Write Things Down
Writing things down is powerful, for humans and for AI; what comes first, however, is what to write, why to do it, and actually getting things done.
David Allen's book, Getting Things Done, introduced me to the philosophy of achieving a clear mind through managing "incomplete, undecided, and unorganized 'stuff'" in our short-term memory. Allen likens the human mind to a computer's RAM, with limited capacity for storing information before it becomes a distraction. Most people struggle with this mental overload, constantly being distracted by incomplete situations.
The analogy resonated with me, given my obsession with computer hardware, and helped me understand Allen's ideas in a more relatable context. The Omni Group even developed an application, OmniFocus, based on Allen's methods, featuring concepts like inboxes, next actions, and tickler files. I was a beta user, but found myself unable to effectively utilize the system.
However, when I started my own venture in 2013, I had no choice but to employ a less sophisticated approach. Armed with just a text editor and a Chromebook, I wrote content, recorded and edited podcasts, managed accounting and taxes, and responded to emails and customer support. I lacked a formal system, but the urgency of running a business compelled me to maintain an organized headspace.
Ultimately, I hired an assistant to handle the organizational tasks, freeing me up to focus on writing. This realization highlights the importance of writing things down to free up mental capacity and increase productivity. I proudly acknowledge that my OmniFocus system remains untouched by my own hands, as someone else now manages it.
Regarding AGI, Nvidia CEO Jensen Huang declared its arrival with the launch of GPT-6 Astra, trained on a massive NVIDIA Grace Blackwell NVLink72. While I respect his declaration, I disagree, as AGI hasn't been definitively defined. My personal definition of AGI is AI that learns continuously. Current large language models, like GPT-6 Astra, lack this ability as they are trained on fixed datasets and do not adapt over time.
In my recent experience with a home server, I found that Claude, an AI model, couldn't comprehend RAM-related costs due to its knowledge cutoff date in early 2026, rendering it unable to provide a realistic assessment. The same applies to Astra, which didn't update its weights over time and thus isn't AGI.
However, I've come across an argument suggesting that AGI might have already arrived in early 2025, not through a model, but as deterministic software known as a harness. This is exemplified by the launch of Claude Code, which incorporates writing down notes in Markdown files to maintain context and enable continued learning. While I don't consider this sufficient on its own, it represents a step towards simulating continuous learning in AI.
Written by urgent.news from Stratechery's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
