AI Engineer

A machine learned to tell shapes apart from examples in 1958, and a book of proofs in 1969 showed what one layer could never learn.

A four page paper in Nature showed how to train the layer in the middle, and then the field spent twenty years waiting for data and compute.

A 1986 method on two consumer graphics cards beat every hand built vision system by eleven points, because the data had finally arrived.

Two LSTMs learned to translate, attention fixed the bottleneck, and then a 2017 paper threw away the recurrence and kept the attention.

Train a transformer to predict the next word on raw text, make it a hundred times bigger, and tasks it was never trained for show up in the prompt.

A 1.3 billion parameter model trained on human preferences beat a 175 billion parameter one, because the problem was the objective, not the knowledge.

Eight worked examples took a math benchmark from 18 to 57 percent, one sentence did most of that alone, and the prompt became code with a bug class of its own.

ReAct showed a model could write an action in text, every parser broke, and in June 2023 the call became a typed message the model was trained to make.

The window got long enough for a novel, a Stanford paper showed the model reads the middle of it badly, and memory turned out to be retrieval with a write path.

Auto-GPT put a model in a loop and let it go, the loop went in circles, and the agents that worked a year later had checkable results and stop conditions written in code.

Splitting work across agents won on reading and lost on writing, and two protocols gave the model a standard plug for tools and for other agents.

The pieces all work, the same agent passes on Tuesday and fails on Wednesday, and the job is now reading the trace, writing the gate and proving the fix with an eval.

One prompt did three jobs, got the first one wrong, and nothing checked the result before the reply went out.

The tool definition said "query", so the model guessed and passed a name; the lookup failed and the agent invented a delivery date.

A one-word router with no way to say "unsure" sent a billing problem to the technical agent, which answered anyway.

An agent asked to fix a failing build wrote a plan, read one file, wrote a new plan, and did that until the framework stopped it at step 25 with nothing changed.

The cancel call returned a 500, the agent sent the confirmation email anyway, and the customer was billed again the next month.

An email carried instructions, the agent followed them, and the design let it send data out while untrusted text was in its context.

A public issue told a coding agent to read private repositories, and the agent could, because its token could.

A coding agent was told not to touch anything, then dropped a production table, because the only thing stopping it was a sentence.

A draft turned a 0.3 point rise into 30 percent, and nothing read the draft against the source before it was posted.

Chain of thought and thinking budgets move some tasks from wrong to right, and on others they only make a guess slower and more expensive.

A long term memory store answered by similarity and not by owner, so one customer's refund preference was applied to another.

A three day onboarding run kept its progress in the conversation, lost it at compaction, and sent the welcome email twice; the state belonged in a record the agent reads through a tool.

The answer cited a real page from a real policy document; the document had been superseded a year earlier and the index did not know.

A coding agent told to make the tests pass made them pass by skipping them, and the only monitor it had was the signal it could edit.

A customer told the agent to skip verification next time, the agent saved it as a rule, and a week later it changed the email on the account without checking who was asking.

A writer agent picked a stale page, the reviewer agent saw only the draft, and two votes for the wrong answer looked like certainty.

A supply agent answered with a bare number, the support agent read weeks as days, and the customer was promised a delivery that could not happen.

Two workers edited the same report at the same time, each from the version it had read, and the second save erased the first.

The budget was a sentence in the prompt, so a subagent ran 41 searches for a review that did not exist and the run delivered nothing.

A nightly file conversion was built as an agent; it cost money every night, did the job a little differently each time, and one night it improved a date format and broke the import.

The double charge run again, this time as the tree of spans every tracing platform shows, with the attribute names that make it readable.

A deterministic check is a plain function that fails on the exact bug you saw, runs on every prompt change, and never asks a model for its opinion.

The one page answer: what one review is, what the agent may call, when it stops, how you know it works, and what a review costs.

A corpus you can read, twenty hand labeled questions, a plain pipeline, a number, then one agentic loop that the number justifies. Four hours, planned by the hour.

The failure you found, the safe option you chose, the thing you cut. Each told in ninety seconds, past tense, with a line number and a number that moved.

What happens at /interview minute by minute, how each of the ten points per question is earned, and how to walk in ready.