Posts

I was perusing the Deloitte Tech Trends of 2026. Some notes of things I found interesting: Conversations with leaders are moving from “what can we do with AI?” to “how do we move from experimentation to impact”. Things with AI feel urgent, but this is driven by the speed of the change more than anything - the broad use of AI by over 50 million people happened in just 2 months, 300x times faster than the adoption of the telephone and 84x faster than adoption of the internet. This is driving faster time to value, faster scaling and faster technology adoption. ...

Was looking at some of the new tooling and frameworks around enabling Agents to generate UI. Just text-based becomes limiting. With agents being able to render more complex UI, across multiple platforms you could even end up running Doom right in your chat experience: https://x.com/rauchg/status/1978235161398673553?s=20 Google just released an open-source A2UI sdk to enable this. But there are others too. I was taking a look at MCP Apps (formerly MCP-UI). Ostensibly, these frameworks overarching goal is to enable a richer experience via chat interfaces. A lot of today’s interactions have been limited to primarily text only, but with MCP Apps and others this is changing. It allows the MCP Tool to now define UI components to render back to the calling client, in a standard way that the client can then render. This could start to empower AI agents to be able to generate their own UI, on the fly. Seems like a cool concept and it opens up several more rich experiences. ...

Imagine if you could run a powerful LLM model locally. This would enable you not need to be online all the time to get usefulness. Even more interesting is the idea that it could be more cost-effective - if you have the right hardware already you don’t have to pay for a host with premiums. In addition, you will own the data you send and receive from the model. It won’t be storing your questions somewhere you don’t allow it to, it won’t use your data to go train the next model. ...

Every 6 months or so, Ben Evans creates a big presentation around tech industry trends. His most recent one (updated for Nov. 2025) clocks in at a whopping 90 slides. You can see him talk about it at the recent Slush conference. Titled: “AI Eats the World” its a play on the older adage of “Software eating the world” first coined by Marc Andresson in his op-ed in the WSJ back in 2011 and explores AI trends. I think its got some really great content and nuggets, capturing the state of AI at this point in time. ...

I came across this article the other day, Chatbots Can Go Into a Delusional Spiral. Here’s How It Happens. which details how a 47 year old, corporate recruiter, was convinced by ChatGPT that he had discovered a novel mathematical formula. Except he hadn’t. Over the course of 3 weeks, his interactions with ChatGPT led him down this spiral and he only got out of it when he used a different AI system to see if it made sense. It reminded me of one of the first chat bots, Eliza, from back in the 1960s. This bot was pretty effective, and all it did was reflect back to the user their own words. Some users began forming emotional attachments and shared personal and private data with it during their interactions. ...

Everyone’s talking about AI transforming how we work but how often do we actually use it to think better? America’s AI Action Plan was unveiled earlier this month. It focuses on 3 core pillars: Accelerate AI Innovation Build American AI Infrastructure Lead in International AI Diplomacy and Security. The full pdf is here: https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf. Rather than write a post about the interesting things in this plan, I wanted to pivot this a bit and talk about how I use AI to interpret this plan to get a better understanding of it. We hear a lot about AI and how it can supercharge learning and understanding, but we see less examples of this in action. My approach works for me, and will evolve as my understanding and the tools evolve. This is also pretty basic, but it is intentionally so because some of this doesn’t have to be complex to immediately get better insights. ...

AI is going to fundamentally change the fabric of our society. Dr. Eric Schmidt suggests that within a year the vast majority of programmers will be replaced by AI programmers. Satya Nadella shared that Microsoft engineers are using AI to help write software (up to 20-30%) in a recent chat with Mark Zuckerburg. Even DARPA is creating exploratory models of Human-AI teams to understand how Humans+AI will work together. AI is forcing us to reflect on the work we do and ask ourselves the question of whether it has value or not. ...

AI, Agents, Workflows, LLMs - it all gets a bit confusing in today’s AI landscape around what all these mean. I want to break this down into a fairly simple framework to think about 4 distinct ways of using LLMs that get increasingly more complex. 1. Standalone LLM (Prompt in, Information Out) Standalone LLM This is the simplest and easiest way to engage a LLM. You enter a prompt into an LLM and it will respond based solely on its training data and the input you gave it in the prompt. This is most useful when you want a quick, self-contained answer that doesn’t require up-to-date data, personalization, or external reasoning, e.g. summarizing a paragraph, rewriting text, generating titles, or writing social media copy. You can also give it in-prompt context to help improve the answers - a concept often referred to as one/few-shot learning. ...

Model-Context-Protocol, or MCP as it’s more commonly referred to, came about in November 2024 when Anthropic introduced it to the world. But what is it really, why do we need it and how to get started? What is it? In its name, there are 3 elements: Model: This is the brain (the AI). It understands the question, figures out what it needs and when to ask for help. Context: Context of the situation (e.g. historical chat transcript) ...

The UK government recently released a report on their cross-department adoption and usage of M365 Copilot with some interesting results. It was an experiment that ran for 3 months across 20,000 government employees (Sept 2024 - Dec 2024) with a main goal being to understand how AI could impact productivity. Top-Level Insights From a positive persective, the key findings are as follows: Average Time Saved: Users reported saving an average of 26 minutes per day using M365 Copilot—equivalent to 13 working days per year. ...