Series

AI log

Learning AI from the ground up, in public, one entry at a time.

July 8, 2026

AI log, day 01: tokens, embeddings, and how a model actually generates text

Starting a day-by-day log of learning AI from the ground up. Day one covers the four ideas everything else sits on: tokens, embeddings, context windows, and how generation actually happens, revisited against a RAG pipeline I already built.

aiai-log

July 9, 2026

AI log, day 02: temperature and sampling, what turning the dial actually changes

Day two of the AI log picks up where generation left off: the temperature, top-p, and top-k knobs that decide which token gets picked next, what they actually do to the math, and why the wrong setting silently made a project less reliable.

aiai-log

July 10, 2026

AI log, day 03: a prompt is an API contract, not a magic spell

Day three of the AI log stops treating prompting as phrasing tricks and starts treating it as an interface: system prompt as config, few-shot examples as test cases, structured output as a return type, and why sharing one text channel for instructions and data is a security problem, not just a quality one.

aiai-log

July 11, 2026

AI log, day 04: fine-tuning is the last resort, not the first move

Day four of the AI log works through fine-tuning versus RAG versus just writing a better prompt, and argues the popular order is backwards: reach for a structured prompt first, RAG second, and fine-tuning only once both have genuinely run out of road.

aiai-log

July 13, 2026

AI log, day 05: evals, or how you tell better from felt-better

Day five of the AI log is about evals: the difference between a prompt change that actually made things better and one that just felt better on the single example you tried by hand, and the smallest honest setup for telling them apart.

aiai-log

July 14, 2026

AI log, day 06: an agent is a loop, not a personality

Day six of the AI log strips the buzzword off "agent": underneath, it is a plain loop of model, tool call, result, repeat, with the model choosing which tool to call, and most of the engineering is in the boring plumbing around that loop rather than the model itself.

aiai-log

July 14, 2026

AI log, day 07: a model call is a billed API call, not free compute

Day seven of the AI log looks at cost and latency through a backend engineer's cost-per-request instincts: tokens are the unit of both the bill and the wait, an agent loop multiplies both, and the context you re-send every turn makes a long loop cost more than it looks.

aiai-log

July 15, 2026

AI log, day 08: hallucination is a sampling problem, not a lying problem

Day eight of the AI log goes back to hallucination for the mechanical version day one only sketched: why next-token prediction has no "I don't know" state, why retrieval reduces but never eliminates it, and which fixes actually lower the rate versus which ones just make the wrong answer sound more confident.

aiai-log

July 15, 2026

AI log, day 09: prompt injection is what happens when data can talk

Day nine of the AI log covers prompt injection, the security failure specific to LLM-based systems: why data the model reads can override the instructions it was given, why it is a different bug than hallucination even though both share the same root cause, and why "just tell it not to listen to that" is not a real defense.

aiai-logsecurity

July 16, 2026

AI log, day 10: chatbot vs. agent vs. workflow, what genuinely separates the three

Day ten of the AI log pulls apart three words the current AI wave uses interchangeably: chatbot, agent, and workflow. Not a vibes-based distinction but a structural one, built on the model-call loop, tool use, and cost mechanics covered in the earlier entries, and what that structural line actually predicts about where each one breaks.

aiai-log

July 17, 2026

AI log, day 11: evaluating an agent, not just its final answer

Day eleven of the AI log returns to evals now that chatbot, workflow, and agent have separate definitions. A single-call eval checks an answer. An agent produces a trajectory, a sequence of decisions that can go wrong at step three and still land on a plausible-looking answer at step seven, so the eval has to check the path, not just the destination.

aiai-log

July 18, 2026

AI log, day 12: what changes when more than one model is involved

Day twelve of the AI log looks at what breaks once a second model enters the loop: an upstream hallucination becomes a downstream model's ground truth with no warning label, and every model-to-model hop multiplies cost, latency, and the same trust gap day nine already found no clean fix for.

aiai-log

July 19, 2026

AI log, day 13: guardrails are a UX problem, not just a safety one

Day thirteen of the AI log argues that a guardrail which blocks silently is indistinguishable, to the person on the other end of it, from the agent just being wrong or broken, and that the confirmation gate day nine and day eleven both argued for is a UX surface as much as it is a permission check.

aiai-log

July 20, 2026

AI log, day 14: context windows are memory, and memory is not free

Day fourteen of the AI log treats a long-running agent's context window as its entire memory, with nothing else backing it, and walks through why the two practical fixes, summarizing older turns and retrieving only what's relevant, both trade one failure mode for another instead of solving the problem.

aiai-log

July 21, 2026

AI log, day 15: giving an agent long-term memory is harder than bolting on a database

Day fifteen of the AI log looks at memory that is supposed to outlive a single conversation, and why storing a fact once and retrieving it later is the easy ten percent of the problem, not the whole thing.

aiai-log

July 22, 2026

AI log, day 16: the agent did not need a memory, it needed a source of truth

Day sixteen of the AI log finds a third shape of the memory question from days 14 and 15: some facts should never have been stored as memory at all, because a live system already owns them.

aiai-log

July 23, 2026

AI log, day 17: a tool schema is an API contract nothing enforces

Day seventeen of the AI log covers what happens when a tool definition changes shape: renaming a parameter or adding a required field does not throw a compile error, it just makes the model guess, and the failure looks like a wrong answer instead of a broken build.

aiai-log

July 29, 2026

AI log, day 18: retrying a tool call retries a side effect, not just an HTTP request

Day eighteen of the AI log covers what happens when an agent times out waiting on a tool call and retries it: if the tool has a side effect, a retry after an ambiguous failure can double a charge, an email, or a database row instead of just resending some bytes.

aiai-log