Field notes from Jackdaw
Plain-spoken pieces on retrieval, memory, evaluation, and the models underneath—written for people building with AI, not selling it.
A memory system can post an excellent recall score and still be unsafe to rely on. The missing evaluation is the full path from what gets remembered to what happens because of it.
Read the essayAsk a frontier model how many R's are in "strawberry" and it has historically gotten it wrong. The reason isn't reasoning. It's that the model was never shown the letters in the first place.
A hallucination isn't a malfunction or a separate mode. It's the same machinery that produces every correct answer, running on a question where plausible and true happen to diverge.
A reasoning model isn't a different kind of intelligence. It's a model spending more compute by writing out its work — which genuinely helps on hard problems and is pure waste on easy ones.
A model can have hundreds of billions of parameters and run at the cost of a model a fraction the size. The trick is that most of those parameters sit idle on any given token.
Quantization is the most common cost lever in AI serving, and it means "the same model" can quietly behave differently depending on how it was compressed. Here's the mechanism and the trade.
A small model can approach a much larger one's quality at a fraction of the cost — by learning to imitate it. Here's how that works, why it's everywhere, and what it can't do.
A language model is a stateless function. Everything that feels like memory is the application replaying the transcript on every call. Here is what that costs and where it breaks.
If you've heard both "we use RAG" and "RAG is dead" in the same week, it's because the phrase covers systems that share almost nothing. Here is what's actually inside each one.
"Memory" is one word covering at least three different systems. Knowing which one you're using explains why the assistant remembers some things perfectly and drops others completely.
"It remembered my name" is the easy case and a misleading success metric. Here is the taxonomy of what's actually hard, and what the benchmarks built to test it reveal.
The specific model that wins changes every quarter. The forces that shape the choice — how these models are built, what they cost to run, and where they break — don't.
Every assistant builds its own private model of you, and none can read the others. The result is that you are the integration layer between your AIs. Here is the shape of the fix and the part of it that's now standardized.
