Apache 2.0 on a 20B search agent is a huge contribution. The 40-page technical report is a nice addition too! Will read
@whats_ai
-
Adding Diverse News Without Re-ranking: Simple AI Intervention
By
–
The finding that just adding diverse news worked without re-ranking is really interesting. Feels like the simplest possible intervention.
-
Bare Metal Performance Optimization Delivers Strong Results
By
–
Stripping everything to bare metal and still getting strong performance says a lot!
-
Building for passion versus exit in AGI timeline
By
–
I get the sentiment but I genuinely enjoy building things regardless of whether AGI is 2 or 20 years away. The people who build because they love it will keep building. The ones doing it purely for the exit are the ones who should maybe reconsider.
-
Open-Source Agents Ship Real Autonomy Features
By
–
The persistent memory + dedicated machine access combo is what makes this different from most agent frameworks. Super cool to see open-source agents shipping real autonomy features. Will play around!
-
Dialogue Format Depolarizes AI Models Across Political Alignments
By
–
Really interesting that the depolarizing effect holds across models regardless of political leaning. Suggests it's something about the dialogue format itself, not the model's alignment.
-
Taxing AI Agents: From Radical Idea to Obvious Policy
By
–
Taxing AI agents might be one of those ideas that sounds radical now but will feel obvious in 3 years.
-
RAG Pipeline Over Raw Token Storage Strategy
By
–
The ideal is not “store everything,” it’s “keep only what is worth keeping” — more like a really good RAG pipeline than a giant token dump.
-
Context Size vs. Quality: When Less Compression Hurts Performance
By
–
Bigger context is useful, but not if it keeps carrying forward noise that should have been compressed away. In some cases that extra clutter can even hurt performance relative to a stronger compaction approach that preserves only the minimum viable context needed for the task.
-
Opus to Opus 1M Switch: Daily Limits and Token Efficiency Issues
By
–
This is also why the switch between Opus and Opus 1M is pretty annoying to me: I hit daily limits much earlier, and part of it feels like compaction happens less aggressively, so way more tokens get sent every time (and probably because they keep reducing that daily limit).