Articles

In September, I wrote about testing and fine-tuning Laya. I liked the idea of giving a model information and a decision, then receiving a structured result. But the output format alone does not tell you whether the decision is useful. I have now built BranchOne, my own System One decision model for choosing among candidate descriptions supplied with a request. Give it a situation, a question, and choices, optionally with one image. It returns a probability per choice and the selected candidate’s ID. ...

I recently shared a short post on LinkedIn: “If an agent can do the job with an existing CLI, I do not think we need to build another MCP integration.” Let me explain the thinking behind it. Start with the Tools You Already Have Before building an integration, I would ask what the agent actually needs to do. Maybe it needs to list pull requests, inspect a deployment, or read a file from cloud storage. We already have command-line tools for these operations. An agent becoming the caller does not automatically create a need for another interface. ...

“Necati, how can I burn more tokens?” Someone asked me this recently. I had to pause. I understand wanting engineers to use AI. Give them access. Give them time to learn. Let them try approaches that might fail. But why would consuming more tokens become the objective? ...

I do not think AI cost is just a finance problem. It is mostly an architecture problem. If you are building an AI product, the cost starts with decisions your engineering team makes every day. Which model handles a request? How much information does it receive? How many times does it run? What happens when the first answer is wrong? By the time the monthly bill arrives, those decisions are already in production. You can negotiate a better price, and that may help. But you still need to understand why your system is doing that much work. ...

There are still lots and lots of organizations treating LLM usage during coding interviews as cheating. I think that needs to change. If you expect your engineers to work with AI, your hiring process should evaluate how effectively candidates work with AI. That is becoming part of the job. This is still far from standard practice. In CoderPad’s 2026 survey, 34% of surveyed hiring leaders said AI was not allowed in interviews and assessments. That is a vendor survey, so I would not treat it as a precise measurement of the entire industry. But it gives us an idea of the disagreement. ...

I recently read Hamel Husain and Shreya Shankar’s article about similarity metrics. Their argument is that metrics like ROUGE and BERTScore often miss the failures that matter in an LLM application. I agree. But I think we need to separate two things: evaluating what a model learns during training and evaluating whether a RAG application works. Both involve an LLM, but the questions are different. If you do not understand the question, choosing a metric will not help you much. ...

1. From Jev to Laya I recently read TypeSafe’s announcement of Jev. The idea is interesting: give a model some information and predefined questions, then get structured decisions with probabilities. A support workflow, for example, could use it to choose which team should handle a ticket. But a structured answer can still be wrong. I wanted to see how this kind of model performed on a concrete task. Then I found Laya, an open source decision model. Its author, Nandakishor M, claims he worked on this direction “a year ago.” His March 2025 SalesRLAgent paper documents earlier sales-prediction research. His write-up describes developing the general-purpose model afterward, so I would not say today’s Laya checkpoint was released a year ago. ...

I lost my father in June. I feel lucky that I had my father until I was 43. But at the same time, he was 71. You can also consider that young. He spent the last three months of his life in the hospital. I flew to Turkey in April. I knew this was probably the last time I would see him. That is a strange feeling. You are going to see your father, but at the same time, you know you may also be going there to say goodbye. ...

Many founders think a CTO is just a CTO. I do not think that is true. I have seen many startups look for the wrong kind of technical leader. Usually this happens because the founder knows they need technical leadership, but they do not clearly understand what kind of technical leadership their company actually needs right now. That difference matters. If you are a non-technical founder, maybe you raised some money, maybe you have an idea, and now you are looking for someone to lead the technical side of the company. It is natural to say, “We need a CTO.” ...

I was recently reading a Wall Street Journal article, and one sentence in it reminded me of Jack Welch, the CEO of GE from the 1980s to 2001, roughly 20 years at the top. I want to talk about that, because I think there is a real connection between what happened at GE and what is happening right now with AI coding agents and vibe coding. The article was titled The AI Superstars Who Say a ‘Vibe Slop’ Crisis Is Coming, and it argues that vibe coding is generating code we cannot really trust. As of now, I agree. ...