Era 15 / 15 · Generative & Agentic 2022 — now

Generative & Agentic

Models generate and act — and the feedback loop returns.

Beat 1 · Concrete

The output becomes the next input

One token at a time — each generated token is fed back in to predict the next.

Autoregressive generation A row of generated tokens, with an arrow feeding the newest token back into the model as its next input. the loop comes back ▍ fed back as next input
autoregressive · token stream

Beat 2 · Abstract

The agent loop is the same loop

Propose an action, the world responds, observe, repeat — and reward tunes the policy.

The agent feedback loop A closed circular loop: model proposes an action, the world responds, the agent observes, and the cycle repeats. A reward signal tunes the model toward the central goal. GOAL MODEL ACTION WORLD OBSERVE reward tunes policy
agent loop = cybernetic loop · Era 01 returns

Beat 3 · Interactive

Close the loop yourself

Step the loop or nudge the reward; watch behavior settle toward the goal.

Settling toward the goal An agent moving along a track from a far position (coral) toward a goal (teal), arriving at rest (chartreuse). GOAL at rest
control & communication — the arc closes
error — · step 0 · running

Drag inside the stage to move the goal · inject error, watch it resolve

The threads that tie back

2022 · ChatGPT

RLHF

Human preference becomes a reward signal; the policy is tuned by the same loop it runs in.

Generation

Diffusion

Image models denoise from pure noise back toward a sample — generation as iterated correction.

Agents

Tool use

The model proposes an action, a tool or the world responds, it observes, and acts again.

Era 01 · 1948

The loop returns

Wiener: "control and communication in the animal and the machine." The arc closes where it opened.