Piyalitt Ittichaiwong

AICamp London is kicking off March with another meetup! Come join us with over 170+ other people in the AI space. Topics : 1 ) Agent2Agent protocol by Holt S. (Google) 2) What Exactly Is Going On Inside an AI's Brain? by Xiangpeng Wan (NetMind.AI) A2A is a standard that would enable collaborative agentic work at scale, I'm pretty excited for this one. We also previously discussed AP2 (Agent Payment Protocol) at the last meetup with XiangPeng ... seems like a natural combination here. NetMind.AI is kindly hosting the community, so please give them some love. They have superb OSS model APIs and all the cool agentic stuff including a piece on OpenClaw. Catch you there tomorrow ... https://lnkd.in/e6xGPreg Bill Liu, Lorentz Yeung, Fatemeh Mosleh, Tibor O. Shenghui Tao, Alice Mao, Stacie Chan, Hao Wang

26

5 Comments

Google DeepMind's research uncovers a fundamental limitation in Retrieval-Augmented Generation (RAG): embedding-based retrieval cannot scale indefinitely due to fixed vector dimensionality. Their LIMIT benchmark (from the Aug 2025 paper "On the Theoretical Limitations of Embedding-Based Retrieval") demonstrates that even state-of-the-art embedders like GritLM, Qwen3 Embed, Promptriever, and E5-Mistral fail to consistently retrieve relevant documents, achieving only ~30–54% recall on small versions and dropping below 20% (e.g., GritLM 12.9%, E5-Mistral 8.3%) on the full, more challenging dataset. In stark contrast, classical sparse methods such as BM25 achieve 93.6% Recall@100. So, how do you fix it? Introducing NUMEN A simple yet powerful response to the "geometric curse" of low-dimensional embeddings. NUMEN fixes it without training, without learned embeddings, and without vocabulary limits :     • Tokenize text into character n-grams(typically 3–5 grams, with start/end markers like ^ap, app, ppl, ple, le$ for "apple")    • Compute CRC32 hash for each n-gram (deterministic & fast)    • Map those hashes to indices in an arbitrarily high-dimensional sparse vector (e.g., 8k, 16k, 32k+ dims)    • Use standard cosine similarity for retrieval The result? A truly dense retriever that scales its capacity geometrically by increasing dimension; no neural network required. Results on the LIMIT benchmark (Recall@100):   - NUMEN @ 32768d : 93.90%   - BM25 baseline : 93.6%   - Modern dense SOTA (E5-Mistral 7B, GritLM 7B, etc.): 8–19% in typical 3072–4096d spaces This makes NUMEN the first dense approach to officially beat BM25 on this notoriously difficult benchmark designed to expose embedding limitations. Why it matters:   The DeepMind work argues the core problem isn't model size or training data. It is forcing infinite linguistic nuance into tiny fixed vectors. NUMEN sidesteps learned embeddings entirely, proving that simply giving dense retrieval more dimensional "room" via high-dim hashing can match sparse lexical search on hard tasks. This concludes that sometimes the solution isn’t piling on massive models but quietly giving the “geometric curse” a knowing wink and saying: why torture language in a tiny fixed box when you can just hand it a much wider playground? 😉 💻 Code :  https://lnkd.in/gYaZwwBr    📄 Paper : https://lnkd.in/gTGqr7RY #InformationRetrieval #DenseRetrieval #Embeddings #RAG #VectorSearch #NUMEN #LIMITbenchmark

4

1 Comment

These exits matter. Not because they make headlines, but because they reveal something deeper about direction and integrity inside organizations. Yann LeCun, Meta’s longtime Chief AI Scientist (Turing Award winner, and considered one of the founding fathers of modern deep learning) is reportedly leaving Meta to launch a new company focused on world models that learn from video and spatial data instead of text. Behind the scenes, this move follows Meta’s sweeping AI reorg, where LeCun began reporting to Alexandr Wang and FAIR (Meta’s AI research lab) saw over 600 roles cut. His departure is symbolic. In every sector, whether we’re talking about Big Tech, Medicaid transformation, or public-sector AI, these moments mark a fork in the road: Do we double down on scale, speed, and short-term returns? Or do we stay anchored to long-term scientific integrity, transparency, and human benefit? For those of us working at the intersection of AI and social systems — Medicaid, behavioral health, aging services — these leadership shifts aren’t just gossip. They’re reminders of what happens when values and vision diverge. MettaHealth Partners we believe integrity at scale isn’t just a nice-to-have. It’s the only way AI can serve the public good. — 💡 If you lead teams or design systems in public health or social services, this is the moment to ask: Are we aligning our AI direction with our mission — or just following the market? #AIethics #PublicSectorAI #MedicaidInnovation #Leadership #Integrity #DigitalTransformation https://lnkd.in/gpqyQxbZ

