Writing
Notes from the work.
Research, implementation details, and what I learned when the first idea was wrong.
2026
19 essaysModel behavior
Jev Beat the 4B Model. Then the 4B Model Beat Jev.
Jev beat untouched small LLMs. After split-safe fine-tuning, a 4B model learned the task and transferred better on BeaverTails and C-SafeQA.
Sep 22AI safety
Your fraud model has never seen the attack that will work
A fraud model trains on fraud that already worked. When a new ring hit an unsecured lender, the behavioural anomaly model with no fraud labels caught 85% of it and the supervised model caught a third.
Sep 22Applied AI
Why construction estimates lose their margin after the bid is won
Construction estimates lose margin between bid and closeout because competitive bidding selects the low side of your own estimating error. Here is the mechanism, and the four questions that remeasure the rule.
Sep 22Model behavior
Why Couldn't an LLM Pick Its Best Expert?
A hindsight oracle could name the best expert in an LLM. A trained judge, a linear policy and a bounded nonlinear check could not find it live. Why a real opportunity never became a usable one.
Sep 22AI strategy
Why thin-file approvals look fine in backtest and fail in production
Thin-file approvals look fine in backtest because repayment labels only exist for applicants you already approved. Here is why the cutoff fails in production, and the four questions that remeasure it.
Sep 15AI strategy
What travels around the loop?
We swapped hidden states in Coconut and a small recurrent transformer to test what their loops carry and how they use it.
Sep 14Model behavior
Deleting a Model's Thoughts Changed Nothing. Then I Changed the Task.
Deleting a model's thoughts barely changed graph-search accuracy but cut arithmetic from 33.3% to 5.1% in this Coconut replication.
Sep 13Model behavior
I Sent a Small AI Model to School. Here’s Its Report Card.
A Small AI Model spent a school year with a larger teacher. It scored 25/78 on unseen puzzles, versus 8/78 for its unchanged copy. Here’s what stuck.
Jul 28Model behavior
I forced a Mixture of Experts to specialize by topic. It got worse.
Five Mixture of Experts models trained from scratch. Routing correlates more with a token's surface form than with the document's topic, and the pattern is fixed in the first 2% of training. Forcing topic specialization three ways made every model slightly worse.
Jul 22Model behavior
What does a Mixture-of-Experts router actually read?
A Mixture of Experts router picks its experts by reading the residual stream. A Jacobian lens on OLMoE and Qwen-MoE shows it sorts token form, brackets and verb tense, not meaning.
Jul 20Model behavior
Mixture of Experts: the sparse model that lights up 93% of itself
A Mixture of Experts should use a fraction of itself per token. Measured on a 4090: 628 of 1,024 expert slots fire every step, and 93% of them at batch 32.
Jul 15Model behavior
J-Scope: before a model answers, it thinks in concepts
The prompt never says Italy, but twelve layers into Qwen3.5-4B the model is already thinking it. J-Scope lets you watch that happen, inject France and get Paris, or drag one vector and watch the answer walk from euro to yen. A 72-second tour here; the live tool is one click away.
Jul 8Model behavior
I Fed a Model Facts Through Its J-Space Instead of Its Prompt
I wrote facts directly into a 4B model's J-space and KV cache on an RTX 4090. Three cache tokens matched full text-RAG, survived 48 distractors, and showed where retrieved facts actually live.
Jul 2AI agents
Can you tell a good AI agent plan from a bad one before you run it? I spent 6 million tokens finding out.
Can a cheap reward model tell a good AI agent plan from a bad one before you run it? I reproduced Orch-RM on a consumer RTX 4090 and RTX 3090: the verifier barely beats a coin flip on the signal that matters, yet still beats majority vote. Cheap to run, expensive to teach.
Jun 28AI agents
I Gave My Coding Agent Karpathy's Discipline Rules. It Got Too Careful to Fix the Bug.
