OpenAI's Dev Day Was a Disappointment & T3 Code's New Orchestrator Shines
Summary
Theo and Ben discuss OpenAI’s Dev Day announcements, noting a lackluster reception compared to recent Anthropic releases. They dive into T3 Code’s new Orchestrator V2, which enables cross-provider agent delegation, and share workflows using "Soul Loop" and "Full Send" skills for automated code review. The hosts also analyze the performance gap between top-tier closed models and open-weight alternatives like GLM and DeepSeek, while Ben shares his experience using the Tesla’s "Grok" integration for hands-free work.
Topics discussed
Intro, Tesla Grok joke, and PostHog sponsorship
AI subscription costs and usage limits
T3 Code features: visualizations and PR watcher
Cache management and compaction strategies
Model capability gaps and product decisions
Agent feedback endpoints and Linux episode tease
Grok FSD speed limits and Waymo discussion
Gemini potential and open-weight model landscape
Small model performance and local inference use cases
Orchestrator V2: multi-agent delegation in T3 Code
Agent skills, prompts, and the Tasks XKCD analogy
Voice input accidents and agent history auditing
G-Stack controversy and P-Stack skills
Model tier classification and lab profit margins
OpenAI pricing strategy and cache read discounts
AI adoption myths and T3 Chat business model
Usage data analysis and the $200 plan economics
Ultra Fast mode, Cerebras, and NVIDIA hardware
Cost breakdowns and subscription normalization goals
Reasoning effort settings and thread management UI
Home view features and Ultra Fast workflow
OpenAI vs Anthropic momentum and Ship Week
Semi-analysis data and Magic Mouse design debate
Cost per task efficiency and model benchmarks
Long-context capabilities and compaction quality
OpenAI Dots: autonomous agents and communication
T3 Code open-source philosophy and Google integration
Product coherence and the 'everything app' problem
Car theft challenge and skater gear anecdotes
Listen ad-free on Castria