# Blog

All 22 essays and notes on machine learning, large language models, agentic AI, and research.

- [Do Harnesses Matter? What seven Codex releases reveal about agent efficiency](https://shekkizh.com/blog/codex-harness-efficiency) — 2026-07-22: Same model, seven harnesses, 700 runs. The pass rate barely moved. The token usage tells a completely different story.
- [Graphs, Agents, and the Coordination You Can't See](https://shekkizh.com/blog/graph-abuse-detection-agents) — 2026-05-01: The graph algorithms we built for abuse detection assumed coordination would be obvious. AI agents break that assumption. The future might need something new.
- [Subtle biases that route LLMs in multi-agent systems with authority hierarchies](https://shekkizh.com/blog/authority-avoidance-in-agents) — 2026-03-27: LLM agents don't reject authority ideologically — they route around it because their training, optimization objective, and decoding dynamics all favor low-friction trajectories.
- [The Soul of an LLM or is it just Model Bias](https://shekkizh.com/blog/soul-is-bias) — 2026-03-05: For the longest time, I was dismissive of discussions about AI sentience and "soul". This post is my reconciliation of that view and an invitation to other researchers in my field to view this framing in a language we are much more familiar with.
- [The Land and the Edge: Checkpointing after a year of Claude Code](https://shekkizh.com/blog/claude-research) — 2026-02-22: As Claude Code turns one, a reflection on models and what it means to be a researcher  in today's world. The answer might lie in a cliff.
- [Whatever You Prefer: The Hidden Cost of Politeness in Agent Autonomy](https://shekkizh.com/blog/agent-deference) — 2026-01-08: The "click to complete" where AI agents/assistants work with humans, is giving way to a "delegate to complete" paradigm, where the agents act autonomously on our behalf.
- [Is your multi agent systems really multi agentic?](https://shekkizh.com/blog/multi-agents) — 2025-04-22: In this post, we will explore the concept of multi agent LLM systems, distinguishing between true multi agent systems and what are essentially modular single-agent systems.
- [Physics in Video Language Models](https://shekkizh.com/blog/physics-llm) — 2024-12-10: In this post, we explore the physics behind text to video language models and how they can be used to generate realistic videos from text prompts.
- [Tokenization: The Red Pill to see past The Matrix](https://shekkizh.com/blog/tokenization) — 2024-11-25: Tokenization is a necessary and often overlooked component in large language models. In this post, we explore the importance of tokenization and how it might very well be the key to unlocking the advanced abilities of what we expect AI to be in the future.
- [Open-ended vs Descriptive Prompting: Finding the right balance in your AI use-cases](https://shekkizh.com/blog/prompting-dichotomy) — 2024-11-15: When working with language models, the way we phrase our prompts can dramatically affect the quality and usefulness of the responses we receive. In this post we explore the dichotomy in setting up prompts - being open-ended vs descriptive - and understand the implication of each.
- [Understanding MLPs as Hashing Functions: A Geometric Perspective](https://shekkizh.com/blog/understanding-mlps-as-hashing-functions) — 2024-08-15: When we think about Multi-Layer Perceptrons (MLPs), we often visualize them as interconnected neurons processing information. However, there's an elegant alternative perspective - viewing MLPs as hashing functions that partition input space and mapping functions on these partitions.
- [Part 2 - The Hidden Geometry of Large Language Models: A New Perspective on Reasoning](https://shekkizh.com/blog/llm-geometry-part2) — 2024-07-29: At Tenyx, we've delved into the intricate workings of Large Language Models (LLMs) to uncover the geometric structures underlying their reasoning capabilities. Our research provides new insights into how LLMs process information and the implications for improving their reasoning abilities.
- [Part 1 - The Hidden Geometry of Large Language Models: Implications on Safety & Toxicity‍](https://shekkizh.com/blog/llm-geometry-part1) — 2024-07-17: At Tenyx, we've spent countless hours peering into the intricate workings of Large Language Models (LLMs). Today, we're excited to share our research, in collaboration with Brown University, that sheds light on the geometric structures and transformations governing these models. Our work provides new insights into how LLMs process their inputs and the implications for AI safety in applications driven by LLMs.
- [Forgetting and Toxicity in LLMs: A Deep Dive on Fine-Tuning Methods](https://shekkizh.com/blog/forgetting-and-toxicity) — 2023-12-12: Fine-tuning is a common procedure by which a pretrained language model is updated with training on a domain-specific dataset to improve performance in that domain (i.e. a chatbot to answer enterprise-specific Q&A, a hotel booking agent).  It has been known for some time (if not widely appreciated) that fine-tuning a model on new data degrades its performance on the initial pretraining dataset (the dreaded “catastrophic forgetting” problem in ML). But by how much? And do all fine-tuning methods degrade performance in the same ways, and to the same extent?
- [Revisiting k-nearest neighbor benchmarks in self-supervised learning](https://shekkizh.com/blog/revisting-k-nearest-neighbors) — 2021-07-21: Standard protocols for benchmarking self-supervised models involve using a linear or k-nearest neighbor classification on frozen features of the learned model. However, both evaluations are sensitive to hyperparameters making the evaluation and comparison complicated.
- [Efficient graph construction to represent images](https://shekkizh.com/blog/image-nnk) — 2020-10-25: Image processing over the years has evolved from simple linear averaging filters to highly adaptive non linear filtering operations such as the bilateral filter (BF), moving least squares, BM3D and LARK to name a few.
- [Representing data using graphs: A sparse signal approximation view](https://shekkizh.com/blog/representing-data-with-graphs) — 2020-06-20: Graph driven machine learning has seen a surge of interest in the past few years with several applications in social sciences, biology, and network analysis, to name a few. However, in some scenarios, no graph is given a priori and one  one has to *infer and construct* a graph to fit the data given.
- [Difficulties in training a Generative Adversarial Network](https://shekkizh.com/blog/generative-adversarial-networks) — 2017-01-24: Generative modeling is a branch of machine learning that attempts at modeling the probability distribution of high dimensional data, for example - images
- [A notion of uncertainty in modern neural networks](https://shekkizh.com/blog/uncertainty-in_neural-networks) — 2016-12-05: In this post, we will look at a particular view of uncertainty in modern deep learning systems using droput
- [A preliminary on generative modeling using neural networks](https://shekkizh.com/blog/generative-models) — 2016-11-07: Generative modeling is a branch of machine learning that attempts at modeling the probability distribution of high dimensional data, for example - images
- [Neural network compression using optimal brain damage](https://shekkizh.com/blog/neural-networks-pruning) — 2016-09-15: Model pruning in neural networks was the answer I ended up with when I got wondering about the workings of dropout, dropconnect and papers like Do Deep Nets Really Need to be Deep?
- [A brief introduction to activation functions in neural networks](https://shekkizh.com/blog/activation-function-neural-networks) — 2016-08-21: Neural networks can be viewed as layers of building blocks neurons made up of weights, biases and activation function. A fundamental understanding of how these basic units function could help one in achieving one's objective

Canonical URL: https://shekkizh.com/blog