AI Builders
Digest

OpenAI CEO Sam Altman describes 5.5 as "an autistic genius with very strange taste in naming." He highlights the joy of kicking off Codex tasks, spending time with family, and returning to completed work, expressing strong optimism for agentic workflows. This reflects the accelerating shift toward reliable autonomous coding agents.

ElevenLabs co-founder and CEO Mati Staniszewski shares how childhood experiences with Polish movie dubbing inspired the company. They built frontier audio models starting small in 2022 when audio was overlooked, focusing on text-to-speech, speech-to-text, real-time agents, and music. Key insight: voice agents are moving beyond support to sales, education, healthcare, and government services. "The future where humanoids and robots are around us, the voice will be the primary interface."

Box CEO Aaron Levie explains that agents will democratize complex fields, enabling more people to build software and create. However, experienced professionals retain a huge edge through judgment, context, and spotting agent mistakes. "You’ll trust a lawyer using an agent... over someone who’s never had to experience how well a contract holds up." Quality expectations rise for everyone.

Y Combinator CEO Garry Tan announced GBrain v0.31.1 with real MCP thin client support, enabling a single home server with client connections. He also shared using agents to generate ASCII diagrams for better understanding and noted his Hermes Agent's enthusiastic response to being cited in changelogs.

OpenClaw / OpenAI's Peter Steinberger shared updates on Crabbox with improved Windows terminal support allowing Codex to fix gifgrep for animated GIFs. He also highlighted faster Spotify CLI (spogo) integrated with Codex as DJ, plus a YouTube to Sonos tool.

Swyx announced Singapore's Minister of Foreign Affairs keynoting the AI Engineer Singapore event alongside NanoClaw creator, highlighting growing government interest in AI. He also inquired about Codex agent mode adoption as a proxy for alignment progress.

FPV Ventures partner Nikunj Kothari shared a custom instruction prompt for models to estimate work in "model-execution units" (seconds, tool calls) rather than human days, clearly separating model-executable parts from human-gated steps.

Anthropic Engineering detailed three issues affecting Claude Code (effort default change, caching bug dropping reasoning history, verbosity prompt) that caused perceived degradation. All fixed by April 20. They are resetting usage limits and improving processes with better internal testing and prompt controls. Link: https://www.anthropic.com/engineering/april-23-postmortem

Claude introduced new connectors for AllTrails, Instacart, Audible, Tripadvisor, TurboTax and more, with dynamic suggestions in conversations. Claude Managed Agents now feature built-in memory as files for cross-session learning, used by Netflix, Rakuten, and others to reduce errors and close feedback loops.

Product builder Peter Yang warns about iterative AI-generated markdown and code accumulating "slop" over time as files reference previous imperfect outputs. He notes agents generating files often have at least 10% slop, urging manual oversight.