Summary of "Algorithms to Live By | Brian Christian & Tom Griffiths | Talks at Google"

Summary of Algorithms to Live By

Brian Christian & Tom Griffiths | Talks at Google

This talk by Brian Christian and Tom Griffiths, authors of Algorithms to Live By, explores how fundamental computer science algorithms provide practical insights into everyday human decision-making and life problems. The central thesis is that many human dilemmas—ranging from dating and housing to organizing closets and managing time—mirror classic computational problems. By understanding these algorithms, we gain not only optimal strategies but also a new vocabulary and framework for thinking rationally about complex, uncertain situations.

Main Ideas and Concepts

1. Introduction to the Book and Authors

Brian Christian is a writer with a background in cognitive science and literature.
Tom Griffiths is a professor of psychology and cognitive science focusing on computational models of cognition.
Their book connects computer science algorithms to human life challenges, emphasizing computational thinking rather than coding skills.

2. Optimal Stopping Problem

Example: Searching for housing or a mate involves evaluating options sequentially and making an irrevocable choice.
The mathematically optimal strategy is the 37% rule:
- Explore the first 37% of options without committing.
- Then select the next option better than all previous ones.
Variations include:
- Rejection probability: If offers can be rejected, shorten the exploration phase.
- Recall: If you can return to previous options, extend exploration and revise thresholds accordingly.
- Full or partial information: Knowing or estimating the distribution of options changes the strategy.
Other examples of optimal stopping:
- Deciding when to sell an asset.
- When to quit a risky activity (like burglary).
- Finding a parking spot (threshold depends on occupancy rate).

3. Explore-Exploit Trade-off

Balancing trying new things (exploration) versus sticking with known good options (exploitation).
Applies to everyday choices like restaurants, music, social relationships, business ads, and clinical trials.
The Multi-Armed Bandit problem models this trade-off: choosing among slot machines with unknown payoffs.
Key insights:
- Exploration is most valuable early on because discoveries can be leveraged later.
- Exploitation becomes more important as the time horizon shortens.
- The concept of a “horizon” (how long you plan to continue) shapes the strategy.
Variants:
- Finite horizon: Use dynamic programming to calculate optimal pulls.
- Infinite horizon with discounting: Use Gittins index to decide which option to try next.
- Minimizing regret: Choose actions to minimize cumulative regret over time; “optimism in the face of uncertainty” (Upper Confidence Bound algorithm) is an effective heuristic.
Psychological implications:
- Babies explore extensively early in life when supported by caregivers.
- Older adults tend to exploit known pleasures, which is rational given limited remaining time.
- People tend to over-explore compared to the model when environments are non-stationary (payoffs change over time).

4. Caching and Cache Eviction

Problem: How to decide what to keep and what to discard when storage space is limited.
Computer science analogy: Managing fast (cache) vs. slow memory.
The optimal eviction algorithm (clairvoyant) evicts the item needed farthest in the future (requires future knowledge).
Practical algorithms:
- Random eviction
- First-in-first-out (FIFO)
- Least Recently Used (LRU) — empirically best and theoretically near-optimal.
Real-life applications:
- Organizing closets or offices by discarding least recently used items first.
- The Japanese economist Yukio Naguchi’s filing system mirrors LRU caching principles.
- Messy piles on desks can be seen as self-organizing lists implementing LRU.
Human memory analogy:
- Memory is not a fixed-size box but more like an infinitely large library.
- Forgetting is about retrieval difficulty, not loss of storage.
- Items more likely needed soon are easier to recall, matching temporal locality principles.

5. Broader Implications

Many human problems correspond to canonical computational problems, allowing us to apply proven optimal or near-optimal strategies.
Rationality is not about perfect, exhaustive computation but about trading off effort, error, and delay.
Approximation, probabilistic reasoning, and heuristics are essential parts of rational decision-making.
Computational thinking helps us rethink rationality and provides concrete, provable advice for everyday life.

Methodologies / Instructions Highlighted

Optimal Stopping (Secretary Problem / 37% Rule)

Determine total number of options (N).
Explore first ~37% of options without commitment.
Set the best observed option during exploration as a threshold.
Commit to the first subsequent option better than this threshold.
Adjust the percentage if rejection or recall is possible.

Explore-Exploit (Multi-Armed Bandit)

Identify the horizon (finite or infinite).
Early in the horizon, prioritize exploration to learn about options.
Later, prioritize exploitation of the best-known option.
Use algorithms like:
- Dynamic programming for finite horizon.
- Gittins index for discounted infinite horizon.
- Upper Confidence Bound (UCB) heuristic to balance optimism and uncertainty.
In changing environments, maintain some exploration to track changes.

Caching / Cache Eviction

Evict the item least likely to be needed soon (LRU heuristic).
Organize physical or digital items by usage recency.
For filing, place newly used items at the front to reflect recent use.
Accept that perfect clairvoyance is impossible; use near-optimal heuristics instead.

Speakers / Sources Featured

Brian Christian – Author, writer on cognitive science and algorithms.
Tom Griffiths – Professor of psychology and cognitive science at UC Berkeley.
Peter Norvig – Introduced the talk; Director of Research at Google.
Martin Gardner – Popularized the secretary problem in the 1960s.
Merill Flood – Mathematician who originated the secretary problem with a romantic context.
Michael Trick – Operations researcher who applied the secretary problem to dating.
Johannes Kepler – Astronomer who experienced a recall variant of the secretary problem.
Allison Gopnik – Psychologist who relates explore-exploit trade-offs to infancy.
Laura Carstensen – Psychologist studying aging and social exploitation.
John Gittins – Mathematician who developed the Gittins index for infinite-horizon bandit problems.
Jeff Bezos – Mentioned for regret minimization framework.
Morris Wilkes – Mathematician who contributed to caching algorithms.
Yukio Naguchi – Economist who devised a filing system based on LRU caching.
Robert Tarjan & Daniel Sleator – Proved theoretical guarantees for self-organizing lists.
John Anderson – Cognitive scientist who modeled memory as an infinite library.
Hermann Ebbinghaus – Early psychologist who studied memory recall patterns.
Eugene Lawler – Scheduling theorist quoted about book writing.

This talk blends rigorous mathematical theory with relatable life examples, showing how algorithmic thinking can improve decision-making, organization, and understanding of human cognition. The authors emphasize that these computational strategies are not just abstract concepts but practical tools for living better in an uncertain world.