Codex vs Claude, Seedance Ads, Local Qwen on 5090 — AI Daily May 12

559 messages · 81 active members

559
messages
81
active members
@jasonakatiff, @arielletolome, @mb29266
top contributors

Overview

Today's chat opened with a hot debate on coding harnesses: @samb69 argued Claude is fine for creative but bad for coding, running OpenCode with custom code+review agents that loop until bugs are fixed — Opus 4.7 for planning, GPT-5.5 for code/audits. Some reported Opus going 'dumb' after 11-day sessions with 17 compactions, prompting talk of weekly resets and memory tools like Honcho, Hindsight, and @jcartu's new Cerebellum repo. Meanwhile @jcartu archived his memory project (beaten by Mem0) and showed Qwen 27B running well on a single 5090, with TP4 setups targeting 200tps on 397B models — enabling local plan/review workflows at 4-6x faster than Opus-only setups. Creative automation dominated the middle of the day. @arielletolome's Seedance-generated ads are outperforming with $200/day budgets, sparking discussion on @jasonakatiff's 20% scaling rule with 24-hour bake time. Builders compared gpt-image-2 vs nano banana (gpt-2 winning for copy-heavy statics and UGC), and @instanetworks shared a workflow of transcribing top ad videos into reusable Claude skills. @jasonakatiff launched thecreativebriefer.com, built in CMUX with self-rating output. On GTM, @iannagy kicked off an agentic B2B SDR thread — Instantly emerged as the 'Shopify of cold email,' with Kieran strongly recommending Claude Code over Clay for enrichment. @jasonakatiff also walked through the chaos of his 60+ decision-point lead router (credit vs prepaid, CAPI, dispositions) with the candid advice: don't build one. @navuud reported back from SaaStr that this group's alpha is well ahead of the broader SaaS scene.

Topics

@samb69 argued Claude is bad for coding, running OpenCode with looping code+review agents (Opus 4.7 plan, GPT-5.5 code/audit). @sibunting reported Opus going dumb after an 11-day session with 17 compactions — builders split between 4.6, 4.7, and GPT-5.5 with weekly resets recommended.

@arielletolome's Seedance ads outperformed at $200/day; @jasonakatiff laid out the 20% scaling rule with 24-hour bake and launched thecreativebriefer.com. @instanetworks shared transcribing winning videos into reusable Claude skills, while gpt-image-2 emerged as the new default for copy-heavy statics over nano banana.

@iannagy asked about working agentic SDR setups; Instantly dominates sending, and Kieran pushed Claude Code over Clay for enrichment (cheaper, better). @jasonakatiff confessed his 60+ decision-point lead router (credit, CAPI, dispositions) is brutal — easier ways to make money exist.

@jcartu spent 4 days optimizing a vLLM fork (MTP-3 to D-Flash to D-Tree) on Blackwell, with 27B running well on a single 5090 and TP4 targeting 200tps on 397B. Claims Qwen 27B + Opus planning delivers 4-6x faster coding than human-readable speed for just GPU + electricity cost.

@sibunting detailed an orchestrator managing 4 Opus agents and 5 Hermes workers via Telegram/Slack with Jira workflow. Memory tools discussed include Honcho, Hindsight, and @jcartu's Cerebellum (RASPUTIN) repo with causal memory. Key/API rotation and session compaction remain unsolved.

Key Takeaways

  • Use Opus/Claude for planning and creative, Codex/GPT-5.5 for coding and audits — OpenCode with looping code+review agents fixes bugs autonomously, and reset weekly to avoid Opus degradation after many compactions.
  • Scale winning ads with the 20% rule + 24-hour bake, but require real statistical significance before going big — don't mistake luck for signal.
  • Replace Clay with Claude Code for B2B enrichment, pair with Instantly/Smartlead for sending, and avoid Microsoft inboxes which are now nearly impossible to land in.
  • A single RTX 5090 can run Qwen 27B locally for plan/review coding; TP4 Blackwell setups target 200tps on 397B models — local Opus-style workflows are now viable at hardware-only cost.
  • Transcribing top-performing ad videos into reusable Claude skill libraries is becoming standard practice — combine sub-skills per creative type with hard DON'Ts for brand/legal compliance.

Hot Threads

@jasonakatiffstarted

Lead router complexity confession + creative brief generator launch at thecreativebriefer.com

22 replies7 participants
@sibuntingstarted

Opus 4.7 going dumb after 11-day session with 17 compactions — orchestrator and memory architecture

18 replies6 participants
@iannagystarted

Agentic B2B SDR stack: Instantly, Claude Code over Clay, inbox sourcing problems

18 replies7 participants

Linked Items