A few years ago, the biggest leap in AI didn't come from a bigger model. It came from a simple instruction: think step by step. Asking a model to show its work before answering, what researchers call chain of thought, turned large language models from fluent autocomplete into something that can actually reason through a problem. Today it sits at the core of every serious frontier model. I want to use my first post here to lay out where I think it goes next, and why it matters for anyone building with AI or hiring the people who do.

From prompt trick to core architecture

Chain of thought started as a prompting technique. It is now a training target. The leading reasoning models are trained to think before they speak, and they get better the longer they are allowed to think. That has created a new scaling axis: instead of only spending compute to train a bigger model, we now spend compute at inference time to let the model reason longer on harder problems. Math, code, planning and multi-step analysis have improved dramatically because of it.

The practical consequence is that "how smart is the model?" is no longer a fixed number. It is a dial. You can buy more reasoning for a hard task and less for an easy one, and the economics of AI products increasingly come down to how well you set that dial.

Five shifts I expect over the next few years

1. Reasoning budgets become a product decision. Teams will route simple requests to fast, cheap responses and reserve deep thinking for the problems that justify it. Knowing when a model should think for two seconds versus two minutes will be a core skill, much like knowing when to cache or when to call a database.

2. Thinking moves partly out of language. Today most chain of thought is written in human-readable text. There is active research into reasoning in compressed or latent form, which can be faster and cheaper. The tension is that readable reasoning is one of our best windows into what a model is doing. I expect the industry to fight hard to keep that window open, and I think it should.

3. Monitorability becomes a safety requirement. When a model writes out its reasoning, we can check it for mistakes, shortcuts and unwanted behavior. That is enormously valuable, but only if the written reasoning faithfully reflects what the model is actually doing. Research has already shown that stated reasoning does not always match the real drivers of an answer. Measuring and preserving faithful chain of thought will become a standard part of evaluating and deploying models, not an academic side topic.

4. Agents turn reasoning into action. The most important use of chain of thought is no longer answering questions. It is planning and executing multi-step work: calling tools, reading results, correcting course and finishing a task. Long-horizon agents live or die on the quality of their reasoning between steps. This is where I spend most of my own time, and the gap between an agent that reasons well and one that doesn't is the gap between a useful colleague and an expensive liability.

5. Reasoning becomes verifiable. In high-stakes settings such as finance, law, medicine and onchain systems, "trust me" is not good enough. I expect reasoning traces to be checked by other models, by formal tools and, in some cases, anchored to tamper-evident records so decisions made by autonomous systems can be audited after the fact. Verifiable reasoning is where AI and blockchain infrastructure meet naturally.

What this means for builders and teams

• Design for inspectable reasoning. Log it, evaluate it and use it to debug. The reasoning trace is the best telemetry you have.
• Treat thinking time as a cost line. Measure accuracy against reasoning spend and tune per task.
• Don't confuse a confident explanation with a correct one. Build checks that test outcomes, not just how convincing the reasoning sounds.
• Hire for judgment, not just prompting. The valuable skill is knowing how to structure problems, evaluate reasoning and decide where a human needs to stay in the loop.

The bottom line

Chain of thought was the moment AI started to show its work. The next phase is about making that work cheaper, more reliable and more honest, and about letting it drive real action through agents. The teams that win will be the ones that treat reasoning as something to measure, budget and verify, not just something to admire.

I'd love to hear how others are using reasoning models in production. What's working, and where is the reasoning breaking down?