All Dispatches

Tokenizers v1 cuts BPE latency by up to 30× on Apple M4 Max, turning a CPU bottleneck into a rounding error. Here’s what actually changed under the hood—and why it matters for your GPU-fed pipelines.

DeepSeek’s new 552B-parameter MoE promises 75% less KV cache memory and half‑price off‑peak API calls. But does the shrink‑wrap architecture deliver real savings for SaaS agents, or just a tighter leash on your bill?

The Yangtze River Delta AI Safety Anhui Lab unveiled three integrated safety suites for LLMs, agents, and AIGC. Here’s what SaaS and IT leaders need to know about their real‑world impact.

Samsung plans to more than double HBM4 and HBM4E output next year, pushing 12-layer stacks to 80% of shipments. Your AI infrastructure bill just found a new reason to cry.

Google's open-agentic orchestrator AX promises to tame runaway agent costs and state management with declarative primitives. But does it solve real ops pain or just add another YAML layer?

The AI-GRACE preprint promises to turn lofty agentic AI ambitions into deployable architecture. We tear apart its layers to see if it solves real enterprise woes or just adds another slide to the hype deck.

Classic benchmarks measure throughput on empty disks. This project tests what actually kills production: snapshot aging, ENOSPC behavior, corruption self-healing, and fsync tail latency.

Vendor SLAs promise uptime; OONI's crowdsourced probes show you which governments actually flipped the kill switch. Run it or guess.

A seven-year Rust veteran reimplements JSONPath in Zig and discovers why 'explicit allocators everywhere' sounds clean until your error paths leak, double-free, or orphan allocations — and Rust's Drop trait makes those bugs impossible by construction.

A GET-only API proves model weights walk out the door through URL parameters while safety researchers debate whether exfiltration is "overrated."

PlanetScale's new TIN extension promises order-of-magnitude speedups over ParadeDB and pg_textsearch by using Postgres ctid values directly. The benchmarks look impressive — but they're all from PlanetScale's own fork of a competitor's benchmarking tool.

OpenAI published a six-pillar youth safety blueprint in Australia — a month after ChatGPT for Teens went live — and every pillar is a promise, not a commitment. Here's what the document does and doesn't buy Australian regulators, schools, and families.