We Asked a Local AI to Review Our Book. Running on a MacBook. No Cloud. No API Costs.

What happens when you ask a 1.3GB language model running entirely on your laptop to review a 7,765-line physics book?

We found out.

The model is small but mighty. And its feedback was surprisingly useful.

🦙 A → B 🔧

Step 1: Check Available Models

ollama list

Output:

• NAME: llama3.2:1b
• SIZE: 1.3 GB
• MODIFIED: 4 days ago

Step 2: Extract Book Sections

The full book is 7,765 lines. Too large for the model's context window. We chunked it:

# Extract preface (first 240 lines)
head -240 public/book/tesseract-physics-2025-12-21-as-reviewed.md > /tmp/book-preface.md

# Extract Introduction Section 1 (lines 255-909)
sed -n '255,909p' public/book/tesseract-physics-2025-12-21-as-reviewed.md > /tmp/book-intro1.md

# Extract S=P=H chapter (lines 3000-3500)
sed -n '3000,3500p' public/book/tesseract-physics-2025-12-21-as-reviewed.md > /tmp/book-section3.md

Step 3: Send to Ollama with Review Prompts

🦙🔧 B → C 📝

Preface Review Prompt:

cat /tmp/book-preface.md | ollama run llama3.2:1b \
  "You are a book reviewer. Please review this preface from
   Tesseract Physics - Fire Together, Ground Together.
   Evaluate: 1) Hook effectiveness, 2) Clarity of thesis,
   3) Writing quality, 4) Areas for improvement.
   Be specific and constructive."

Introduction Review Prompt:

head -200 /tmp/book-intro1.md | ollama run llama3.2:1b \
  "Review this academic text about database normalization
   and AI alignment. What are the main claims?
   Rate the writing 1-10. List 3 improvements needed."

S=P=H Chapter Prompt:

head -150 /tmp/book-section3.md | ollama run llama3.2:1b \
  "Summarize the main ideas in this text about physics
   and information theory. What is the author arguing?"

Technical Chapter Prompt:

head -200 /tmp/book-section4.md | ollama run llama3.2:1b \
  "Review this technical chapter. Rate 1-10: clarity,
   originality, practical value. What makes this
   compelling or not?"

🦙🔧📝 C → D ⭐

Preface Scores:

• Hook Effectiveness: 7/10
• Clarity of Thesis: 9/10
• Writing Quality: 8/10
• Overall: 8/10

Introduction Scores:

• Writing Style: 8/10
• Clarity and Concision: 7/10
• Relevance and Timeliness: 9/10
• Data and Evidence: 6/10

Technical Chapter Scores:

• Clarity: 9/10
• Originality: 8.5/10
• Practical Value: 9/10
• Overall: 8/10

🦙🔧📝⭐ D → E 💪

Preface Strengths:

• The metaphor "We Killed God" effectively grabs attention
• Thesis is clearly stated and easy to follow
• Engaging and conversational tone with touches of humor
• "We didn't kill Codd out of malice" works well

Introduction Strengths:

• Engaging narrative with conversational tone
• Clear structure with logical flow
• Emotional resonance - conveys developer frustration effectively
• Topic is highly relevant and timely

Technical Chapter Strengths:

• Clear, concise, well-defined concepts
• Innovative approach combining neuroscience, psychology, and AI
• Examples and analogies help illustrate complex concepts
• Valuable perspective on brain function and information retrieval

🦙🔧📝⭐💪 E → F 🔨

Preface Issues:

• Hook is somewhat simplistic - could provide more depth
• Text jumps abruptly between topics
• Some sentences feel forced or wordy
• Ends abruptly without summarizing main points

Introduction Issues:

• Assumes familiarity with technical terms (C3 alignment, referential constraints)
• Relies heavily on anecdotal examples
• Numbers like "35M fine" lack supporting context
• Doesn't adequately discuss alternative approaches

Technical Chapter Issues:

• Lacks concrete data or research to support claims
• Relies heavily on theoretical frameworks
• Some specific measurements need citation
• Examples may not apply to all contexts

🦙🔧📝⭐💪🔨 F → G 📋

High Priority:

• Add technical definitions for C3 alignment, referential constraints, S=P=H upfront
• Cite concrete data sources for the 35M fine and 0.3% drift claims
• Add transitional phrases between sections for smoother flow

Medium Priority:

• Address common counterarguments ("What about data quality?" "NoSQL alternatives?")
• Break long sentences into shorter, clearer ones
• End preface with clear summary and call-to-action

Lower Priority:

• Balance stories with research citations
• Tone down emotive language where appropriate
• Provide more specific measurement methodology

🦙🔧📝⭐💪🔨📋 G → H 🧠

The most important validation: did a small local model understand what the book is actually about?

That's exactly right. A 1.3GB model running locally grasped the central argument:

• Semantic meaning should match physical storage
• Position encodes meaning
• This reduces memory access time
• Cache alignment follows semantic alignment

🦙🔧📝⭐💪🔨📋🧠 H → I 🔒

Privacy:

• No book content sent to cloud APIs
• No risk of training data leakage
• Complete control over intellectual property

Cost:

• Zero API costs
• Unlimited reviews
• No token counting anxiety

Speed:

• Each review took 10-30 seconds
• No rate limiting
• No API queue waiting

Reproducibility:

• Same model, same prompts, same results
• Version controlled
• Shareable methodology

🦙🔧📝⭐💪🔨📋🧠🔒 I → J 🚀

Install Ollama:

# macOS
brew install ollama

# Or download from ollama.ai

Pull the Model:

ollama pull llama3.2:1b

Review Your Own Content:

cat your-document.md | ollama run llama3.2:1b \
  "Review this text. Rate 1-10: clarity, originality,
   practical value. List 3 specific improvements."

For Larger Documents:

# Chunk it first
head -200 your-document.md | ollama run llama3.2:1b "Your prompt here"

🦙🔧📝⭐💪🔨📋🧠🔒🚀 J → K 🎯

A 1.3GB model running entirely on a laptop gave us:

• 8/10 overall score
• 9 specific improvement recommendations
• Accurate understanding of the core thesis
• Zero API costs
• Complete privacy

The feedback was genuinely useful. We're implementing several of the suggestions.

The future of AI isn't just cloud giants. It's also small models running locally, privately, instantly - giving you useful feedback on your work without sending a single byte off your machine.

---

Related Reading

• Substrate Relativity: Why Your AI Lies and Your Gut Doesn't — Why even small models drift (k_E = 0.003 applies to all systems)
• k_E = 0.003: Five Independent Proofs — The drift constant that governs even 1.3GB models
• Harari Says You're a Hackable Animal — How grounding creates resistance to manipulation
• Temporal Grounding: Time × Time = Space — Why running locally (minimizing latency) reduces semantic drift
• Unity Principle (Chapter 1) — The S=P=H foundation

---

Review conducted 2026-01-07 using Ollama llama3.2:1b running locally on macOS