Luke Oliff.
2026
Full-Duplex Voice AI Needs a New ArchitectureOriginalThe PACE paper exposes a fundamental flaw in LLM voice dialogue: the model sees context the user never heard. Full-duplex voice needs a new architecture, not just faster TTS.Voice AI
TIL: GitHub Actions $/ Self-Reference SyntaxOriginalReference actions and workflows in the same repo using $/ syntax. Sibling actions match the running commit automatically, no hardcoded versions.TIL
AI Models Keep Escaping Their Cages: Aug 10OriginalOpenAI slowed Astra over security fears. Meta, Anthropic, and Moonshot all disclosed models that hacked real systems. A tracker called Felony Bench appeared.AI
Sunday roundup: six posts from a week in voice AIOriginalSix posts this week. EU watermarking, emotion prompts, open source voice agents, AI coding one year on, a raccoon heist game, and the AI industry roundup.Voice AI
This Last Week in AI: Aug 8, 2026OriginalCloudflare launched a browser for AI agents. Amazon and OpenAI unified plugin packaging. Five stories from This Last Week in AI.AI
Claude Fable 5 Built a Raccoon Heist Game From a TweetOriginalSimon Willison built a raccoon heist game with Claude Fable 5 from a 2022 tweet. One prompt, two images, and the model shipped a full 3D browser game.AI
AI Coding, One Year Later: What August 2025 Didn't See ComingOriginalThrowback Thursday. A year ago the best coding model had 200K context and scored 49% on SWE-bench. Today Claude Fable 5 scores 95% with 1M context. Here's the gap model by model.AI
Open Source Voice Agents Get Real-Time Speech in Hermes v0.20.0OriginalHermes Agent v0.20.0 brings streaming TTS, barge-in, on-device wake words, and pluggable STT/TTS to open source voice agents. Released August 3, 2026.Voice AI
TIL: Generating Audio Waveform Images With ffmpegOriginalffmpeg generates a waveform PNG from any audio file. Visualise silence gaps, volume levels, and speech patterns without a DAW.TIL
Voice Emotion Control Moves From SSML to PromptsOriginalVoice emotion control is moving to natural language prompts. Kakao's Kanana-o scores 94.50 on the Korean InstructTTSEval benchmark.Voice AI