Journal7 entries
Writing on
machine
intelligence
Investigating the intersection of computational theory and real-world application. Technical notes on language, learning systems, agent architectures and the hardware they have to run on.
- NLP8 min read
Latest
Tokenization Decides What Your Model Can Count
The vocabulary you never think about sets a hard ceiling on arithmetic, multilingual cost, and how well Urdu is treated.
Read entry
- Machine Learning7 min read
Your Classifier Is Measuring Your Sampling, Not the World
A ticket classifier at 94% offline and 71% live was not implemented wrong. The test set was drawn from the same snapshot as the training set.
Read entry
- Agentic AI9 min read
Agentic AI and the Compounding Cost of a Wrong Turn
Per-step accuracy is a seductive metric. At 95% per step, a twenty-step task succeeds only about a third of the time.
Read entry
- Agents8 min read
Multi-Agent Systems Are a Communication Problem
Adding agents does not add intelligence. It adds edges to a graph, and every edge is a place where context goes missing.
Read entry
- LLM Developments7 min read
Prompts Are Production Code and Should Be Versioned Like It
A string that determines system behaviour, can be edited by anyone, and has no history is not configuration. It is an outage waiting for a calendar slot.
Read entry
- Edge Computing8 min read
Edge Inference Is a Memory Bandwidth Problem
Teams optimise FLOPs and wonder why the model is still slow on device. For single-batch generation, arithmetic was never the bottleneck.
Read entry
- Physical AI9 min read
Physical AI Cannot Retry
Software agents fail into a log file. Embodied ones fail into the world, where there is no exception handler and no undo.
Read entry
Research
Notes from research
to production.
New entries cover NLP, machine learning, agentic systems, LLM developments, edge inference and physical AI — written from what actually held up in deployment.