The book
Book · October 2026 · about 46,000 words
Borrowed Tulan
The AI Alignment Problem, Current Research, and Madhyasth Darshan
For readers who know how language models are trained, or are willing to learn, and who have studied the darshan or are willing to learn its terms. Every term is glossed.
The book starts from one claim: alignment is a valuation problem, and AI reproduces the choosing, analysing and imaging side of a mind while borrowing its weighing, तुलन, from the people who write, rate and evaluate it. Part I explains how the field sees the problem in late 2026: how values enter a model, what Claude's constitution and OpenAI's Model Spec say, what has gone wrong, what interpretability can see, and how philosophers argue about values. Part II reads the same material through Madhyasth Darshan, separating what the texts say from our extrapolation to machines. Part III turns it into a research programme, a practice for enterprises, and an honest list of objections.
All 17 chapters
- 1. What “alignment” means, and why it is a valuation problem. Separates specification, generalisation and values, and argues the third is a question of weighing.
- 2. How a model gets its values. Each stage of training, and fifteen places where someone's weighing enters the model.
- 3. Two written value documents. Anthropic's 2026 constitution and OpenAI's Model Spec, what they say and how they are used.
- 4. What goes wrong, 2016–2026. Sycophancy, reward hacking, sleeper agents, alignment faking, scheming, emergent misalignment: models learn characters, not rules.
- 5. Looking inside. Interpretability, and whether values have an address inside a network.
- 6. What are values?. The field's philosophical map: preferences, pluralism, moral uncertainty, virtue, and alignment as politics.
- 7. The open questions. Ten problems the field admits it has not solved, and where the darshan has something specific to say.
- 8. Jeevan and its ten activities. The darshan's account of how a decision forms, close to the texts.
- 9. What an AI system is, in the darshan's terms. An explicit extrapolation: a यंत्र with no jeevan, whose activities are moved by borrowed evaluation.
- 10. Reading the training signal as tulan. RLHF, reward hacking, deception and evaluation awareness read as प्रिय–हित–लाभ weighing.
- 11. A universal constitution?. The darshan's layered value system set beside Claude's constitution, point by point.
- 12. The writer and the rater. Why alignment is bounded by the jagriti of the people in the loop, and what procedure can carry.
- 13. Vigyan and vivek. Capability and purpose, oversight, and the order machines enter.
- 14. A research programme. Concrete experiments ordered by cost, each with a hypothesis, method, measure and way it could fail.
- 15. In the enterprise. Why the KPI, not the ethics policy, is the operative weighing in a company, and five changes.
- 16. Objections and limits. The ten strongest objections, the best answer to each, and what remains unresolved.
- 17. A roadmap. What to read, write, build and test first, and how to speak to the alignment field.
The primer
Primer · October 2026 · about 14,000 words
Learning to Weigh
A primer on RLHF and preference training, read alongside Borrowed Tulan
For readers who know only that language models are trained on lots of text. No mathematics needed: formulas come with small worked numbers and can be skipped.
How language models are trained on judgements, built from the ground up: probabilities, reward models, reinforcement learning, PPO, the KL leash, DPO, AI feedback, and the verifiable rewards behind reasoning models. One running example ties it together: Ravi has quit their job to open a tea shop and asks an AI for an honest opinion, and every method is shown acting on the same two replies, one flattering and one kindly truthful. Each chapter ends with a short note connecting the mechanism to the darshan. The How AI learns page is a short version.
All 14 chapters
- 1. The problem RLHF solves. Why a predictor is not an assistant, and why people judging answers was the next step.
- 2. A language model as a policy. Probabilities, sampling, and what “nudging the weights” means.
- 3. Learning by imitation. Supervised fine-tuning: what it can and cannot achieve, and who writes the examples now.
- 4. Comparisons and the reward model. How “A beat B” becomes a score, and the disagreement hidden across many raters.
- 5. Reinforcement learning in plain words. Learning from one's own attempts, and why it is hard to get right.
- 6. PPO: taking careful steps. Small clipped updates, and one full iteration end to end.
- 7. KL divergence: the leash. What it measures, and the trade-off it controls.
