AI Research
Giving Fable 5 a prime directive to mitigate sloppy domain language coinage.
Verbatim transcript where I hand Fable 5 a prime directive concerning ( ironically) its use of language. ( Edited for clarity)
Read more →Insights — category
AI Research
Notes from our internal R&D efforts
AI Research
Claude Fable 5: Past, Present and Future Stories of Informational Retrieval at Scale
Ask "who is the funded, empowered owner of this schema in five years, and what percentage of our actual query log requires a join?" If either answer is mumbling, you've found your line.
Read more →
AI Research
Betting against Claude Opus and winning big!
I propose double-or-nothing on our 100GB bet .. I bet we can fix this and get the Average Precision above .80 !!
Read more →
AI Research
The Naming Nemesis: Teams with reduced productivity, AI with reduced accuracy... all because of naming!?
Please note, in this article, the term 'names' is meant to include the cognitive / linguistic aspects of naming in the software realm. 'Concepts' is perhaps a better term... i.e., the behavioral, cultural, political, operational dimensions of our organizational conceptual frameworks.
Read more →
AI Research
Success Story: Cost-effective AI models answer data analytics, business intelligence and data science questions
See the results of a 2-year research project to use AI inference and NLP interfaces to answer complex data analytics and data science queries.
Read more →
AI Research
When building AI apps, coding is less and less the important part. Plan, Spec, (Code) Test, Document!
When developing with AI, it's what you do before and after coding that really matters... And how you manage LLM context space.
Read more →
AI Research
When developing AI based apps and services: Always know the answer before you ask the question.
This sounds like a contradiction but before coding an AI agentic pipeline to handle a certain type of question, you should already know representative answers. Don't rely on your inline agent to manufacture your answers on it's own. Known-good answers (e.g., 'goldens') are a key accelerator for AI development.
Read more →
AI Research
Inference-time AI Evaluation Methods: Loop through validation code, judge models and human monitors
Production AI systems need layered test-time evaluation driven by model confidence scores. Code-based business rules validate every response. Model judge panels catch edge cases when confidence dips. And for the highest-stakes scenarios, a virtual situation room of human experts and peer reviewers provides the ultimate quality gate. The key: not every answer needs the same scrutiny. Let confidence drive the escalation.
Read more →
AI Research
Why is psychometrics a white-hot human skill for the AI field!?🔥
Psychometrics is showing up in AI engineering because evaluating and shaping the behavior of large models now looks a lot like evaluating and shaping human abilities, traits, and biases—which is exactly what psychometrics was built for
Read more →
AI Research
AI Model Complexity Boundaries... (AKA the indefinitely ongoing need for Humans-in-the-Loop)
No matter how large a language model is and how well it’s trained there’s ‘laws of physics’ that limit the combinatorial complexity of the inference and retrieval that AI can accomplish
Read more →
AI Research
Parsing and classifying PDFs and unstructured docs: Python libraries vs Online Services
A practical guide to choosing between Python libraries like Marker, pypdf, pdfplumber, and Docling versus cloud services like LlamaParse, Mistral OCR, and Mathpix for parsing unstructured documents in AI applications.
Read more →
AI Research
Architecting the Tabular GRAPHRAG Paradigm for Enterprise Knowledge
As organizations move beyond experimental pilots into mission-critical deployments, a stark realization has emerged: the vector map is blurry... not accurate enough for many production transactional and operational apps.
Read more →
AI Research
Low vs. High Dimensions in Transformers
This high-dimensional space is where the model “thinks” or processes facts. By projecting data into a higher dimension, the model can “untangle” complex relationships and perform non-linear operations (using the activation function) that wouldn’t be possible in the compressed lower dimension.
Read more →
