Notes
Two minutes. Then you know a little more.
Short observations on building, evaluating, and operating AI systems.
Topics
16 notes
GPU utilization is the wrong question
Seeing 95% GPU utilization feels reassuring. The expensive hardware is busy, so the workload must be well optimized.
Retries are a correctness problem
A retry sounds like reliability plumbing until the operation being retried has consequences.
The evaluation set becomes part of the product
The first evaluation set for an AI system is usually built before anyone knows how the system will fail.
Agents need budgets
An agent loop needs more than a stopping condition. It needs a budget.
False positives are an architecture decision
A similarity threshold looks like a model parameter until the system gives it consequences.
When the model stops being the bottleneck
A machine learning project can change character without changing its model.
Speaker embeddings
Speaker embeddings make voice matching look almost trivial.
Direct Preference Optimization (DPO)
Suppose you show someone two responses to the same prompt and ask which one they prefer.
Prompt injection and guardrails
Prompt injection becomes much more consequential once a model can do something.
Serving LLMs in production
LLM inference is partly a scheduling and memory problem disguised as a model call.
Structured outputs
Free-form language is an excellent interface for people. It is usually a bad interface between software components.
RAG failure modes
RAG failures become easier to debug once “RAG quality” stops being treated as one thing.
Evaluating LLM products
The useful unit of evaluation for an LLM product is usually not the completion. It is the task.
Tool-calling agents
Function calling gives a model an interface to software. An agent appears when you put that capability inside a control loop.
A cache key is part of the AI system
Caching looks deceptively easy around AI workloads.
Authorization belongs before retrieval
A RAG system can produce a perfectly grounded answer and still have a severe security bug.