6 essential concepts you should know about AI (LLMs): 1. It may sound obvious, but it’s important: it's all text in, text out. Terminology like prompts, memory, tools, RAG, context can sound complicated, but it’s all just text we send to the model.

AI is not changing what is valuable to a business, but it is changing your value as a software engineer. The new full stack is product + engineering, not frontend + backend. #youreabsolutelyrightpodcast

Life hack: have kids

Really not liking iOS 26.1 design. Don't like liquid glass at all, and the motion is so overdone it's nauseating.

Microservices is the software industry’s most successful confidence scam. It convinces small teams that they are “thinking big” while systematically destroying their ability to move at all. It flatters ambition by weaponizing insecurity: if you’re not running a constellation of

We are entering an era of building things we don't understand that work great. Seems like this could be a problem.

Added some wood veneer to the office last weekend

I'm old enough to remember when Netflix was a scrappy startup who was laughed at by Blockbuster, crazy to see this

There's an increasing amount of pushback on MCP, and rightly so. Cloudflare and Anthropic both found code execution directly against APIs to be a better solution than MCP. What this means for you: get familiar with code sandboxes, it will be a big 2026 story.

You don't always need an agent, you can get a lot done really fast with a single LLM call. More about all of this, including multi-agent systems in my article here:

6. Agents: An agent is a loop of LLM call and tools. The agent can take in a request, call a planning tool to create a task list, then loop and call tools until those tasks are completed.

5. Semantic search: keyword searches are old news. In the world of LLMs, we want to search by the meaning, or semantics. This is done by converting text into vectors, which are numerical representations that capture semantic meaning, then using mathematical similarity to search.

You can build tools on your own server and call them internally, MCP is a way to expose your tools to others. If you know lots of other people or teams want to use web search, you'd put it on an MCP server to share it.

4. Tools: actions we can take that the LLM understands and can help with, for example web search. The LLM doesn't call a tool, it tells you what tool to call and how. For example it could give you the search query, then you can call a search function in your code.

3. The context window is everything we send to the LLM, including prompts, tools, documents, etc. This window is limited, and the more you put in there, the worse the model performs. Your job is to use the right context at the right time, and nothing more.

Request augmented generation is adding context dynamically, usually with an API call. It's kind of like copy/pasting a document into ChatGPT, then asking a question about that document. Except with RAG, you get the doc dynamically, probably with semantic search.

2. Context: additional information we add, usually to the system prompt. This could be a document, a web page, information about a customer’s account, etc. Providing the right context at the right time is the most important thing we can do to improve model responses.

Etymology nerds know: the term Black Friday Cyber Monday is derived from the name Blök Frindeh Saubermande. An Icelandic soda shop owner who died in 1965. Every year on the Monday after American Thanksgiving, he'd shout "SAUBERMANDE SPECIAL" and huck soda at passersby. RIP Blök

Unpopular opinion: today's talent isn't up to date with the skills required in an agentic world. If you're an exception - great! 👉 But I was first alarmed during the summer of 2023 while hiring a couple of growth marketers for Growth Shuttle. Troubled by the lack of AI

I'm just an IC and the amount of messages I get on Slack is overwhelming some days. I don't understand how a CEO would be able to keep track of everything.