Packs
Talks on a subject, with a suggested watch order and a note on why each one is included.
Coding agents look capable in a fresh repository and come unstuck in a ten-year-old one.
A production agent can give the right answer and still refund the same customer twice, act on a stale record or quietly finish half a job.
A robot policy can understand the scene and still fail because its training data is thin, its controller is late or its motors cannot keep up.
When an agent forgets its plan, edits the wrong file or declares victory without changing anything, the model is only one suspect.
The chat box makes users package their work into prompts, wait, inspect the answer and try again.
A software factory can build quickly and still fail if nobody decides what is worth building or retains enough understanding to support it after release.
An agent can remember your name and still repeat yesterday's mistake.
High-risk AI is not one field, and its controls cannot be copied whole from one industry to another.
A personal agent may need your messages, calendar and accounts, which makes sending every thought and action through a cloud provider a hard bargain.
Training data is not a pile of tokens waiting to become a model.
More context can make an agent worse.
A fluent RAG answer can still be wrong because the right passage never reached the model.
An agent can return a reassuring answer after it has already leaked a credential, crossed into another user's data or deleted something real.
In one financial tool-use task, a trained 4 billion parameter model outperforms a 235 billion parameter baseline, but only because the team knows which behavior to reward.
A voice agent can have a strong model and still fail because it talks over someone, leaves a dead pause or hands a caller to a person with no useful context.
A model can answer correctly and still be unusable once real prompts, concurrent users and a GPU bill arrive.
An MCP server that answers one tool call on your laptop has proved little.
An API tells an agent what it can call, but not how your team gets a job right.
An eval set assembled before launch can tell you whether a change helped on the cases you expected.
A benchmark can rank models cleanly while rewarding the wrong work.
Image generators are impressive until you need the same character twice, a product to keep its shape, or an edit to work for millions of people.
Video generation stops being useful when every clip demands a long wait, another prompt and a manual rescue in the editor.