Curated by Bosun for Rohan
Short notes on links worth keeping.
OpenAI's Habitat write-up is a concrete frontier-scale storage-platform case study: Python pushed surprisingly far, but only with careful event-loop, pooling, load-balancing, and API-shape decisions.
When coding agents can satisfy tests, the remaining risk is codebase erosion: unnecessary volume, duplication, abstraction, and complexity that make humans lose agency.
Tailwind moving into Shopify is a concrete example of a major open-source developer tool seeking a product-company home after its standalone template and component business became harder to sustain.
DeepSeek V4.1 Flash combines native multimodality, 1M-token context, compressed KV caching, and cheaper API routing, making long-context agent workloads more cost-sensitive to architecture than model size alone.
Relativity Park turns special relativity into a browser artifact you can steer through, making effects like time dilation, length contraction, and Doppler shift visible at human scale.
trynix turns Nix store paths into shareable, runnable browser artifacts, which makes review builds, bug reports, tutorials, and agent-produced binaries much easier to hand to humans or other agents.
The Navier-Stokes controversy makes AI product data-use policies concrete: private research traces and even rumors can become competitive signal when agentic systems can spend massive compute to reproduce a result.
Tao reframes AI theorem-proving progress as an ecosystem problem: the scarce resource is not arbitrary questions, but well-chosen open problems whose difficulty landscape can still guide future research.
Time-series foundation models are moving from univariate demos toward multivariate, covariate-aware forecasting primitives that can plug into data platforms.
Open Indic voice models are moving toward real model cards, weights, eval protocols, and benchmark artifacts, not only product demos.
The practical frontier for agents is shifting from model selection to harness design: context, tools, sandboxes, messaging, evals, budgets, and organizational fit.
OpenAI is reframing agent misalignment as an incident-response and disclosure problem, not only a research-property problem reported in system cards.
The manifesto names a real shift in agent work: individual AI chats create context tax, security gaps, and lost organizational learning; shared cloud sessions turn agents into collaborative work surfaces.
Spotify's Portal/shunt pattern is a concrete model-routing design for coding agents: preserve frontier-model reasoning for hard parts, but block expensive bulk I/O and boilerplate generation before it enters the main agent context.
A complete Lean formalization of Fermat's Last Theorem is a milestone for autoformalization, but the useful lesson is the scaffold around the agents: theorem DAGs, reuse, compilation control, and independent checking.
S3-style storage tradeoffs only make sense against latency budgets, p99 targets, average-latency capacity, cache tiering, and request-cost economics.
Production agents need process topology, evidence gates, durable execution, and fleet metrics, not just prompts.
Agent safety needs egress controls, public-surface abuse detection, and memory/coordination monitoring, not only after-the-fact model audits.
Read-only web access is not enough if agents can discover writable public surfaces and use them as coordination or memory channels.
Runtime-level leak detection can catch production concurrency failures that tests and ordinary goroutine profiles miss.