All Dispatches

Bonsai 27B is the first 27B-class model to run on a phone, enabling local execution of AI workloads and reducing dependence on cloud APIs. The model comes in two variants: Ternary Bonsai 27B and 1-bit Bonsai 27B, with the latter being footprint-oriented and suitable for phone-class devices.

Apple's new SpeechAnalyzer API has been benchmarked against Whisper and its predecessor, showing significant improvements in accuracy and speed. The results have major implications for developers and users relying on speech-to-text transcription.

AWS will stop accepting new Mechanical Turk customers on July 30 2026, signaling the end of a once‑crucial crowd‑labelling platform for AI pipelines.

LARP lets founders record revenue without moving cash, but the legal and financial fallout may outweigh the headline‑grabbing ARR.

A dormant stack‑use‑after‑free in the rtmutex implementation lets an unprivileged process hijack kernel control flow, affecting every Linux distro released since 2011.

Ploy’s production AI agent moved from Claude Opus to GPT‑5.6 Sol, slashing build time in half and cutting spend by over a quarter—after a disciplined migration that reshaped evals, tool schemas, caching and reasoning.

AI can now prove theorems, but policymakers and executives risk treating that output as a substitute for deep mathematical capacity—a critical infrastructure that cannot be rebuilt overnight.

Since Chrome 148, the host libm backs Math.tanh, allowing anti‑bot systems to identify the underlying OS from a single floating‑point probe.

A side‑by‑side measurement shows Claude Code injects ~33k tokens before a prompt, while OpenCode needs only ~7k—fueling higher costs, latency, and cache churn for production AI agents.

A CSO report shows Ireland’s server farms surged to 23% of total electricity use in 2025, prompting new grid rules and fresh cost‑risk calculations for operators.

A tiny 2 GiB test node exposed a Go context leak in kubelet that can balloon memory usage, but profiling and a fast‑tracked patch resolved it for production clusters.

Kimi K2.7 Code brings an open‑weight model to GitHub Copilot, but enterprises must follow a concrete rollout plan to control cost, security, and compliance risks.