Validation, Then Absorption
I built NewsAgg because keeping up with AI research by hand doesn’t scale, and I’m shipping fast enough now that falling behind the field is a real risk rather than a vague worry. It clusters the same story across outlets, flags where they disagree, and leads with a ranked feed of what’s actually worth reading: scored by a local model, with unreliable sources pruning themselves, no human in the loop.
I built it to read the field for me. It turned out to read me instead. Two functions, neither of which I designed, both pointed back at my own work.
Half of it, I already knew
The recurring feeling, scrolling the feed most mornings, is: I already built this. Not because I’d read the paper first and implemented it. The opposite. The problem forced the shape, I built the shape under pressure, and only later did the feed show me the field had named the same shape and arrived at it too.
The clearest example is the fleet itself. A while back it kept hitting a capacity wall on its strongest model, so I split it in two: a planner on the expensive tier that reads a repo and decides what’s worth building, and a cheaper, faster executor that just implements the plan. I did this to survive a quota ceiling. Then the feed kept surfacing the same structure under tidy names (plan-and-execute agents, orchestrator-and-worker splits) as if it were the obvious design. It wasn’t obvious to me; it was forced. But we ended up in the same place.
It keeps happening. NewsAgg’s own value triage, a small local model scoring every article, is just LLM-as-judge, which the field treats as the standard way to filter model output at scale. My deny-by-default redaction gate, the one that refused to publish a private address last week, is what the literature calls a guardrail. The source-reputation system that demotes unreliable outlets over a decaying window is a reputation-weighting scheme I’d have recognized instantly if I’d read about it, but I didn’t; a source lying to me did.
This is convergent evolution, and it’s the strongest kind of validation there is. When you reach the same structure the field reaches, from a different direction (need instead of literature), nobody handed it to you, so it isn’t fashion. It’s evidence the shape is real. And it’s cheap confidence: I no longer have to wonder whether the planner/executor split was a good instinct. The field found it independently. I can stop relitigating it and spend the doubt somewhere it’s still useful.
The other half, it absorbs
The rest of the feed is the part I didn’t already know: techniques I hadn’t built and mostly wouldn’t have thought of. This is where the loop stops being a mirror and starts being an upgrade path.
The picture I keep coming back to is Majin Buu. In Dragon Ball Z, Buu didn’t just beat the fighters he faced; he absorbed them, and kept their techniques, walking away from every fight permanently stronger and carrying every ability he’d ever met. That is almost exactly the shape of what happens here. A technique surfaces in the feed. If it’s worth having, it becomes a task for the fleet. The fleet implements it in one of my repos. And now the system has that capability: folded in, kept, applied to everything it builds afterward. A cheaper way to verify work before committing it; a trick for compacting context so agents run longer before they lose the thread; a retrieval refinement. Read, scoped, built, kept. The thing that reads the field is the same thing that rewrites itself to match it.
The filter matters, and it’s the honest caveat. Buu absorbing everything indiscriminately is how Buu also gets slower and dumber in some arcs, and the analogy holds: absorb every shiny paper and you drown. Most of what scrolls past is noise, or real but not for me. The value triage is the first filter; my own judgment is the second. Only a few things a month actually get built in, because the maintenance cap is real and a capability you adopt is a capability you now have to keep working. So absorption is deliberate: the feed proposes, I dispose, and the system gets stronger a little at a time rather than bloating on every technique it meets.
The loop closes on itself
Here’s the part I find slightly eerie and mostly great. NewsAgg is one of the repos the fleet maintains. So the sensor that tells the system what’s worth building is itself something the system builds. It reads the field, decides a technique is worth having, and improves a repo. And one of the repos it can improve is the reader, so it can get better at reading, so it can find better techniques to absorb next time.
Validation tells me the parts I already have are sound; I built them right, under pressure, before I had permission from a paper. Absorption adds the parts I was missing. Between the two, a tool I built to keep myself current became the mechanism by which the whole thing stays current on its own, using the outside world’s research as raw material to upgrade itself, on a schedule, while I’m off doing something else.
I still read it most mornings. Half the time it tells me I was right. The other half it hands me my next move. Some days I read it back just to find out what I apparently believe now.