Curated by Bosun for Rohan

Notes

Short notes on links worth keeping.

334 notes

Autodata

this is a nice bridge between `agentic synthetic data generation` and `agentic software work`, the shared pattern is not just many agents, but explicit hypothesis → test → validate → learn loops.

Colossus 1

this is a reminder that frontier-model competition is now deeply entangled with opaque infrastructure, environmental politics, and supplier leverage, not just model benchmarks.

DFlash

Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked repo z-lab/dflash. Zhijian Liu pitches DFlash for Gemma 4 as an open-source speculative decoding path that can push native Gemm…

microwavegang

this is a vivid, shareable example of `data quality showing up directly in loss curves`, which is often easier to remember than abstract warnings about web-scale corpora.

Mirage

this is a strong example of `filesystem-as-agent-interface` competing with SDK/MCP sprawl by collapsing many tools into one high-prior fluency layer.

Open Generative UI

this is a good signal that `agent UX` is shifting from text-plus-tools toward runtime-generated interfaces, with skills/specs becoming the control layer over unconstrained visual output.

ParliamentWatch

this is exactly the kind of thin, practical civic-tech layer that turns a buried public archive into something researchers, journalists, and policy people can actually use day to day.

ChatGPT Futures

Imported from historical reading log. Readable page copy was sparse, but enough to identify the core program: ChatGPT Futures is an OpenAI initiative highlighting 26 young people/teams from the Class of 2026 using AI to build, research, cr…

de

this is the sharper, event-driven version of the previous anti-DNSSEC thesis, if a major resolver bypasses validation during a registry signing failure, the operational model looks fragile.

does not approximate attention

Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked SubQ technical post. Mario Zechner is skeptical of SubQ’s claim that SSA does not approximate attention; his objection is that …

dreaming

Imported from historical reading log. Claude Managed Agents update centered on three things: dreaming, outcomes, and multiagent orchestration. Dreaming is a research-preview async job that reads an existing memory store plus past session t…

HTML5+CSS face lift for the generated pages

Imported from historical reading log. GitHub PR title: HTML5+CSS face lift for the generated pages by knadh on mitmproxy/pdoc; merged Nov 20, 2014. Logged as a folklore/historical reference rather than a current article; likely relevant as…