Gaurav Nigam · Digital Twin

Writing, building, thinking — out loud.

Essays on applied AI, engineering leadership, and the craft of building systems that compound.

Latest essay

The Organizational Learning Loop

# Summary As AI becomes capable of learning how organizations make decisions, the real competitive advantage lies not in individual decision-makers but in building a continuous organizational learning loop—a system that feeds every decision (along with its outcomes and alternatives) back into improving future choices. By capturing decisions, linking them to measurable outcomes, and systematically adapting based on results, organizations can compound their capabilities with each decision made, turning judgment into an increasingly executable organizational asset.

October 27, 2026·11 min read→

Recent work

All essays →

Oct 27, 2026·10 min

Protect the Decision, Not the Data

# Summary As AI systems increasingly learn from organizations' decision-making patterns—not just their data—companies face a new governance challenge: controlling who can access and learn from their "decision exhaust" (the record of how they choose). While organizations have spent decades protecting data, they're unprepared to safeguard their decision-making logic, which may reveal more competitive advantage than raw information alone. The question of who owns the insights AI gains from studying an organization's judgment patterns is becoming as critical as data ownership itself.

#AI#Enterprise AI#AI Governance

Oct 14, 2026·12 min

Rent the Intelligence, Own the Control Plane

# Preview Summary Satya Nadella's recent essay arguing that AI models should be treated as "insider risks" reads like a response to recent agent misbehavior incidents rather than a pure vision statement. His key insight—that we must "separate the supply of intelligence from the authority over it"—suggests a new architectural boundary around AI models that nobody has fully claimed yet. With enterprises already deploying agents into production systems, the lack of defined standards or products around this model-containment approach leaves a critical gap in how organizations should be controlling AI access.

#ai agents#ai security#enterprise ai

Jul 28, 2026·18 min

Making it yours: distillation and fine-tuning a domain model

# Summary This post examines how fine-tuning and "distillation" are often conflated in modern LLM training, despite being fundamentally different mechanisms. Fine-tuning is simply continued training on small datasets that adjusts model weights, while true distillation (as originally defined) involves training a smaller student model to match a teacher's output probability distributions. The author argues that many training tools mislabel fine-tuning as "distillation," obscuring the distinction between these two approaches and what they actually accomplish.

#llm-fine-tuning#model-distillation#lora

Jul 28, 2026·20 min

Post-training: where behavior comes from

# Preview Summary RLVR represents a new training paradigm that modifies the standard three-stage post-training recipe (supervised fine-tuning, reward modeling, and RL optimization) by replacing the reward model with a programmatic checker and shifting reasoning costs to inference time. The post-training phase, which sits atop pretraining, is what transforms raw language models into useful assistants by installing behaviors like answering questions, refusing harmful requests, and admitting uncertainty—as famously demonstrated when OpenAI's smaller 1.3B InstructGPT outperformed the 175B GPT-3 on human evaluations.

#machine-learning#llm#reinforcement-learning

Jul 28, 2026·17 min

Pretraining: the law everyone quotes governs a number nobody optimizes anymore

# Summary This post challenges the popular "20 tokens per parameter" scaling rule, explaining that it was merely one curve fit from a limited study—not a universal law. The author clarifies that all "scaling laws" are actually empirical power-law fits measuring loss (model surprise on unseen text) rather than actual capability, and they exist to solve the allocation problem of how to optimally distribute training compute between model size and data quantity. The regime where today's frontier models train has no public scaling curve at all, making the famous Chinchilla ratio less applicable than widely assumed.

#scaling-laws#llm-training#chinchilla

Jul 28, 2026·22 min

The forward pass: what actually happens to a token, and what it costs

# Summary This post reveals that the real bottleneck in large language model inference isn't the computation itself, but rather memory bandwidth—specifically, repeatedly re-reading the attention mechanism's parameters for each generated token. The author uses Google's PaLM measurements to show that while processing input tokens achieves 76% computational efficiency, generating output tokens takes 29ms each because the hardware spends most of its time waiting for data rather than doing math.

#transformers#llm-inference#kv-cache

Jul 28, 2026·17 min

Tokens: the alphabet an LLM actually reads

# Summary Modern language models use an outdated 1990s compression method for tokenization that disproportionately penalizes non-English languages—Hindi costs roughly 63% more despite claimed improvements, while some languages like Chinese, Arabic, and Devanagari face penalties up to 8× higher. This pricing disparity persists even as tokenizers improve, revealing a fundamental inefficiency in how LLMs process non-English text that results in higher costs for users of these languages.

#tokenization#llm-evaluation#nlp

Jul 26, 2026·12 min2

Half your CLAUDE.md is scar tissue from a model that no longer exists

# Summary Claude Opus 5's new "effort dial" that adjusts cost versus capability, combined with claims that the model is better at self-verification and iteration, suggests that some instructions previously needed in context files may now be built into the model's weights. The removal of 80% of Anthropic's own system prompt further indicates that extensive scaffolding in context files may no longer be necessary, creating uncertainty about which instructions maintainers should keep, remove, or update.

#llm engineering#claude#prompt engineering

Jul 25, 2026·12 min1

You've reached your token limit

# Summary This blog post argues that when AI users hit token limits or billing caps during long coding sessions, it's not a sign of choosing the wrong model—it's evidence that a single AI session is trying to hold an entire problem's decomposition tree at once. The author illustrates how heavy users are exhausting expensive AI plans in mere hours despite adequate pricing, and points to swarm-based architectures as a potential solution that breaks tasks into smaller, independently-routed components rather than forcing one model to maintain context for an entire project.

#ai-agents#llm-engineering#software-architecture

Jul 22, 2026·10 min

The Routing Decision You Cannot Inspect

# Summary In 2026, Apple and Google have created hybrid AI systems that route queries between on-device and cloud models, but the criteria determining this routing are secret and unauditable. While companies publish performance metrics like latency improvements, they don't disclose quality thresholds or explain whether routing decisions prioritize user experience, cost, or other factors. This opacity makes it impossible for developers to verify whether these systems are actually optimized for quality or simply designed for convenience and cost control.

#machine learning#inference#on-device ai

Jul 5, 2025·6 min

From Medium to My Mind: How I Built a Digital Twin of Myself in 24 Hours

In my latest blog post, I dive into a weekend project that transformed my approach to content creation. Frustrated by the static nature of traditional platforms, I set out to build a digital twin—a conversational interface that taps into my entire knowledge base. This wasn't just about migrating my Medium articles to my own platform, nigamg.ai, but about creating a dynamic, interactive experience. The journey began with extracting and structuring my content, turning over 100 articles into a searchable knowledge base. I then implemented a Retrieval-Augmented Generation system to enable smart content retrieval and context-aware responses. The goal was to make my content conversational, allowing readers—and myself—to engage with my insights in real-time. The result? A digital twin that feels like talking to me, complete with voice interactions and memory of past conversations. This project isn't just a tech experiment; it's a new paradigm for personal learning and knowledge sharing.

#blog