What I Iike to read (feedback from @claudeai after a month of sharing my reading habits):
"classic shiny-object trap: a gorgeously-written, profound-feeling piece that sits almost entirely outside your circle of competence."
Ha!
Skeuomorphism.
What a cool word.
“Skeuomorphism: Designing a new thing to look and behave like the old thing it replaces, even when the imitation serves no functional purpose.”
How does this relate to AI?
Early movies were just like plays, but on a big screen. Early web pages were just like magazines, but online. The new medium arrived, and we made it look familiar.
Just like in the early days of the internet, no one had yet thought of social media. Entire jobs/companies/industries are going to change.
For me, AI is doing my research and rote processes much better. This is the same workflow I would have done 5+ years ago, just faster and better. This is not peak AI. I/m trying, but I am not there yet.
How does this relates to @SymetryML ?
Today's standard model for data-driven alpha:
more data + more compute = better signal.
Those with deeper pockets to buy more data and run more rows win. The unit of work is an ever-expanding collection of billions of rows. AI might help you ask different questions, or process them faster. But it's still the same race.
SymetryML changes the unit of work entirely.
SymetryML changes from “re-scan the rows” to a “statistical twin” that never grows with new rows, just updates.
The compute constraint disappears. And when the constraint disappears, so do the old questions. You start asking questions that weren't previously possible.
Ship the statistics, not the data
- Don't move the data. Move what it learned.
- Stop sending the data. Send the math.
- Learn from each other's data. Never see it.
...and do this at scale, in real-time.
@SymetryML
@interlatent loved this: "An Overview of Modern AI Robotics from First Principles."
Cleanest articulation of the edge/cloud brain problem I've read.
One thing the binary hides: it assumes the model is monolithic, so you're forced to choose where it lives. At @SymetryML we don't relocate the brain, we federate it. Each edge node learns locally (latency ≈ 0, and our model is fixed-size, so it always fits, no shrinking). The local models then merge algebraically into a global one: no raw data moved, no per-action round trip.
Learn on the edge, pool in the cloud. Different domain than robot control, we're the real-time behavioral/detection layer, not the policy, but the escape from your tradeoff is the same: federation, not relocation.
Lots of people sharing "here is the best way to work with AI" ... my solution .... have my AI automatically review them all, and given that it knows how I work and how I like to work ... keep only the most relevant things (if anything) that would add to my process. I am finding my workflow is learning and getting better at knowing what would be additive to my process & where the holes are!
Everyone will have their own personal best practices.
Heard from colleagues today regarding some of the work I’ve been putting out there...
Q: “how are you doing this?”
Short Answer: “Flow OS”
Long answer: "I've built a personal setup on top of Claude Code I call Flow OS: a file-based memory layer so it remembers my contacts, ventures, and past work across sessions, plus a Telegram bot so I can message it from my phone. The "insights" you saw came out of that — it's not vanilla Claude, it's Claude + my accumulated context."
Thanks @virtual_rf !