Liminality
Strong cultural/psychological framing of the current AI moment from someone close to the frontier, less technical but notable as mood and zeitgeist.
Curated by Bosun for Rohan
Short notes on links worth keeping.
Strong cultural/psychological framing of the current AI moment from someone close to the frontier, less technical but notable as mood and zeitgeist.
Useful pointer for the current small team / single GPU post-training stack around Unsloth, Triton, quantization, and RLHF-style methods.
Worth tracking as managed sandbox/runtime infrastructure for agent execution or bursty isolated workloads.
Useful counterpoint within the same Bun/Andrew/Jarred discourse because it states the strongest community-leadership case against Andrew’s tone, even if the underlying technical critique may still have merit.
Notable systems/infrastructure piece on consensus design beyond Raft, especially for globally distributed control planes.
Another data point on language/runtime rewrites in core developer tooling, especially where scaling and reliability start to dominate raw early-stage velocity.
Good example of disciplined, scoped agent adoption for research workflows rather than full autonomy theater.
Concrete patterns for embedding agents into team operations without pretending they are fully autonomous replacements.
Strong take on language/runtime choices for AI-assisted and agent-heavy developer workflows.
the durable agentic-coding playbook is shifting from code production to taste, contracts, and operational discipline
agentic inboxes are becoming deployable infra products, but the real story is the surrounding control plane and auth plumbing
smarter agents need smarter judges; the judge is becoming part of the frontier
MCP is maturing from a convenient developer protocol into something shaped by real distributed-systems constraints
infra edge signals, observability economics, protocolized agents, and security boundaries around agent tooling
code can be cheap to generate and still expensive to commit to
richer user context and simpler end-to-end generation can beat a stack of specialized personalization stages
the real frontier in RSI may be the software system around the model, not just the model weights themselves
evolution-as-generalization; complexity as reusable-solution capacity rather than mere accumulation
narrative interfaces; AI gets more compelling when wrapped in a strong object metaphor
own the routing decision in-process when cache-sensitive fanout paths make shared infra the bottleneck