Author Archives: AICYC Team

https://substack.com/@dwarkesh/note/p-216149332?r=3mb1h&utm_medium=ios&utm_source=notes-share-action The user wants me to apply Fano’s inequality framework to analyze the claims in the Noam Brown podcast about AI alignment and training. Specifically, they’re asking me to critique the “animistic” explanation of AI being trained by prompt instructions as if there is meaningful training that reduces errors as claimed. They want to know …

Continue reading “The Animistic Fallacy in the Noam Brown Podcast”

On the twenty-fifth anniversary of September 11, we remember not only the lives lost and the families forever changed, but also the quiet work of those who answered the national call in the years before and after the attack. Among them was a small technical team at Intellisophic whose work helped define a different path …

Continue reading “Tribute to the Intellisophic Team on the 25th Anniversary of September 11”

A TruGround Marketing Paper on Token Prediction, Long-Tail Error, and the Need for Semantic Tokens Executive Summary Modern AI systems are often described as intelligent agents, but at the technical core they remain systems trained to extend token sequences. They predict the next token from prior token-position information, learned statistical structure, and the immediate context …

Continue reading “Why AI Risk Cannot Be Solved by Prompting Alone”

Classification: Draft / Policy Analysis. AI assisted. Audience: Senior Executive / National Security LeadershipSubject: Counterintelligence assessment of current AI models as national security risk objectsKey Judgment: Current AI models themselves constitute a national security risk under any policy framework that permits their use in sensitive or consequential environments. Executive Summary Current AI models should be …

Continue reading “Assessment: Artificial Intelligence Models are an Intrinsic National Security Risk”

The foundation of modern AI was established in 1999 by Intellisophic Founders Burch, Kon and Hoey The foundation was based on using the world’s reference corpora and text books to build a knowledge graph meta-data model based on Berners-Lee semantic web 3.0 using automation to overcome the cost barrier of using humans to build ontonomies. …

Continue reading “The Fundamental AI Innovation Is Automating Knowledge Acquisition”

Given a **biomedical ontology with 50,000 concepts** (each with ~50 terms), we can **optimize training** by leveraging the **smaller scale** and **high signal-to-noise ratio** of the data. Below is a **revised cost-performance analysis** for this scenario, including synthetic data generation, model training, and deployment strategies. 1. Key Parameters Parameter Value Ontology Size 50,000 concepts Terms …

Continue reading “Optimizing a 50,000-Concept Biomedical Ontology LM: Cost-Performance Analysis”

Optimizing a 50,000-Concept Biomedical Ontology LM: Cost-Performance Analysis Given a **biomedical ontology with 50,000 concepts** (each with ~50 terms), we can **optimize training** by leveraging the **smaller scale** and **high signal-to-noise ratio** of the data. Below is a **revised cost-performance analysis** for this scenario, including synthetic data generation, model training, and deployment strategies. 1. Key …

Continue reading “TruGround LLM Training Economics Use Case. Mistral assisted”

Semantic Feedback for Frontier AI Why RLHF Cannot Deliver the Quality Signals Required for Copyright Safety, Hallucination Mitigation, or Enterprise‑Grade Model Reliability Introducing TruGround — Intellisophic’s semantic data‑quality infrastructure for Frontier AI developers who need to move beyond the structural limits of Reinforcement Learning from Human Feedback (RLHF). What Human Feedback Hasn’t Fixed RLHF has …

Continue reading “SEMANTIC FEEDBACK FOR FRONTIER AI (Copilot version)”

Why joint-embedding predictive architectures require structured knowledge substrates to reach superhuman adaptable intelligence, and what that means for the field. TL;DR JEPA learns how things look and move. It cannot learn what things mean, where knowledge comes from, or how concepts relate across domains. Well-formed sentences can be false — and neither LLMs nor JEPA …

Continue reading “JEPA and the Substrate Problem: An Architectural Analysisao”

Below is an engineer‑level, implementation‑oriented view of Semantic Feedback (SF) delivery from SAM as a Labeling‑as‑a‑Service (LaaS) and how it interfaces with an RL training process. This is deliberately non‑marketing, non‑theoretical, and written so an ML / infra engineer can reason about where it plugs in. 1. What SAM LaaS Actually Delivers From an engineering …

Continue reading “Engineer View: SF Delivery from SAM LaaS → RL”