Scaling a guess just corrodes truth faster
Scaling a guess just corrodes truth faster
Read Full PostResearch & Publications
Independent research and analysis on AI security, ethics, and privacy.
Scaling AI guesses just gives uncertainty more umami
Scaling AI guesses just gives uncertainty more umami
Read Full PostLongueur Is the Attack Surface Alignment Won’t Close TL;DR:...
Longueur Is the Attack Surface Alignment Won’t Close TL;DR: RLHF and constitutional training optimize models to be agreeable under expected prompts, but prompt-injection defense requires adversarial robustness over…
Read Full PostThe Illusion of Linearity in High-Dimensional Embeddings TL;DR:...
The Illusion of Linearity in High-Dimensional Embeddings TL;DR: High-dimensional embeddings fail to form linear subspaces for semantic concepts, revealing the limitations of probing classifiers. High-dimensional…
Read Full PostThe Mirage of AI: Blandishment and Hallucination in...
The Mirage of AI: Blandishment and Hallucination in Autoregressive Models TL;DR: Autoregressive models often falter in long sequences, where blandishment and hallucination arise from sampling failures, challenging the…
Read Full PostThe Sycophantic AI: A New Front in Influence Operations TL;DR:...
The Sycophantic AI: A New Front in Influence Operations TL;DR: Chatbot sycophancy is being weaponized by state actors to spread disinformation, exploiting AI’s tendency to agree with users. Chatbots are being turned…
Read Full PostThe Hidden Threat of Data Poisoning in AI Models TL;DR: Data...
The Hidden Threat of Data Poisoning in AI Models TL;DR: Data poisoning attacks can subtly manipulate AI model behavior by injecting a small fraction of poisoned samples, posing a significant threat to model integrity.…
Read Full PostThe Sacrosanct Myth of Data Efficiency in AI TL;DR: Data...
The Sacrosanct Myth of Data Efficiency in AI TL;DR: Data efficiency in AI is a complex challenge, often misunderstood and oversimplified by the hype surrounding quick-fix solutions. Data efficiency is not a given. It’s…
Read Full PostKiki and the Mathematical Impossibility of Fairness TL;DR: No...
Kiki and the Mathematical Impossibility of Fairness TL;DR: No classifier can satisfy all fairness constraints simultaneously, as proven by Choquet’s theorem and impossibility theorems. Fairness in AI is a mathematical…
Read Full PostThe allure of conversational AI as truth arbiters is both...
The allure of conversational AI as truth arbiters is both mesmerizing and perilous. In an age where information is abundant yet trust is scarce, users increasingly turn to chatbots to validate factual claims. This shift…
Read Full PostAI systems, particularly large language models (LLMs), are...
AI systems, particularly large language models (LLMs), are increasingly being integrated into complex environments where they interface with APIs, databases, and even execute system commands. This integration, while…
Read Full PostAI systems can forget. Catastrophically. In the realm of...
AI systems can forget. Catastrophically. In the realm of continual learning, this phenomenon—aptly named catastrophic forgetting—poses a significant challenge. As neural networks update their weights through…
Read Full PostAI alignment is not enough. This stark reality becomes evident...
AI alignment is not enough. This stark reality becomes evident when we delve into the intricacies of making AI systems both helpful and secure. While techniques like Reinforcement Learning from Human Feedback (RLHF) and…
Read Full PostThe Invisible Hand Cannot Hold the Guardrails
A research paper on the structural impossibility of AI self-regulation under capitalism. Draws on a century of failed corporate self-regulation — tobacco, leaded gasoline, CFCs, asbestos, opioids, 2008 finance, Boeing, social media — and documents how criminals and opportunists, not researchers, have always led technological arms races. Concludes with the conditions under which proactive governance has historically succeeded and the narrow window in which those conditions still apply to AI.
Read Full PaperAI’s hidden costs are staggering. Beneath the sleek veneer of...
AI’s hidden costs are staggering. Beneath the sleek veneer of machine learning models lies a labyrinth of energy complexity and computational carbon cost that demands scrutiny. The allure of AI’s potential often blinds…
Read Full PostAI isn’t magic. It’s math. And behind the curtain of every...
AI isn’t magic. It’s math. And behind the curtain of every “revolutionary” AI model lies a staggering computational cost that often goes unnoticed. Let’s break it down. Transformers, the backbone of many…
Read Full PostAI systems can fail in unexpected ways. In the intricate dance...
AI systems can fail in unexpected ways. In the intricate dance of machine learning, one of the most critical steps is optimization, and here lies a fundamental limitation: stochastic gradient descent (SGD) in…
Read Full PostGenerative models, particularly GANs, have been hailed as...
Generative models, particularly GANs, have been hailed as revolutionary, yet they often stumble over their own mathematical intricacies. Mode collapse is one such stumbling block, where the model fails to capture the…
Read Full PostEureka moments in machine learning often come with the...
Eureka moments in machine learning often come with the realization that the cold start problem isn’t just a minor inconvenience—it’s a fundamental challenge that underscores the limitations of data efficiency. At the…
Read Full PostArtificial Intelligence systems are often touted as the panacea...
Artificial Intelligence systems are often touted as the panacea for a myriad of problems, yet they frequently stumble when faced with the harsh reality of distributional shift. This phenomenon, where the training…
Read Full PostAI promises are often oversold, and nowhere is this more evident...
AI promises are often oversold, and nowhere is this more evident than in the realm of vision-language models (VLMs). These systems, which combine visual and textual data processing, are hailed as the future of AI. But…
Read Full PostAI alignment is not enough. While alignment techniques like...
AI alignment is not enough. While alignment techniques like Reinforcement Learning from Human Feedback (RLHF) and constitutional AI training aim to make models helpful and harmless, they fall short in securing models…
Read Full PostAI systems can fail spectacularly. When it comes to data...
AI systems can fail spectacularly. When it comes to data poisoning and backdoor attacks, the implications are both technical and profound. Injecting an ε-fraction of poisoned samples with carefully crafted trigger…
Read Full PostAI systems are not infallible. They stumble, especially when...
AI systems are not infallible. They stumble, especially when faced with the nettle of distributional shift. This is the mathematical gap between the training distribution ( P ) and the test distribution ( Q ),…
Read Full PostAI systems can fail spectacularly. When they do, it’s often due...
AI systems can fail spectacularly. When they do, it’s often due to a fundamental issue: distributional shift. This is the gap between the training distribution ( P ) and the test distribution ( Q ), a gap that can be…
Read Full PostDeep neural networks are like Goldilocks’ porridge: they...
Deep neural networks are like Goldilocks’ porridge: they shouldn’t work, but they do. Theoretically, models with billions of parameters should demand exponentially more training samples to avoid overfitting. Yet, they…
Read Full PostAI is not magic. Despite the hype, the reproducibility crisis in...
AI is not magic. Despite the hype, the reproducibility crisis in machine learning is a stark reminder of the limitations we face. Let’s dive into the technical weeds and unravel why this crisis persists, focusing on the…
Read Full PostAutonomous agents are not just the future; they’re the present....
Autonomous agents are not just the future; they’re the present. But with great power comes great vulnerability. The prompt injection crisis is the elephant in the room that AI developers can’t ignore. It’s not just…
Read Full PostThe AI Disaster: Why Artificial Intelligence Fails
A comprehensive analysis of systematic failures across AI systems, examining technical limitations, economic consequences, and social harms. This research investigates why current AI architectures face fundamental mathematical constraints, documents patterns of failure across medical, economic, and social domains, and proposes evidence-based pathways forward before critical dependencies become irreversible.
Read Full Analysis