LLMs will become more powerful but they won’t magically become easier to use as advancements are made Like any tech, applying LLMs is a skill that takes time and practice. Wrote more in depth about how to be effective with LLMs here

at work I have unlimited tokens and a leaderboard. the past 6 weeks I've been on paternity leave exploring consumer tools instead. a few thoughts: 1. real usage limits sent me back to pencil and paper. I used to dump every idea straight into an LLM session to explore; this is more cognitively taxing than I realized, much better to capture and prioritize first 2. claude code and codex are genuinely good. but the harness engineering I actually want to do, my own orchestration, instructions, tool surface, needs direct API calls, and that gets expensive fast. I burned $20 in an hour. cost prohibitive for personal projects. 3. unlimited tokens a leaderboard (top 1% btw) is a hell of an incentive to learn and experiment. I get that its not-sustainable (especially for non mega big techs) but I support it.

Your skill level acts as a ceiling for the agent So every skill you gain multiplies your output Never been a better time to git gud

Incredible alpha right now in buying a small notebook, high quality pencils and work through ideas on paper. While everyone else submits to the LLM's in-built judgement and ideas, you can easily stand out by ideating outside of an LLM context.

Seems like Anthropic team did not scale their web search services to accommodate Fable’s intense desire to research

After using Opus as an agent, there is no going back. Nothing quite like it. But for my hand-rolled agents, you have to use the API for $$$ yet codex lets you use the subscription. So I’ve landed on the perfect balance of Opus brain with Codex sub-agents

Most people prompt the AI, I correct it. Dozens of times a day the smartest AI in the world agrees with me. It’s lonely having this responsibility but I will do my best to use it for good.

This app is undefeated when the wife rattles off a marathon list of tasks, packing list and all of her hopes and dreams in one breath

eyyyy, I've been working like this forever. Mock => polish => backend.

Every Opus or GPT model release results in the same performative benchmarking. The skills that compound are on either side of the model. Knowing what you want before you start matters more than which model you picked. Wrote about how I do this x.com/hexednobility/status/2…

There is a healthy balance required when using AI to make things. Not trusting the LLM to produce amazing output on its own nor skipping the LLM all together. I wrote a bit about my mental model here.

AI is a force multiplier. You know what a force multiplier needs? A force. It makes your people better. It doesn't replace your people. That's not how *any* of this works. 2

I don’t know why people talk in absolutes about agentic design. Especially for coders and builders. We just invented this stuff. It’s wild to think it won’t radically change over the next year. We’re at the very beginning.

The most underrated ai engineering adjacent space is software ops & maintenance How can you effectively apply agents/LLMs to enable one person to operate a much larger surface area of software?

Working on resume Using claude to capture probe on work projects / anecdotes then carefully whittle down into clean resume bullets prepared interview stories. Claude has this remarkable ability to help think deeply, something I haven't really felt in GPT models.

adopting coding agents has made me way faster as an IC it also makes me appreciate good (full-time) eng managers more, not less a team of superpowered builders without alignment is a great way to blow tokens with no product to show for it

I enjoy using LLMs in my work, they’ve erased tedium and enable me to build FAR faster. But I do miss the craft and collaboration from the before times. When multiple talented engineers working closely together was assumed as necessary to build high quality software.

idk chief im having more fun than ever by a wide margin