19

1 Comment

🚀 Boosting Deep Learning Workflow with the Newly Launched Google Colab (Cloud GPU) + VS code (local system) Working on a medical MRI dataset project has really shown me the value of scalable compute.  * My Local System (MacBook Pro w/ MPS): ~ 𝐚𝐩𝐩𝐫𝐨𝐱 𝟑𝟓 𝐦𝐢𝐧𝐮𝐭𝐞𝐬 per epoch. 🐢  * Google Colab (A100 GPU): ~  𝐚𝐩𝐩𝐫𝐨𝐱 𝟒𝟎 𝐬𝐞𝐜𝐨𝐧𝐝𝐬 per epoch. ⚡ This setup not only saves hours of experimentation time but also makes the entire workflow smoother — coding locally in VS Code while leveraging cloud-grade GPUs. If you’re dealing with heavy models, large datasets, or medical imaging tasks, this combo is a game-changer. #DeepLearning #MedicalImaging #GoogleColab #A100 #VSCode #MachineLearning #OpenSource #MachineLearning #ArtificialIntelligence #ComputerVision #DeepLearning #MachineLearning #AppliedAI #AIProjects #OpenSource #RoadSafety #AutonomousVehicles #AIForGood #Python #TechInnovation #Hiring #AIEngineer #MLEngineer #SoftwareEngineer #AIResearch #CVPR #NeuralNetworks #StartupTech #ProductEngineering #BuildInPublic #India #US

43

1 Comment

Clinical AI explanations often tell us what influenced a prediction — but not why in the causal sense clinicians reason. In our latest proof-of-concept (https://lnkd.in/eeXhY6w6), we developed a framework that generates structured, evidence-linked causal explanations for clinical decision support. Using MIMIC-III and standardized ontologies (SNOMED CT, UMLS), the system maps AI predictions to domain causal graphs, assembles temporally ordered pathways, and links them to guidelines and literature. Across 100 acute myocardial infarction cases, the framework produced consistent causal structures (fidelity 0.85; completeness 1.00), with knowledge-graph integration expanding identifiable pathways. This is not a clinical tool — but a technical demonstration that causal explainability can be operationalized alongside attribution and similarity methods. Prospective clinician validation is the next critical step. #ClinicalAI #ExplainableAI #CausalInference #HealthInformatics https://lnkd.in/eeXhY6w6

21

1 Comment

🚀 Understanding Large Language Models (LLMs): A Theoretical Breakthrough A recent study bridges the gap between the impressive performance of LLMs and the lack of understanding around how they actually work. The authors propose a groundbreaking framework that models autoregressive LLMs as finite-state Markov chains. By leveraging the finite nature of an LLM’s vocabulary and context window, they prove that these models operate like Markov chains on a vast yet enumerable state space. Key Takeaways: 1️⃣ Stationary Distribution: LLMs reach a unique long-term equilibrium (stationary distribution). Convergence depends on the context window and temperature settings. 2️⃣ Temperature Matters: Higher temperatures improve convergence speed, enabling LLMs to explore state spaces more efficiently. 3️⃣ In-Context Learning (ICL) Laws: The team derived predictions showing ICL error scales with the context length. Experiments confirmed this "scaling law," demonstrating LLMs' ability to generalize effectively. 4️⃣ Outperforming Traditional Methods: For complex tasks with large state spaces, LLMs showed superior sample efficiency over traditional frequentist estimators. Why It Matters: This study provides the first rigorous theoretical foundation for LLMs, showing they are not just "black boxes" but efficient learners of Markov chains. These findings highlight how LLMs generalize to unseen data distributions, such as random walks, and explain their success in in-context learning. By treating LLMs as structured systems, this research advances our understanding of their behavior and sets the stage for more robust, transparent AI development. 