I tested the viral Karpathy coding-discipline rules on 100 real SWE-bench bugs, three times. They didn't help: the agent fixed fewer of them, trading fixes for caution.
Jun 25Model behavior
I bolted MiniMax's MSA sparse attention onto a 3B model on a single 4090
MSA sparse attention, retrofitted onto a 3B model on a single RTX 4090 with no training: the 28.4x compute cut holds, quality survives, plus caveats the paper skips.
Jun 23AI strategy
I ran DFlash on a MacBook. The 4x is a data-center number.
I tested DFlash speculative decoding on an M5 Max and RTX 4090. The 4x headline is real, but workload and concurrency decide whether you get a speedup or a slowdown.
Jun 11Applied AI
Why your adaptive controller might be causing the defects it’s trying to prevent
A field note on over-cautious process control, and the margin hiding in your own sensor data.In one engagement: the force the controller watches stays low across speeds, while sidewall wear spikes at the crawl.“Slower is safer” is the most expensive assumption on your shop floor.Most adaptive controllers are built on it. When cutting force climbs, the controller slows the tool; when force eases, it speeds back up. That logic is sound for one failure mode: snapping a tool or gouging a part under
May 1AI strategy
The Geometry of Intelligence: What Lagrange Can Teach Modern AI
In 1772, Joseph-Louis Lagrange solved a piece of the three-body problem that had stumped everyone before him. He didn't solve it by tracking forces. He solved it by finding the five points in the gravitational field of two large bodies where a third, smaller body can sit stably - held in place not by any single pull, but by the balance of all of them.These are the Lagrange points. They are where constraints carve stable regions out of a space that would otherwise be chaotic.That is a useful imag
2025
24 essaysAI strategy
The Real State of Generative AI in the Enterprise in 2025
A practical analysis of generative AI in the enterprise, covering pilots, production failures, governance challenges, and what actually scales.
Nov 11AI in finance
AI in Finance: Transforming the Industry
Discover how AI transforms finance with automation, risk analysis, and customer insights. Stay ahead in the evolving financial landscape.
Nov 5AI strategy
When Computation Surprises Itself: How Emergent Uncomputability Might Explain AI, Consciousness, and the Unive
Emergent unpredictability may link physics, AI, and consciousness—suggesting the universe is a self-computing system discovering itself.
Nov 4AI strategy
Custom AI Solutions for Business Growth
Discover the benefits of custom AI solutions tailored to your business. Achieve efficiency and innovation with personalized AI strategies.
Oct 27AI safety
Fraud Detection Solutions to Protect Assets
Learn about cutting-edge fraud detection solutions to safeguard your business. Protect assets and build trust with innovative technology.
Oct 18AI strategy
When a Few Pixels Redrawn Change Everything: The Battle Over Human Authorship in the Age of AI
Understanding Intellectual Property in the Modern AgeAt its core, intellectual property (IP) is not merely about ownership of ideas. It revolves around incentive design. Societies discovered long ago that creativity and invention flourish when individuals can reap benefits from their mental labor. However, pure ideas are non-rivalrous: once shared, they can be copied infinitely. Therefore, IP law carves out temporary monopolies on specific expressions or applications of ideas. The logic is utili
Oct 7AI strategy
Animals vs Ghosts by Karpathy: A Realistic Path for Enterprise AI
Most AI systems today are ghosts: trained on human text, fine-tuned for alignment, and limited to observing and predicting.The future may lie in animals: AI systems that act, learn, and adapt from experience. But the path from ghost to animal is farmore complex than most realize. This article explains why enterprises should focus on perfecting "ghosts with tools" whilecarefully experimenting with "animal instincts" in controlled environments—and why this transition will take years, not months.1
Sep 30AI in finance
Finance x AI: Key Trends in AI in Financial Services 2025 - Phenx Tech
Discover the future of AI in financial services and its impact on finance workflows. Explore key trends and breakthroughs in AI in financial services today.
