Learning to Inject: When Prompt Injection Becomes an Optimization Problem
AutoInject shows why prompt injection should be tested as an adaptive optimization problem, not merely as a list of hand-written attack templates.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Speculation, But With Standards: Training Draft Models That Actually Get Accepted
VSD shows why speculative decoding improves when draft models are trained for accepted paths, not merely probable tokens.
Tokens, Watts, and Waste: The Hidden Energy Bill of LLM Inference
A mechanism-first reading of why LLM inference energy is shaped by prefill, decoding, prompt length, and unnecessary generation—not merely model size.
Ultra‑Sparse Embeddings Without Apology
CSRv2 shows that ultra-sparse embeddings fail less because sparsity is impossible, and more because we have been training them badly.
When Words Start Walking: Rethinking Semantic Search Beyond Averages
A comparison-based reading of why Word Mover’s Distance with GloVe outperforms centroid-style semantic search in statement-level retrieval, and where that lesson actually applies in business systems.
Benchmarks Lie, Rooms Don’t: Why Embodied AI Fails the Moment It Enters Your House
A mechanism-first reading of TEA, an in-situ task-generation framework showing why embodied AI needs environment-specific evaluation before deployment.
Beyond Cosine: When Order Beats Angle in Embedding Similarity
A business-focused reading of recos, a Rearrangement Inequality-based similarity metric that tests whether embedding similarity should care about ordered structure, not only vector angle.
First Proofs, No Training Wheels
Why unpublished research lemmas expose the difference between fluent mathematical performance and proof-grade AI reasoning.
Hallucination-Resistant Security Planning: When LLMs Learn to Say No
A mechanism-first reading of how abstention, lookahead, and feedback turn LLM incident-response planning from fluent guessing into calibrated decision support.
When AI Forgets on Purpose: Why Memorization Is the Real Bottleneck
A mechanism-first analysis of how attention sinks can reveal and suppress harmful learning during LLM fine-tuning.