11

Bold prediction from Julien Chaumond , CTO at Hugging Face - 90% of LLMs will run locally in 18 months. At Lancashire Teaching Hospitals NHS Foundation Trust from where we lead the Lancashire and South Cumbria Secure Data Environment Programme , we have had the foresight to invest heavily in building both the expertise and the technology stack required to deploy agentic workflows locally as well as to undertake research in this area. 🚀🚀🚀 This is especially important for domains with highly sensitive data such as health. Unstructured text and certain types of imaging data are virtually impossible to fully de-identify. If we are to accelerate both research and deployment of AI and especially LLMs while preserving public trust in our work, the ability to run LLMs in air-gapped environments becomes important. Watch this space for another Lancs #dark_horse event. 🎠🤔🖥️🖥️🖥️ Grateful for the support of Stephen Dobson and Saeed Umar as well as Andy Wicks, CIO at University Hospitals of Morecambe Bay NHS Foundation Trust and Mark Singleton at Blackpool Teaching Hospitals NHS Foundation Trust. And my fantastic team of engineers and students. Mike Harding Alwin K Thomas Saurav Nair Sudhar Ian Farr Niko Möller-Grell Shihao Shenzhang Hevin Patel and Younus Rawat who I should get on LinkedIn. This is a regional effort and a first in this part of the NHS as far as I am aware. What is coming next should fundamentally change how we do agentic research/deployments at Lancs including our academic partners at Lancaster University and University of Lancashire . SCAN COMPUTERS (UK) LIMITED NVIDIA

11

3 Comments

I’ve uploaded a preprint and full reproducibility package to Zenodo and submitted the manuscript to arXiv (currently in moderation). Title: λQ = 0.336354: A Universal Information–Curvature Coupling Constant Derived from First Principles and Validated on Quantum Hardware The work includes (i) a first-principles derivation of λQ from four independent mathematical/physics inputs, and (ii) an experimental estimate from runs on IBM quantum hardware with full code and artifacts provided for reproduction. I’m inviting the scientific community to review, critique, and attempt independent replication. If you spot errors, have stronger alternatives, or can run follow-up checks on other platforms, I’d love to hear from you. Zenodo (paper + code + artifacts): https://lnkd.in/eg6vM-uM (arXiv submission is pending moderation; I’ll post the arXiv link as soon as it is live.) Welcome to Q-Bond Network! SIMPLER. SAFER. SURREAL. Q-Bond Network, where you're always 100. Founder, Kevin Miller Q-Bond Network DeSCI DAO, LLC Kevin@qbondnetwork.com Quantumblackswan@protonmail.com #QuantumComputing #QuantumInformation #quantph #QuantumHardware #Qiskit #QuantumErrorCorrection #OpenScience #ReproducibleResearch #ScientificComputing #InformationTheory #InformationGeometry #QuantumGravity #Holography #Langlands #NumberTheory #Cosmology #Astrophysics #TheoreticalPhysics IBM Quantum, Qiskit, Institute for Quantum Computing, Quanta Magazine

6

everyone wants a faster database ... https://lnkd.in/guMDU4MV Here's a summary: DBPlanBench: A Test Harness for LLM-based Query Optimization Researchers introduced DBPlanBench, a test harness that enables LLMs to optimize database query plans. Key contributions: - DBPlanBench: A novel test harness with a token-efficient serialization format and patch-based editing interface for LLMs to refine execution plans. - LLM Optimization: LLMs outperform DataFusion optimizer in query plan optimization. - Semantic Cardinality Estimation: LLMs use domain knowledge to identify structural optimizations (e.g., join reordering). - Small-to-Large Workflow: Optimizations at small scales transfer to larger scales via rule-based procedures. Code available on GitHub 😊