Sep 22AI strategy
Custom AI Solutions for Your Business
Discover the benefits of custom AI. Learn how custom AI solutions can address unique challenges in your business.
Aug 28AI strategy
Build vs Buy: Navigating Software Decisions in the Age of AI
Executive SummaryFor decades, executives have debated whether to build or buy software. AI has changed the dynamics, but the fundamental principle has become clearer than ever: • Buy more often than you build. Mature, complex, regulated systems are safer to source from vendors. • Build only when it increases your competitive gap or creates a unique value proposition. That is where differentiation, margins, and long-term survival live. • Augment with AI agents to bridge gaps, integrate platforms,
Aug 14Model behavior
I Read the Chat GPT-5 Prompting Playbook. Here’s My Hot Take.
CTO-friendly hot take on GPT-5’s Prompting Playbook: five levers that move metrics—eagerness, Responses API, verbosity, instruction hygiene, repo rules.
Jun 20AI strategy
AI Trends 2025: What Karpathy’s Talk Didn’t Tell You (But You Need to Know) - Software 3.0
Andrej Karpathy says prompts are the new code—but that’s just the beginning. This deep dive reveals what his Software 3.0 talk missed: real-world developer challenges, second-degree insights, and the shift from coding to product judgment.
Jun 12AI strategy
Apple's Illusion of Thinking paper
Apple's “Illusion of Thinking” paper reveals why reasoning models collapse—and why verification, not model size, is the next AI battleground
Jun 3AI strategy
AI Trends 2025: I Read Mary Meeker’s 300-Slide Deck So You Don’t Have To
A strategic breakdown of Mary Meeker’s 2025 AI Trends report for CTOs, CIOs, and CEOs—key signals, second-order insights, and what to act on now.
May 28AI safety
2025’s Ultimate Guide to AI-Driven Credit Risk Models for Climate-Impacted Regions
Explore how AI-driven credit risk models reshape lending in climate-impacted regions. Discover 2025 trends, expert insights and tools to future-proof your portfolio.
May 21AI safety
Beyond the Buzz: Deploying Generative AI in Credit Risk decision with Precision and Accountability
Discover how top lenders are using Generative AI to speed up credit risk, reduce bias, and scale underwriting with trust and transparency.
May 21Reasoning & reliability
Why AI Still Struggles with Negation and How We Can Fix It
AI still stumbles on negation—“not,” “no,” and absence. Discover why LLMs fail and how researchers are solving it with symbolic logic, counterfactuals & more.
May 8AI strategy
AI TCO Framework: Frequently Asked Questions About the True Cost of Enterprise AI
Discover the true cost of enterprise AI with our AI TCO framework. Learn what drives hidden costs, how to reduce spend, and download a free AI cost calculator. Ideal for CIOs, CTOs, and finance leaders planning sustainable AI adoption.
May 4AI agents
Will AI Agents Replace SaaS? Or Make It Invisible?
Discover how AI agents are transforming business software by automating logic, reducing backlogs, and making SaaS systems feel invisible. Learn what leading platforms like Salesforce and Microsoft are doing — and how to get started.
Apr 29Reasoning & reliability
secure AI development, enterprise AI solutions, explainable AI, AI uncertainty
AI-generated code introduces hidden risks for enterprises. Learn how Phenx transforms AI uncertainty into secure, explainable, enterprise-ready AI systems through expert engineering and client collaboration.
Apr 9AI agents
From Tool Soup to Protocol Power: Why MCP Server Might Just Be the API Revolution AI's Been Waiting For
Discover how the Model Context Protocol (MCP) server revolutionizes AI tool integration. Learn how businesses and developers can use MCP to build modular, scalable, and intelligent AI systems—without prompt chaos.
Mar 8Model behavior
The Pizza Paradigm: A Manifesto on the Art and Philosophy of AI Prompting
Four ways to structure AI prompts, explained through pizza, music, visual art, and the creative process.
