Native Ad Pipelines, PR Review Stacks, Fish Audio TTS — AI Daily Aug 24

550 messages · 68 active members

550
messages
68
active members
@jcartu, @tounano, @jasonakatiff
top contributors

Overview

Sunday's discussion was anchored by @tounano's detailed walkthrough of a near-autonomous creative pipeline for native ads: a 100-ad decomposed library feeds headline generation, image prompting, and deterministic linting, with Gemini Flash winning on image prompting, Grok on headlines, and orchestration handled by plain API scripts rather than a coding harness. @iannagy shared a parallel animation pipeline (script → storyboard → VO → clip via Kie → assembly) and a Gemini-based VSL analysis flow for extracting mechanisms suitable for animation. On the coding side, builders compared PR review stacks — Greptile, Cubic, CodeRabbit, and Aether — with @thewildzeno noting per-review pricing burns fast and DIY multi-model review via max subs is increasingly attractive. @jasonakatiff and @seekersight favor Claude calling Codex and Grok directly for adversarial review, and Jason floated layered ad compliance QA (state bar, FTC, platform-specific) as a promising SaaS idea. Model chatter dominated: Opus 5 has visibly regressed, Grok 4.5 is winning on copy, and @tounano flagged a broader creative-writing regression across modern models due to coding-heavy RLVR training. Voice gen got a strong signal boost with Fish Audio unanimously recommended over ElevenLabs — cheaper, better voices, lenient policies. @leewardbound laid out a thesis on the token price collapse: real agentic apps will shift 90%+ of traffic to lightweight models like Luna/Haiku for bounded tasks. Meanwhile @jcartu's OMP orchestrator and Omarchy Linux distro drew converts, and @samtome demoed Qwen running locally on a Pixel phone.

Topics

@tounano detailed a multi-step pipeline: build a 100-ad library decomposed along axes, generate 20 diverse headlines from a lander + 50 random ads, then produce 20 image prompts with deterministic linting and agent review. Gemini Flash wins on image prompting (~$0.14/step), Grok on headlines, and orchestration runs through API scripts rather than a coding harness because the flow is deterministic. @iannagy shared a parallel animation pipeline using Kie for clip generation.

Builders compared Greptile, Cubic, CodeRabbit, and Aether for automated code review. Consensus: per-review pricing burns fast, CodeRabbit offers best hygiene value, and having Claude directly call Codex/Grok for adversarial review is often cheaper and better. @jcartu's OMP orchestrator sparked deep discussion, with @jonmacofficial generating custom Grok-authored skills (Bug Fix, QA, New Feature) from the docs.

Multiple builders report Opus 5 quality has dropped noticeably, while Grok 4.5 is winning on copy and fast on API. @tounano observed modern models have regressed on creative writing because RLVR training is heavily coding/security-focused — GPT-4 and early Claude 4 produced better prose. @basant framed Fable as orchestrator with Opus 5 as grunt worker. Ox Alpha (rumored GLM 5.3 Flash) got mixed reviews.

Builders unanimously recommended Fish Audio over ElevenLabs for TTS — cheaper, better voices, and no bans on financial lead gen ads. @drcopybymatt has produced thousands of videos with it. VoxCPM mentioned as best local option, but consensus is to skip local TTS entirely.

@leewardbound argued current AI usage is heavily skewed toward top-tier models for dev workflows, but real agentic apps use bounded-complexity prompts (spam detection, lead rating) that lightweight models handle at 1/20th the cost. Expect 90%+ of production traffic to move to Luna/Haiku-tier models, driving explosive efficiency gains. @samtome demoed Qwen locally on Pixel with theoretical multi-phone clustering.

Key Takeaways

  • Deterministic creative pipelines run cheaper and faster as plain API scripts than inside a coding harness — reserve harnesses for cases needing orchestration judgment.
  • Gemini Flash currently produces the best image-model prompts; Opus 5 has visibly regressed and Grok 4.5 is winning on copy.
  • For PR review, per-call pricing on Greptile/Cubic favors switching to Claude → Codex/Grok adversarial reviews on max subs once volume scales.
  • Fish Audio is the community pick over ElevenLabs: cheaper, better voices, lenient policies for financial lead gen — just make the switch.
  • Future AI token demand will be dominated by lightweight models (Luna/Haiku-tier) handling bounded tasks at scale, not heavy dev-loop workflows.

Hot Threads

@tounanostarted

Autonomous native ad creative pipeline walkthrough

30 replies5 participants
@jonmacofficialstarted

How to drive OMP effectively for bugs and new features

12 replies3 participants
@nowwatchthisdrivestarted

Better alternatives to ElevenLabs for voice generation

9 replies6 participants

Linked Items