Codex vs Claude, Veo3 vs Seedance, PostHog Hogpill — AI Daily May 08

704 messages · 72 active members

704
messages
72
active members
@jasonakatiff, @arielletolome, @jcartu
top contributors

Overview

A loaded day in the builders channel, with three parallel wars dominating chat: harness selection, video model selection, and analytics tooling. On the coding front, @geilt, @jakubjbln, and others declared Claude 'lobotomized' and migrated to Codex (GPT-5.5 xhigh), with one builder reporting Codex solved in 4 minutes what Claude Code couldn't fix in 2 hours. The Hermes vs OpenClaw debate raged in parallel — @jcartu clocked Hermes at 36 minutes / 29 tool calls for a router config OpenClaw nailed in 30 seconds, while @trevlolwut and @Wootbro reported the opposite. @watchmedropship shipped a TF-IDF skill selector cutting Hermes prompt size 87.5%, addressing the skill-bloat problem several flagged. On the creative side, @arielletolome, @c_1media, @expadz, and @navuud ran a deep comparison of Veo3 (Google Flow Ultra, 50-100 videos/day unlimited), Seedance (40-min queues but better multi-character handling), Runway Explore, and Kling 3. Consensus: audio matters more than the model, and Deepgram beats Whisper for word-timestamp precision in automated edits. @navuud demoed an end-to-end pipeline (Format Finder → GPT Image 2 → Kie/Seedance → yt-dlp b-roll), and @arielletolome leveraged Opus Clip's enterprise API for captions and creative scoring. Analytics saw a collective 'hogpill' moment as the community migrated from Microsoft Clarity to PostHog, with @iamgalba driving everything via API and letting Claude Code auto-promote winning variants. @jasonakatiff shipped an anti-CRM mobile app for insurance brokers and pitched B2B lead sales as e-commerce, while Nous Research dropped Hermes Agent v0.13.0 and Telegram shipped a major bot-revolution update.

Topics

@geilt and @jakubjbln declared Claude 'lobotomized' and fully migrated to Codex (GPT-5.5 xhigh), citing dramatically better instruction-following on complex framework code. @GuruTime pushed back that LLM quality reflects codebase quality. In parallel, @jcartu clocked Hermes at 36 min / 29 tools for a Mikrotik config OpenClaw solved in 30 seconds, while @Wootbro and @trevlolwut reported Hermes 'forgets less.' @watchmedropship's TF-IDF skill selector prunes 351 skills to top 25 (87.5% prompt reduction) addressing widespread skill-bloat complaints.

@arielletolome ran 665 videos MTD on Veo3 via Google Flow Ultra (unlimited, 50-100/day), using Seedance as a backup despite its 40-min queues and stronger multi-character handling. @c_1media recommended Runway Explore mode (480p/720p often beats 1080p for UGC authenticity). @expadz built a Claude Code + ffmpeg + Whisper + Kie pipeline assembling 60-second videos in 8-10 minutes; Kieran recommended Deepgram over Whisper for tighter word timestamps. Opus Clip's enterprise API handles captions, silence removal, and creative scoring programmatically.

Members declared themselves 'hogpilled' as PostHog became the consensus funnel analytics tool. @iamgalba runs everything via API with Claude Code checking statistical significance and auto-promoting winners, while @samb69 built a deep tracking schema across all funnels. Free under 1M events/month and works for both ecom and lead gen — a clear migration path from Microsoft Clarity.

@bartekadamczyk kicked off automating viral hook discovery from IG/TT; @kingofgrowth shared an Apify-based scraper noting 'concept-inspired' hooks beat copy-pasted ones. @navuud demoed a full pipeline: Format Finder for hooks/scripts → GPT Image 2 storyboards → Kie + Seedance 2.0 video → yt-dlp + ffmpeg b-roll scraping from YouTube. @iannagy suggested feeding scraped content to Vertex API for visual+transcript extraction at 1-second intervals.

@jasonakatiff shipped a thin native mobile app extending theLeadRouter for direct brokers who hate CRMs and live in Google Sheets, pitching B2B lead sales as an e-commerce flow with tenant Stripe accounts and affiliate-driven lead generation. Discussion with @chester9383 covered automating B2B lead buyer management via CLI-controlled Meta campaigns, mass tort, telehealth (iSCALE.com), and the shift to NAD+/peptides amid GLP-1 CAC inflation.

Key Takeaways

  • Multiple builders are fully migrating off Claude to Codex (GPT-5.5 xhigh) — one reported Codex solved in 4 minutes what Claude Code couldn't fix in 2 hours.
  • Audio quality matters more than which video model you use; Veo3 on Google Flow Ultra unlocks 50-100 videos/day unlimited, and Deepgram beats Whisper for word-timestamp precision.
  • PostHog is now the community-default funnel analytics tool — drive it via API and let Claude Code auto-promote winning variants based on statistical significance.
  • Audit your Hermes skills — most builders use <10% of installed skills; @watchmedropship's TF-IDF selector cuts prompts 87.5% (34K → 4.3K chars) with 100% recall.
  • Hindsight memory plugin works with any agent harness (OpenCode, Hermes, OpenClaw) — it's the persistent-memory unlock, not Hermes-exclusive.

Hot Threads

@arielletolomestarted

Veo3 vs Seedance vs Runway — character consistency, storyboards, and high-volume production

44 replies8 participants
@geiltstarted

Claude is lobotomized — Codex GPT-5.5 xhigh is night-and-day better for complex coding

38 replies6 participants
@danfeldmanstarted

What's everyone using to monitor and improve funnels? (PostHog vs Clarity)

22 replies10 participants

Linked Items