Feb 22AI agents
Why LLMs Are the Future of IT Automation: Beyond Simple Rule-Based Systems
Discover why LLMs and Future of IT Automation are game-changers for IT systems. Learn how LLMs enhance operations beyond rule-based limits.
Jan 14AI agents
The Future of AI Agent Frameworks: Trends, Predictions, and Business Opportunities
Discover the future of AI agent frameworks with insights into hybrid systems, low-code platforms, explainable AI, and edge computing. Learn how these trends are transforming industries and creating smarter, scalable, and ethical AI solutions.
2024
21 essaysAI agents
How AI agents will Automate Enterprise workflow?
Discover how AI agents are revolutionizing enterprise workflows by automating repetitive tasks, enhancing decision-making, and orchestrating complex operations. Learn about their impact, real-world success stories, and how they’re shaping the future of work.
Dec 31AI agents
AI vs. Agents: Understanding Their Differences and Collaborative Roles
Explore the distinctions and collaborative dynamics between Artificial Intelligence (AI) vs agents. Learn how thinkers and doers in technology shape modern innovations.
Dec 26AI agents
Agents Unleashed: The Top 5 AI Agent Types Powering the Future
Discover the top 5 AI agents revolutionizing technology and business. Learn about their types, real-world applications, and why executives should care. Explore how AI agents like learning and utility-based systems drive innovation in industries from robotics to finance.
Dec 26Reasoning & reliability
Harnessing the Creative Power of AI Hallucinations: A New Frontier for Innovation
Discover how to turn AI hallucinations into a powerful tool for innovation. Learn strategies to control and harness AI's creative potential for breakthroughs in science, technology, and business.
Nov 25Model behavior
The Art of AI Traffic Control: Mastering LLM Routing
Discover how LLM routing is transforming industries by optimizing AI-powered workflows. Learn real-world use cases, from customer support to healthcare, where task-specific models improve efficiency, reduce costs, and enhance performance.
Nov 8Reasoning & reliability
Can AIs Debate Their Way to the Truth? Exploring Experimental AI Debates with a Guidance-Focused Judge
This blog post delves into a fascinating thought experiment where two AI models engage in a structured debate over the nature of a simulated surface—flat or round. Using a debate framework guided by a sophisticated, evidence-focused judge, we explore how AIs can be trained to prioritize truth over rhetorical skill. The post examines the importance of rigorous experimentation, the role of guidance-based judgment, and the broader implications of debate-driven truth-seeking for AI alignment.
Oct 30AI strategy
How Can Fuzzy Matching and NER Solve Complex Business Data Challenges?
ransform unstructured business data with fuzzy matching and NER. This guide reveals how Python can help solve complex data inconsistencies for improved accuracy in machine learning projects.
Oct 29AI strategy
Behind the Curtain: Why AI Explainability Matters Now More Than Ever
Explore the importance of explainable AI in public sector applications, with insights on balancing complexity and transparency. Discover practical templates, stakeholder-focused explanations, and ethical considerations, guiding responsible AI development for trust and accountability.
Oct 23Model behavior
What Happens When AI Faces Uncertainty? Unlocking the Power of Probabilistic Reasoning in Language Models
Discover the power of probabilistic reasoning in AI. Learn how Language Models are evolving to make flexible, nuanced decisions in uncertain environments.
Oct 22AI safety
The Power of RPA + AI for Heavy Industries: Automating for Efficiency, Safety, and Predictive Insights
Discover how RPA and AI are transforming heavy industries like Oil & Gas, Manufacturing, and Construction. Learn how automation boosts efficiency, enhances safety, and delivers predictive insights for smarter decision-making and optimized operations.
Oct 16Model behavior
LoLCATs: Demystifying Linearized Attention in Large Language Models
Discover how LoLCATs (Learnable Linearized Attention Transformers) revolutionize large language models by simplifying the attention mechanism from quadratic to linear complexity. This blog explores the use of learnable linear attention, low-rank adaptation (LoRA), and layer-wise optimization to make LLMs more efficient, scalable, and accessible. Learn how LoLCATs enable models to handle larger sequences with reduced computational costs, while maintaining performance.