4

1 Comment

💥 Yann LeCun and Alex LeBrun has raised $1B to build AMI - Advanced Machine Intelligence. so what's a world model? 🌎 🔹 The thesis is simple - LLMs are trained on text about the world (which are inherently unreliable and does not represent the truth), whereas world models are trained on the world (through sensors, interaction, empirical data). 🔹 Yann has been a vocal advocate of world models over LLMs for a long time, the underlying solution to models that "understand" our world may be a combination of training data, and architecture (JEPA). 🔹 For healthcare this is particularly important - It is the difference between being a medical student who has read all the textbooks vs actually practicing as a doctor. 🔹 There are hundreds of papers demonstrating this "real-world" gap for LLMs - you may have seen the recent paper from nature demonstrated 52% undertriage performance in emergency cases. As someone who went from delivering clinical AI in Ambient voice tech/ AI scribes and now to building world models for health, it's great to see the top scientists in the world focusing on big problems that benefit humanity. -------------------------- 👋 Join Dev and Doc: AI for healthcare for the latest news, education, and deep dives in AI for healthcare!🤖👨🏻‍⚕️ LinkedIn Newsletter https://lnkd.in/eRXECsFC YT - https://lnkd.in/edFehUFy Spotify - https://lnkd.in/ei_kkY4b Apple- https://lnkd.in/ewJ8Yntt Substack- https://lnkd.in/e_XV4Pbz

48

11 Comments

GitHub’s latest shows how AI development is maturing from clever prompts to real engineering. By introducing agentic primitives and context engineering, GitHub outlines a framework for building reliable, reusable AI workflows—treating prompts like source code, context like data, and agents like modular programs. It’s a glimpse into how developers will soon design, test, and deploy AI systems with the same rigor as software. Key Points 1. The article argues that moving from ad-hoc prompt use of AI to reliable, repeatable engineering practices requires a structured framework. 2. The framework comprises three layers: Layer 1: Markdown prompt engineering — using structured Markdown (headers, lists, links) to guide prompts, activate roles, integrate tools, load context, and enforce validation. Layer 2: Agentic primitives — reusable, configurable building blocks (e.g., .instructions.md, .prompt.md, .chatmode.md, .spec.md, .memory.md, .context.md) that encode capabilities, workflows, context, and memory. Layer 3: Context engineering — managing what context the AI sees (session splitting, targeted instructions, memory files, chat-mode boundaries) so it stays focused and reliable rather than overwhelmed or distracted. 3. These layers combine into agentic workflows: end-to-end processes defined in .prompt.md files that orchestrate the primitives, load context, enforce validation gates (human checkpoints), and run either in an IDE or in CI/CD. 4. Practical how-to / checklist: The article provides steps to get started—write instructions, set up chat modes, build prompt templates, build spec templates, practice session splitting—and indicates architecture for instructions, chat modes, and workflows. Implications/Why It Matters For teams building AI-augmented systems, this offers a roadmap to bring structure to LLM or agent workflows rather than one-off prompts. It emphasizes that prompts plus context management plus tooling equals repeatability, not just “tell the model what to do”. It highlights the importance of context boundaries and role definitions (chat modes) to avoid misuse or cross-domain confusion. For environments with safety, regulatory, or enterprise demands, this approach supports auditability (instructions, validation gates, memory) and integration into existing CI/CD practices. Points to Note / Limitations The framework is fairly high-level; while it gives templates and examples, actual adoption will require discipline, tooling setup, and culture change. The article is from GitHub’s vantage point, so many references are tied to GitHub’s ecosystem (e.g., Copilot CLI, MCP servers, APM) which may map differently in other toolchains. https://lnkd.in/gvnnmdgm

3

Working with LLMs: From Usage to Orchestration (6-2) When Parallel Bottlenecks Disappear   After writing about role separation and handoff discipline, I noticed another effect: speed removes natural buffers. That’s where things get tricky.   Working with multiple LLMs removed bottlenecks I didn’t realize were protecting me.   One of the most striking differences between working with humans and working with LLMs is the disappearance of parallel bottlenecks.   Human collaboration inevitably includes pauses: scheduling, alignment, waiting, revision. Work flows partly in series.   With LLMs, that friction almost vanishes. I can run multiple agents in parallel and receive results immediately. Tasks that would take a week with a human team often take a single day.   The problem is not speed itself. The problem is what speed does to the human operator.   Immediate feedback removes natural stopping points. It becomes easier to stay immersed, push further, and postpone rest.   In retrospect, I realized that many pauses in human collaboration were not inefficiencies—they were protective buffers.   When those buffers disappear, productivity increases dramatically. So does the risk of cognitive overload and impaired judgment.   In environments without bottlenecks, slowing down must become an intentional act—not an accident.   Question: Has AI speed changed your sense of limits or stopping points?   #AIProductivity #CognitiveLoad #HumanFactors #FutureOfWork

2