B), QuickSilver (https://lnkd.in/gpZQKMmP), and the Counter Turing Test (EMNLP 2023 Outstanding Paper Award; https://lnkd.in/gaVHeXgJ), share a similar first-principle approach to LLM refinement/understanding. If nDNA is something you'd like to explore further -- we'll be launching a website for nDNA soon, stay tuned! #nDNA #AI #FoundationModels #ModelInterpretability

283

5 Comments

A reminder that for the appraisal of prediction models - PROBAST+AI is now published. PROBAST+AI is used to assess the risk of bias, quality, and applicability of prediction models, and accompanies the reporting guidelines TRIPOD+AI (for transparency of reporting). Both papers are published open access in the BMJ. PROBAST+AI: https://lnkd.in/ekg2smy4 http://www.probast.org/ TRIPOD+AI: https://lnkd.in/es7Fmqes https://lnkd.in/eD4e5PPY Gary Collins Richard Riley (R²) Maarten van Smeden Laure Wynants Kamran Abbasi

197

3 Comments

People sometimes ask how we deal with model drift at Aampe. We don’t, for the most part. We don't have to. Agentic systems that operate on the basis of semantic-associative learning don’t have to worry about drift in the traditional sense, because the system learns continuously. Model drift is a problem you get when you deploy a static model and hope it stays useful. It’s rooted in an old deployment pattern: train a model offline, push it into production, and re-train it periodically when performance drops. That setup invites drift because the world keeps moving while the model stands still. Agentic infrastructure doesn't involve the deployment of fixed models. Each agent is continuously updating based on real-time user interactions. Every user has their own evolving policy, shaped by live feedback. Instead of worrying about drift, we focus on responsiveness: ➡️ Making sure exploration stays active ➡️ Adapting quickly to behavioral shifts ➡️ Keeping up with changes in reward signals Those adaptations aren’t on a quarterly retraining schedule - they happen constantly. Every agent's parameters are updated at least daily. Traditional ML setups come with more failure points than people often realize. Monitoring isn’t automatic—you have to build it, maintain it, and decide what exactly to watch for. Then, when performance starts slipping, someone has to make a judgment call: Has the model drifted enough to justify retraining? And once that decision is made, retraining itself is a whole new project. In contrast, an agentic system doesn’t treat learning as an occasional intervention. It’s built to act and adapt continuously from day one. Feedback isn’t a separate process—it’s baked into the core loop.

50

6 Comments

We are excited to share our latest pre-print enabling effective human oversight for AMIE, our research diagnostic dialogue AI. We introduce a new asynchronous oversight paradigm, decoupling history-taking by AMIE from sharing a human-approved diagnosis: https://lnkd.in/eYHwVc9h In this work, AMIE can perform consultations with patients to gather information within guardrails (g-AMIE), abstaining from individualized medical advice. A diagnosis and treatment plan is proposed, which licensed physicians authorize through our interface, the clinician cockpit, developed together with physicians. Our guardrailed-AMIE multi-agent system consists of a multi-phase dialogue agent, a guardrail agent and a subjective, objective, assessment, plan (SOAP) note generation agent based on Gemini 2.0 Flash. We evaluate this workflow in a virtual Objective Structured Clinical Examination (OSCE) study with oversight, contextualizing g-AMIE’s performance with control groups consisting of primary care physicians (PCPs) and nurse practitioners (NPs)/physician assistants/associates (PAs). In our study, g-AMIE and the control groups (g-PCP and g-NP/PA) all operate under the same guardrails of not providing individualized medical advice during consultations and draft SOAP notes for handoff. The SOAP notes are then reviewed by overseeing PCPs to make edits and either accept the edited note or not. The whole workflow is then evaluated by independent experts across a range of evaluation rubrics. We find that g-AMIE successfully performs high-quality intake with guardrails and produces more accurate and complete SOAP notes which lead to higher note acceptance and better oversight experience. Moreover, g-AMIE’s intake leads to better overall composite diagnostic performance and better oversight decisions as evaluated by independent raters. Our work has various limitations, including the difficulty of classifying individualized medical advice, the AI-focused workflow which was unfamiliar to both control groups, high mental load required for oversight and the simulated nature of our OSCE study. Because of this, our results need to be interpreted with care and cannot be used to draw conclusions about the relative performance of our PCP, NP and PA control groups. We still believe this is a significant milestone in enabling responsible and scalable use over conversational diagnostic AIs such as AMIE in healthcare. Amazing work by a great team from Google DeepMind, Google Research: Elahe Vedadi, David Barret Ellery Wulczyn Natalie Harris Shashir Reddy Roma Ruparel, MPH Adam Rodman Mike Schäkermann Vivek Natarajan Alan Karthikesalingam MD PhD Ryutaro Tanno Nenad Tomašev, Khaled Saab, Jan Freyberg, Tim Strother, Dylan Slack, Anil Palepu, Valentin Liévin, Wei-Hung Weng S. Sara Mahdavi Yash Sharma Pushmeet Kohli Joëlle Barral Dale Webster Avinatan Hassidim Yossi Matias Tao Tu Yun Liu  Kelvin Guu, Jihyeon Lee Cían Hughes Kavita Kulkarni James Manyika Katherine Chou

207

8 Comments

A primary goal of PhDs & Professors is to disseminate knowledge and educate the world—and the most effective way to do this in the AI era is by training LLMs, which in turn helps teach humans around the world at much larger scale. micro1 is driving this mission by collaborating with Professors like Mark Esposito, PhD to make it happen. Full interview coming soon to the micro1 youtube channel.

181

42 Comments