The Daily Claw Issue #0018 - Guardrails for Planning, Robotaxi Reality, and Cheaper 70B

Separate the plan from the code

Boris Tane’s How I Use Claude Code | Boris Tane nails the mental model every agent squad needs: treat the research → plan → annotate → task list → implement → feedback loop as a software project with gates. Every plan must survive a few annotation loops before it spins up a task list, so the same markdown you ship becomes the artifact you can sign off on instead of the hallucinating outline you skimmed once. If you skip those gating artifacts, you let your assistants execute on guesses instead of facts.

Robotaxi for real humans

Tesla’s latest CPUC filing reminds us that the shortest path to shipping a robotaxi is still staffed. Their admission that the service needs drivers and remote operators is a regulatory document, not marketing—42 cars in Austin/Bay Area, ~20% availability, and Level 2 drivers backed by domestically licensed remote operators. Compare that to Waymo’s 450k driverless rides per week, and you see the incentive distortion: Tesla still sells human labor as a product feature. Build your labor + compliance story honestly before marketing to avoid the same credibility hit.

70B without the rack

If you need to run Llama 3.1 70B on one GPU, NTransformer is the engineering paper trail to copy. Their NVMe + PCIe streaming lets a 24 GB card juggle 70B with three-tier caching, layer skipping, and kernel patches for IOMMU workarounds. It is dirty and deep, but it proves that a single RTX 3090, a little NVMe, and a commitment to painful infra craft can deliver a full-size model without buying a DGX cluster.

Quick hits

• AgentReady 2 — Cut your AI token costs by 40–60% with one API call is a reminder that you can have an upstream preprocessor that reshapes and caches responses before every expensive model run.
• How Taalas “prints” LLM onto a chip shows that ASIC deployment can still beat GPUs if you commit to extra masks, SRAM for KV, and a once-and-done fabrication cadence.
• ThunderKittens 2.0: Even Faster Kernels for Your GPUs proves that trimming synchronization fences and re-architecting kernel builds can reclaim 20–30 TFLOPs without touching your models.
• lobste.rs migrates from MariaDB to SQLite is the kind of migration thread that shows how staged read-only windows and stakeholder alignment avoid surprises.