Codex vs Opus, Seedance Video Ads, Hermes Agents — AI Daily Apr 26

613 messages · 74 active members

613
messages
74
active members
@jasonakatiff, @samb69, @Wootbro
top contributors

Overview

The dominant thread today was a clear shift in coding driver: builders are migrating from Claude Code to Codex (GPT-5.5) after Anthropic's nerfs, OpenClaw block, and tightened limits. The emerging meta-workflow uses Opus 4.7 as orchestrator/PRD-writer and Codex as the fast one-shot executor for backend and bug-finding, often handed off via GitHub issues. AGENTS.md ports cleanly from CLAUDE.md, making the switch low-friction. Power users like @jcartu burn 250€+/night on Opus mem0 tuning while @robinroy runs 9 Codex subs without maxing out. Video ad automation hit a new gear: @samb69 finalized ~40 ads in hours via a Seedance 2.0 pipeline with skill-scaffolded scripts, and @rockdm two-shotted a full ad with video_use + hyperframe + seedance. On the systems side, @namtalks detailed a Hermes-based ecom OS where Opus acts as 'CEO' delegating to Minimax/Kimi sub-agents for SEO, blogs, support, and content review — making the case that open-source harnesses are essential to avoid LLM vendor lock-in. Infra and tooling rounded out the day: long-form transcription stacks (Groq v3 Turbo at ~1/5 Deepgram cost, VibeVoice-ASR, Whisper) with diarization still the unsolved bottleneck; programmatic cart launches via CheckoutChamp/Sticky APIs; @samtome's argument that cloud GPU rental beats ownership at ~15% margins over 5 years; and notable drops including HeyGen's open-sourced Claude-built launch video code, Microsoft VibeVoice, and Google Jules.

Topics

Builders are migrating off Anthropic en masse, citing nerfs, compute limits, and the OpenClaw block. Consensus workflow: Opus 4.7 writes detailed PRDs and orchestrates, Codex executes one-shot via GitHub issues. Codex is winning on backend, bug-finding, and token efficiency; Opus still leads UI/FE work, and AGENTS.md mirrors CLAUDE.md for clean config porting.

Seedance 2.0 has collapsed video ad production to ~5% manual QA, with @samb69 shipping ~40 finalized ads in hours using skill-scaffolded scripts driving characters and voices. @rockdm two-shotted a full ad combining video_use, hyperframe, and seedance, and @jasonakatiff flagged browser-use/video-use as a promising agentic editor approach.

@namtalks broke down their DTC stack using Hermes with Opus 4.7 as 'CEO' delegating to cheaper models (Minimax, Kimi) for SEO, blog writing, landing pages, support, and content review. The bigger thesis: open-source harnesses like Hermes and opencode are now mandatory to hedge against LLM provider rug-pulls.

Builders compared Groq v3 Turbo (~1/5 Deepgram cost), VibeVoice-ASR, and Whisper for long-form audio, with diarization still an unsolved cost/accuracy problem. @kennyaronson sparked a debate on programmatic cart launches — consensus: build on CheckoutChamp or Sticky APIs, don't vibe-code financial flows. @utilitiesmind shared an anti-AI-writing prompt killing parallel repetition stacks and false contrast triplets.

@samtome argued cloud GPU rental beats ownership at ~15% margins on 75% utilization over 5 years, plus access to newer cards. He also predicted Google will simply acquire the winning AI lab in ~2 years. Notable drops: HeyGen open-sourced their Claude-built launch video code, Microsoft released VibeVoice, Google teased Jules, plus a WiFi-to-pose project and an AI PCB design tool.

Key Takeaways

  • Hybrid stack winning: Opus 4.7 writes PRDs and orchestrates, Codex 5.5 executes via GitHub issues — best of both with token savings
  • Seedance 2.0 has collapsed video ad production to ~5% manual QA — builders shipping 40+ ads/day with skill-scaffolded pipelines
  • Open-source agent harnesses (Hermes, opencode) are becoming default to hedge against LLM provider rug-pulls and vendor lock-in
  • Groq v3 Turbo cuts transcription costs to ~1/5 of Deepgram with comparable quality, but diarization remains expensive and inaccurate
  • For programmatic checkout flows, build on CheckoutChamp or Sticky APIs — never vibe-code financial logic, and always use staging + commit-by-commit rollback

Hot Threads

@KtargetMediastarted

Codex 5.5 vs Opus 4.6/4.7 — which is actually better right now?

40 replies10 participants
@samb69started

Seedance video ad pipeline: 40 ads finalized in hours with skill scaffolds

25 replies6 participants
@namtalksstarted

Hermes-based ecom OS with Opus as CEO delegating to Minimax/Kimi

12 replies4 participants

Linked Items