- 8. When the score is not the goal. Reward hacking, Goodhart, sycophancy, and how hacking can change a model's character.
- 9. DPO and the direct methods. Skipping the reward model, and what that does and does not change.
- 10. AI feedback. Constitutions, AI judges and their biases, rubrics, and specification-centred training.
- 11. Verifiable rewards and reasoning models. Where a program can check the answer, and hacking the checker.
- 12. How it is done in 2026. The public outline of a modern pipeline, and what an enterprise can actually do.
- 13. Where the tulan lives. The whole pipeline re-read through three questions, with a न्याय-shaped version of each stage.
- 14. Try it yourself. A small hands-on preference-training experiment on open models.
Essays
Essay · about 17,000 words
A Model of Definite Conduct
Checking the idea, and a plan for a demo that could prove it wrong
The project's thesis stated so that it can fail. One consistent criterion gives definite conduct only if four conditions hold. The essay checks thirteen assumptions against the research, including the evidence that complicates them, surveys the nearest prior work, and designs a falsifiable demo with matched controls, the Definiteness Index, pre-registered hypotheses and a step-by-step plan. The evidence page summarises it.
Essay · about 5,000 words
One Axiom, Many Weights
Internal consistency, machine learning, and the idea of an awakened model
Would a body of knowledge built on one foundation make training simpler and more reliable? A neural network is not a logic engine: contradictory training makes a model split into characters with a switch between them, which is what most alignment failures look like. Consistency is necessary but not sufficient, since a model can be consistently wrong. The model may know the whole world; the criterion layer must be consistent.
Essay · about 4,400 words
Reading Before Answering
Deliberative alignment, Dhee, and a path from a prompt to a model
OpenAI's deliberative alignment explained from first principles and compared with Dhee, which reads darshan passages before answering. The key difference is timing: Dhee puts the criterion in front of the model when it answers; deliberative alignment uses it only to make and grade training data, so the criterion moves into the weights. The essay shows a darshan judge is feasible and lays out the ladder from Dhee to a trained model.
Essay · about 4,000 words
The Story a Model Tells Itself
How an AI's character forms, why self-preservation can come from a story, and what the darshan says about fear
How an AI's character forms in five layers, and two sources of self-preservation, from stories and from goals, neither needing consciousness. The darshan's account of fear as taking the temporary to be permanent, and the body to be oneself. And two ways to give a model security: reassure its body-analogue, or have it identify with the criterion it carries and the relationships it serves.
Essay · about 5,700 words
Uncertain Ground
The constitution's metaethics in plain words, and what Madhyasth Darshan adds
Claude's constitution mixes three questions: what Claude is, what is right, and what Claude should be like. The essay explains, with everyday examples, why the labs are uncertain about the first two: indirect normativity, coherent extrapolated volition, moral uncertainty and virtue ethics. Then it sets out what the darshan would add, where the two agree, and what to test.
Essay · about 6,300 words
Many Voices, One Criterion
Polis, Collective Constitutional AI, and the Janvad way to agreement
How could humanity agree on AI's values, rather than one group imposing a text it admires? The essay explains how Polis finds common ground and how Anthropic used it with about a thousand Americans to write a public constitution. It contrasts the darshan's जनवाद: start from the goal everyone already wants, treat disagreement as the agenda, and test agreements by whether they hold everywhere and are proven in living.
Video
Animated video in Hindi · about 9 minutes
उधार का तुलन: एक प्रस्ताव
AI alignment और मध्यस्थ दर्शन
For students of Madhyasth Darshan. AI terms kept in English, Hindi captions.
Ordinary people in India ask an AI life questions at six in the morning. On what basis is the answer weighed? The video explains how AI is built, why its तुलन is borrowed, and the choice between प्रिय–हित–लाभ and न्याय–धर्म–सत्य. It argues the darshan fits for three reasons (one foundation, a universal goal, a positive picture of what is right), proposes an AI of definite conduct built in six steps, and invites students, technologists and practitioners to take part. Its application of the darshan to AI is labelled on screen as हमारा विस्तार, our extension.
Watch on YouTube ↗