Sep 27AI agents
Agents on Rails: The Smartest Way to Keep AI Under Control and On Task
"Agents on rails" refers to a structured framework for autonomous AI agents, where their actions, decisions, and processes are guided by predefined pathways or "rails." This approach limits the unpredictability of agents by constraining them to follow specific rules or workflows, ensuring that they remain aligned with business objectives while carrying out complex tasks. This architecture balances flexibility with control, helping businesses automate processes efficiently.
Aug 26AI safety
Navigating the AI Seas: A CIO’s Guide to AGI Safety and Alignment
Explore how CIOs can navigate the challenges of Artificial General Intelligence (AGI) with a focus on safety and alignment. Discover actionable strategies to harness AGI's potential while mitigating risks, set in a near-future scenario where AGI is on the horizon. AI Safety is key for enterprise AI journey.
Jun 27Model behavior
Streamlining AI: The Essentials of Quantization and Model Distillation for Large Language Models
With Large Language Models (LLMs) like OpenAI's GPT becoming increasingly prevalent, optimizing them for better performance and lower resource consumption has become a top priority. Two techniques stand out in this optimization landscape: quantization and model distillation.
Jun 13AI strategy
Deciphering the Titans of Thought: General AI (AGI) vs. Strong AI (ASI)
Unravel the enigmatic world of AI with our guide to Artificial General AI vs. Artificial Strong AI. Explore the cosmos of AGI and ASI with us!
Jun 7AI agents
AI for Business Automation: Essential Strategies for C-Suite Executives
By harnessing the capabilities of AI, Business can automate mundane tasks, freeing up valuable resources to focus on strategic initiatives.
Jun 6Model behavior
Behind the Curtain: Quirks and Perks of Enterprise AI in Finance - Prompt Engineering, RAG, and More
Unleash the full potential of Enterprise AI in finance through the art of prompt engineering and the transformative power of Retrieval-Augmented Generation (RAG).
Jun 6AI strategy
Generative AI vs. Applied AI in Business
Both Generative AI and Applied AI are integral parts of the broader Business AI landscape but serve different purposes. Generative AI focuses on creating new content and ideas, while Applied AI is about applying AI technologies to improve processes, enhance decision-making, or automate tasks across various industries.
Mar 2Model behavior
Adventures in AI: Duel of the Digital Minds - Exploring the RPG World with Two Unique LLM Architectures, Act 2
Building on the foundation in "Navigating New Realms: Open-Source LLMs in RPG Environments," our follow-up exploration in LLM innovative application series, pushes the envelope further by comparing the outcomes of two different models in controlling a player within a virtual RPG world. The first model, a 7B Dolphin2.2-Mistral, demonstrated impressive capabilities in navigating and strategizing within a complex, dynamically rendered RPG landscape. This model, built on interpreting textual descrip
Feb 21Model behavior
Navigating New Realms: Open-Source LLMs in RPG Environments
The crossroads of AI and gaming are more than just a breeding ground for innovation; they are testing grounds for the future. Our latest venture pushes this boundary by integrating Large Language Models (LLMs) into the heart of role-playing games (RPGs). This project, rooted in open-source and community collaboration, puts LLMs to the test in a dynamic RPG universe. Here, LLMs face a world that changes with every decision, challenging them to devise strategies in real-time. But this is about mor
Feb 5AI safety
Security of AI Models: Navigating Emerging Threats and Solutions
Executive SummaryThis blog post provides a comprehensive exploration of the current security challenges faced in the field of Artificial Intelligence (AI) . This document serves as a critical guide for understanding and addressing the various types of attacks that AI models are susceptible to, including data poisoning, membership inference attacks, model extraction attacks, and the practice of fairwashing. The white paper aims to educate and inform a wide range of audiences, from AI professional