Book · The room of machines

What are machines learning from us at 3 AM?

Two Point Singularity

Coherence precedes intelligence · Venkat Rao · Two Point Singularity Press · Kindle · ₹599

Machines learn our stories. Coherence decides which ones we believe.

Buy on Amazon.in · ₹599Open the interactive page

The book in five minutes

Today’s most capable AI systems are shaped by human feedback: people rating answers as better or worse. This book starts from one plain observation. The same person, rested and steady at 10 AM, judges differently when exhausted and alone at 3 AM.

It writes that observation as h = v + ε(state): the feedback a person gives is their real value plus an error that depends on their state. Current training treats that error as zero. The book proposes Coherence-Weighted Human Feedback, which takes the state of the person into account.

Five people carry the argument: Maya in Mumbai, Amara in Lagos, Russ in Ohio, Haruto in Osaka and Camila in São Paulo. Through them, the book argues that a coherence singularity has to come before the technological one.

Key takeaways

• AI alignment is also a human coherence problem, not only a technical one.
• Human feedback carries the state of the person who gave it, not just their values.
• Weighting feedback by coherence is a concrete proposal that can be tested.
• The order matters: people need to become coherent before machines out-think them.

Who it is for

Read this if you build, use or worry about AI, and suspect the problem is not only in the code.

You may never again think of AI as something separate from the people who teach it.

The opening

Maya Iyer reviews model outputs for a living. Amara Okonkwo rates them at the end of a night shift. Neither is wrong. Neither is the same person at 3 AM as at 10 AM. The machine is learning from both.

What this book does not claim

CWHF is a hypothesis, pre-registered for testing at OSF.IO/W7NYH. The book does not claim it has been proved.

Questions

What is Two Point Singularity about?

Today’s most capable AI systems are shaped by human feedback: people rating answers as better or worse. This book starts from one plain observation. The same person, rested and steady at 10 AM, judges differently when exhausted and alone at 3 AM. It writes that observation as h = v + ε(state): the feedback a person gives is their real value plus an error that depends on their state. Current training treats that error as zero. The book proposes Coherence-Weighted Human Feedback, which takes the state of the person into account.

Who should read Two Point Singularity?

Read this if you build, use or worry about AI, and suspect the problem is not only in the code.

What does Two Point Singularity not claim?

CWHF is a hypothesis, pre-registered for testing at OSF.IO/W7NYH. The book does not claim it has been proved.

Editions

Next room

COHERENCE: What actually holds a life together?