Back to the bench
Like any technologist with an internet connection, I'm feeling the magnitude of the moment, day in and day out. It's a single data point, but yesterday's release of Opus 4.7 appeared to continue the trendline. I did an informal personal eval on the model -- which involves talking to it about some philosophical and technological questions I've been thinking through for the past decade+ -- and was impressed by the sense that I was having a conversation with a very intelligent counterpart.
Claude read my past writing in moments, called up the key relevant thinkers in the literature (many of whom I'm familiar with, and many more who I've never encountered), and most importantly both articulated what I've been thinking accurately and concisely, and helped me uncover new insights, form new connections, and identify actionable paths forward to develop and deliver these ideas.
We're hitting the elbow of the exponential, and I'm feeling mixed emotions, equal parts awe and angst about it. The fact that I've been reading both Superintelligence and The Fall of Hyperion -- concurrently -- is a big contributing factor, as is the closed release of Mythos last week, even if one holds a reasonable sense of skepticism about some of the claims and the significance of the capabilities.
(A diagram from Chapter 4 of SuperIntelligence: Paths, Dangers, Strategies by Nick Bostrom, 2014. It feels like we're in "crossover" territory.
I've been a bit quiet on my research notes for a while. This is mostly because I've been focused on personal things -- most notably welcoming a new family member. I've also been getting Astral into a stable position, and thinking about the future of the research (more on this very soon).
Another reason is that I've been building my "system".
There's a tension here, between "keeping up" and doing actual work. I remember seeing a tweet about Clawdbot, now OpenClaw (don't ask me how I feel about the branding) in early December and thinking, to my point earlier about mixed emotions: this looks incredible // this looks like a security nightmare.
AI-driven software engineering has accelerated the pace of innovation so much, and the bleeding edge is very bloody indeed. Every third post on my timeline is something like "It's over! x just release y!" -- I could spend all my time scouring every new thing, and never get anything done. (It feels very DeFi Summer in more ways than one ... a post for another time.)
Tempted as I am to jump on every new thing, I have chosen to sit a few months behind it, steering clear of the riskiest and most exciting edge of the wave, instead letting these things "mature" for a little while before I dive in. Of course I watch the latest big releases and test them out, but I'm also building AI into my workflows and entire digital life, and that's where a bit of patience is paying off.
I notice that probably by necessity, the innovators are messy, sweaty, a bit unhinged. Innovations, especially in agentic systems, are naturally going to spawn in places where deep consideration, rigorous security practices, and elegant, refined patterns are not a priority -- because those design properties, quite necessarily, take time to form. It usually takes 3-6 months to go from a vibecoded Steve Yegge release (the sweatiest) to a Claude Code feature. My risk-benefit calculus has me waiting a bit; stability is valuable. Plus the rapid speciation of approaches gives me a lot more to consider -- now we have gbrain, hermes, openclaw, and months of writing and code, for reference.
So yes, I'm not riding the skirt of the wave. I'm watching it roll by, sweaty surfers having a blast and getting wiped out. Part of me worries that the early bird gets the worm, but then I remind myself that it's the second mouse that gets the cheese.
In any case, I'm back from a glorious holiday in Morocco with my partner (I finally took some kiting lessons ... more mixed extreme emotions there), and settling in. I hope to be posting more, there's a lot to say.