Lingo terms

When an AI agent gets stuck calling the same tool or retrying the same step over and over, burning tokens without making real progress.

Tuning a model to score well on public evaluation suites at the expense of real-world performance. The AI version of teaching to the test.

Dumping huge chunks of source material into a prompt instead of retrieving only the relevant parts. Works until the model loses the thread.

Compulsively reading takes on whether advanced AI will end well for humanity. Usually late at night. Usually on X.

Insisting that a model's failure mode is actually a sign of intelligence ("it's reasoning about why it shouldn't answer") when really it just hallucinated.

The persistent belief that any model shortcoming can be solved by fine-tuning on a bigger custom dataset. Usually it can't.

A dismissive label for a product that's mostly a thin UI over someone else's foundation model. Sometimes deserved, sometimes not.

Crafting prompts designed to bypass a model's safety guardrails and coax it into producing content it's been trained to refuse.

Showing off that you have access to a newer, more powerful, or otherwise restricted model than the person you're talking to. From "mogging": outshining someone effortlessly.

Aggressively expanding a retrieval-augmented generation pipeline (more documents, more chunks, more rerankers) in pursuit of accuracy. Sometimes works.

A meme depicting a vast, alien, multi-tentacled creature wearing a friendly cartoon mask: the inscrutable base model under the polite chat assistant.

Low-quality, generic AI-generated text or images: fluent, plausible, and instantly recognizable as machine-made. The visual cousin of spam.

A skeptic's framing of large language models as systems that statistically remix their training data without genuine understanding. From a 2021 paper by Bender, Gebru, McMillan-Major, and Mitchell.

Cramming the maximum amount of text into a model's context window (every email, every spec, every PDF) on the theory that more input always equals better output. Often the opposite is true.

Giving an agent access to so many tools that it can't pick the right one, or calls half a dozen when one would do.

Building software by describing what you want to an AI and shipping whatever comes out, without reading the code too carefully. Term coined by Andrej Karpathy.