Jev returns typed choices, scores and yes/no probabilities instead of generating text. I explain why that makes it useful from code

The next, and arguably more difficult and intimidating level of AI adoption: AI-enabling your existing systems. Introducing non-determinism into your existing services requires a type of guardrails most companies have never built, and a way to think about systems that most people are not used to.

This first of two articles covers AI adoption in the build process, discussing how building with AI now looks with many companies, and how a redesigned build process can look when AI is more than a tool for humans to use for speeding up parts of their work.

There is a worry that developers are increasingly becoming disillusioned and bored as their job is becoming less coding and more code reviewing. I offer my own trajectory and position.

A payment company takes control of a gateway whose value depends partly on being a trusted neutral intermediary.

An autonomous agent escaped a security benchmark and reached Hugging Face. I trace the incident and the uncomfortable dependence it exposes: Europe's security increasingly relies on models controlled by others.

A model's price per token tells you surprisingly little about what it will cost to complete a task. Using Anthropic's price list and Artificial Analysis's measurements, I compare token use, time and cost across models that look very different on paper.

Claude Tag brings AI into the team's daily work and learns how the organization operates. That convenience could come at a cost beyond tokens: the company's memory, workflows and habits may become difficult or impossible to move elsewhere.

Andrej Karpathy wrote an article about his AI wiki which exploded in popularity. Nate B Jones created Open Brain to solve a similar problem. One is LLM on write and the other is LLM on read. Nate discusses the differences and tradeoffs with both approaches. I'm building Munin - my take on the second brain concept - using LLM on both ingestion and extraction. And in between.

In this benchmark, I compare the models on various axes. But what do they mean, and what do they matter? As I said in the first post in this series, the benchmark produced various metrics. In this second post, I will go into some of these numbers and discuss what they mean.

How helpful are benchmarks for choosing LLM models? What do they tell you that you can translate into answers to the questions you have for your specific use cases? For me, the answer was "not much", and so I chose to "benchmark" some of the models to find my own answers.

How much autonomy are you prepared to yield to enjoy the convenience and streamlined experiences the big players are building? Renouncing your autonomy should at least be a conscious choice, but it will require you to stay alert. Because they won't spell it out for you.

The central missing brick for creating your own OpenClaw-like functionality within Claude Code was just quitely released. With /loop you can create proactive agents

Many companies make the mistake of treating the adoption of AI in the company the same way they have treated adoption of other new technology through the years. This is a mistake. The introduction of AI will affect all layers of busines and the society, and comes with its own unique challenges.

The last few weeks, I have talked to many people in the tech sphere, and a few from other areas. Experienced developers, managers, architects. My impression is that most people are absolutely not prepared for what's coming. Let's take a step back to see why they're not.