Milestones on the line · INTERFACE

Dialogue and human feedback

Models were taught to answer the way a person finds useful: plain words, on point, following the request. People compared candidate answers, and the model was tuned on those comparisons.

Training on human preference turned a language model into an interlocutor. Natural language became the universal interface to computing.

StatusPASSED
TypeMilestones on the line
Marker2022
What exists today6

Researched

2022

InstructGPT

Instruction tuning — the model starts following orders.

verified
2022

ChatGPT

The fastest technology adoption in history.

verified
2022

Constitutional AI

Training against a written set of principles instead of hand-labelling every answer.

verified
2023

GPT-4

Human-level results on professional exams move AI out of the toy category.

verified

In progress

сейчас

Sycophancy as a defect

A model trained to please tends to agree. Teaching it to push back is an open alignment problem.

NEW

Planned

впереди

Learning from outcomes

Replacing "did the answer feel good" with "did the result actually work".

NEW
Branches
Milestones on the line
To the line

Track the line as it moves

Once a week: which branches advanced, what unlocked, and what turned out to be overstated.