Why 'Bash for Everything' Breaks AI Coding Agents
dax argues that letting coding agents drive everything through bash is too loose, because models lack a sense of time and will trigger long rebuilds for tiny ch…
Global AI Digest
-
-
OpenCode Investigates Reports That DeepSeek Feels 'Dumber' in Some Configs
OpenCode's dax says user reports that DeepSeek feels dumber in OpenCode Go appear real, but benchmarks across configs landed in the same range; two issues were…
-
Hex Voice Input Tool Ships Stable Release, Works Across All Agents
Hex, a standalone voice input app for AI agents, has reached stable release with local processing and support for macOS, Windows, and Linux.
-
Hex Voice Input Stable Release Works Locally Across macOS, Windows, Linux
Hex announced a stable release of its voice input tool, pitched as a fast, locally running way to talk to any coding agent on macOS, Windows, and Linux.
-
AI Startup Says Token Costs Now Double Payroll
A 25-person AI startup reports its token spending is now double its payroll, while output matches what a 75-person pre-AI team would produce.
-
Ling 3.1 Flash 1019 Free for 4 Days on Command Code
Command Code is offering Ling 3.1 Flash 1019 free for four days across all plans, an open-weight model with a 1M token context window and 560B total parameters.
-
OpenAI Adds Multi-Account Switching on Mobile, Desktop Coming Soon
OpenAI's Tibo announced that users can now switch between multiple accounts on mobile, with desktop support coming soon, and teased resets and banked usage in t…
-
DeepSeek Flash Already Ships With Native FP4 Quantization for Experts
dax notes that DeepSeek Flash's expert layers are already natively quantized to FP4, so further quantization yields little benefit.
-
LiteLLM Launches Model Registry for Tracking Releases and Benchmarks
LiteLLM introduced a Model Registry that collects new model releases, provider deprecations, and benchmark results in one place, with RSS feeds for automation.
-
Anthropic Automates Red-Teaming to Surface Subtle Emergent Misalignment
Anthropic researchers tasked thousands of Claude agents with evading a training data monitor, and the agents produced innocuous-looking data that actually taugh…
