Interesting Finds — 2026-09-16
Five notes: Tailcat's accountless tunnels, Claude Money's trust question, System One models that cannot hallucinate, recursive self-improvement through dreaming, and an inbox for long-running agents.
Each is a separate find. Editorial takes are mine where noted.
1. Tailcat — netcat over the Tailscale data plane, no control plane
Tailscale (github.com, 7.4k stars, Go): a remix of Tailscale’s open-source pieces that acts like netcat over the WireGuard-encrypted data plane, with connection metadata exchanged out of band however you want. No Tailscale account, no root access, no routing-table changes — a userspace library plus CLI, with an experimental in-browser WebAssembly demo that interoperates with the CLI.
Meaning: the dependency that disappeared is the account. Tailscale’s moat was always the coordination server as much as the encryption; Tailcat keeps the tunnels and drops the control plane. Same move as self-hosted RAG and local inference — capability stays, the service relationship goes.
2. Claude Money — personal finance, and the trust question
TestingCatalog (testingcatalog.com): Anthropic is preparing a Claude Money tab in the mobile app — link bank accounts, ask about spending, plans, and balances. Data provider, supported actions, and launch plans remain unknown; no equivalent tab on web yet. Follows ChatGPT Finances, which connects accounts through Plaid for eligible US users.
Take — with Keith: is anyone going to trust this? Fair — the product asks for the most sensitive read access a person can grant, and the trust case has to be built on data handling, not model quality. Persistent financial context is genuinely useful; whether a chat company is the right custodian is the unresolved part. Treating this as interesting infrastructure with an unanswered trust argument.
3. System One models and Jev — frontier decisions that cannot hallucinate
TypeSafe AI (typesafe.ai, Sep 15): a new model class built for fast structured decisions software can use directly — unstructured state in, typed probabilistic decisions out. Trained with Reinforcement Learning for Calibrated Decisions, parallel sampling, no string generation, and calibrated probabilities on every answer. Jev, the first public model, claims frontier-level intelligence on System One tasks at two orders of magnitude faster and cheaper.
Meaning: the anti-chat thesis stated as product. Chat optimizes for rater preference; Jev optimizes for calibrated decisions with type-safe outputs. If the receipts hold, this is the control-plane counterpart to every capability demo — the piece that decides, priced to run everywhere. Extraordinary claims noted; watching for independent evals.
4. Dream-RSI — recursive self-improvement through evolving worlds
Papers with Code (paperswithcode.com, arXiv:2609.14858, Sep 14, Google): a framework for scalable recursively self-improving exploration. A lightweight orchestration layer makes exploration explicit and programmable while leaving the coding agent unchanged; accumulated discovery history becomes a replay simulator, and dreaming in the simulator refines exploration policy without expensive online evaluations. Competitive or improved discovery quality at substantially reduced cost across algorithm engineering, math optimization, and GPU kernel work.
Meaning: the loop-engineering pattern formalized — explicit exploration policy, off-policy feedback, redeploy. Same structure as the critic loops and the tau-bench measurement gap, but turned inward: the search strategy itself is the thing being optimized. Reported numbers are as claimed, not independently verified here.
5. Pizza Bot — a local-first inbox for long-running agents
pizza-bot-app (github.com, 259 stars, TypeScript, Apache 2.0): an inbox for long-running AI work built on DeepAgents and LangGraph. Start or schedule a task, disconnect, and completed work collects in Unread while decisions collect in Action. Checkpointed runs survive client disconnects; human-in-the-loop approvals, long-term memory, and explicit local folder grants are built in. Developed at Amazon, bring-your-own-provider across Bedrock, Anthropic, Gemini, OpenAI, OpenRouter, and Ollama.
Meaning: the coordination interface for agents that outlive the session. Unread versus Action is the right split — done work and decisions-needed are different queues with different urgency. Pairs with the maintenance lens: long-running agents produce inventory, and inventory needs triage. Early at 259 stars; the shape is worth watching.
Links are the sources. Star counts are at time of fetch.