liveBook · Manning

Continuous improvement

As explained in Machines that Think

A feedback loop is a process in which a system's outputs or effects alter later inputs or data, thereby influencing subsequent system behavior.

Continuous improvement can arise through a feedback loop: a system’s outputs or effects influence later inputs, changing subsequent behavior. Repeated cycles can reinforce useful capabilities, but they can also amplify harmful patterns. The direction of change depends on what is measured and fed back into the system. section 4.4.2 section 7.3.5

In a mixture-of-experts model, a gating network selects experts for each token. Selected experts receive a learning signal; unselected experts do not receive that token’s direct expert update. An expert repeatedly assigned a type of pattern can improve at it, making the gate more likely to select it for similar tokens. Specialization emerges through repeated assignment and reinforcement rather than explicit labels. section 4.4.2

This loop needs balance. If routing concentrates too heavily on a few experts, others receive too little data to learn useful behavior. If routing is too even, experts encounter mixed patterns and specialization weakens. section 4.4.2

When feedback reinforces bias

A predictive policing system trained on past policing data may direct more attention to neighborhoods already subject to heavy policing. More surveillance produces more recorded incidents there, which can appear to confirm the original prediction and reinforce the cycle. section 7.3.5

AI development can also feed improvements into later AI systems. Automated machine learning uses algorithms to tune hyperparameters, select features, and compare configurations; neural architecture search extends this approach to network structures. These techniques optimize parts of development without making AI research autonomous. section 8.4.4

AI-assisted AI development

AlphaEvolve uses language models in an evolutionary coding process to generate and modify programs that are evaluated automatically. Its reported improvements include parts of data center operations, chip design, and AI training processes, including components related to training its underlying models. section 8.4.4

Figure 8.8 AI-assisted improvement of AI. AI can help improve future AI by contributing to better model architectures, software, computational primitives, and specialized hardware, creating a feedback loop in which research advances are fed back into later systems.

Trace how improvements to architectures, software, computational primitives, and hardware can feed into later AI systems. section 8.4.4

How Machines that Think covers it

How deep it goes
Explains expert routing and its balance requirements, then surveys automated approaches to AI development without presenting a full procedure.
Examples
Illustrates reinforcing cycles with predictive policing, expert specialization, and AlphaEvolve’s automatically evaluated programs.
What you can do after
Evaluate feedback loops by distinguishing useful reinforcement from cycles that amplify bias or starve components of training data.
How it connects
Connects feedback loops to mixture-of-experts models, algorithmic bias, automated machine learning, and neural architecture search.

Also explained in

Compare how 43 books explain continuous improvement →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machines that Think. The section links point to the parts of the book it draws on.