Meta Ads CLI, Seedance 2.0, Claude Outages — AI Daily Apr 29

819 messages · 96 active members

819
messages
96
active members
@Wootbro, @mb29266, @jcartu
top contributors

Overview

Today's headline: Meta quietly shipped an official Ads CLI, opening a first-mover automation lane for agentic media buyers — though builders are bracing for the same rate limits, 250-ads/page caps, and ban-happy moderation that already plague the API. The release dovetailed with @mb29266's viral playbook for 200x'ing ad output: rip winning headlines, feed Claude 50 variations of the same psychological trigger, expect ~10 printers per batch. Consensus: the bottleneck is no longer creative, it's Meta's infrastructure. Don't buy pages (banned), buy profiles, and submit a public app to unlock real rate limits. Claude had a brutal day — outages, 47-minute runs on 4.7, and aggressive usage-policy refusals — pushing more builders toward Codex/GPT-5.5, where reset limits remain nearly impossible to hit. Opus still wins on humanized copy, but the gap is closing. On the creative stack, Seedance 2.0 was crowned video SOTA (Wan 2.2 and Higgsfield aren't close), while @jcartu pushed local iron: a Qwen3.6-27B abliterated NVFP4 multimodal model running on RTX 6000s, with NVFP4's cutlass kernel hardware compression making prosumer inference actually viable. Voice AI for sales drew a long thread, with Retell + n8n as the working stack for qualification and appointment setting — but @rmaigue's rule landed: the closer to money, the less AI involvement. Builders also quantified Claude Max ROI (replacing $50k/mo dev teams, powering multi-million ARR apps), debated 14" vs 16" MacBooks, and traded multi-agent orchestration setups (cmux + tmux + autossh + mosh, with Hermes orchestrated by a cheap frontier or free local model to dodge its self-upgrade slowness).

Topics

Meta dropped an official Ads CLI, letting devs manage campaigns from the terminal and slashing tokens needed to launch. Builders are plotting integrations but warn of inherited rate limits, the 250 ads/page cap, and ban risk on aggressive automation. Workarounds: don't buy pages (banned), buy profiles, submit a public app for real rate limits.

@mb29266's playbook: rip winning ads, feed Claude 50 variations of the same psychological trigger, expect ~10 printers per batch — a claimed 200x speed unlock. Tip: use cosine-similarity querying against curated copywriter corpora (Stefan Georgi's RMBC, classic ad legends) instead of relying on training data. Fundamentals still required to avoid testing slop.

Seedance 2.0 is the consensus winner for AI video ads, leaving Wan 2.2 and Higgsfield behind on quality. The catch: not runnable locally, so builders pay per generation. Happy Horse looks promising but weights aren't public yet.

Claude went down mid-day with 4.7 crawling for days (47-minute runs reported) and refusal storms interrupting local dev. Codex/GPT-5.5 limits reset and remain effectively unhittable, pulling defectors. Meanwhile @jcartu pushed local abliterated models (Qwen3.6-27B NVFP4 multimodal) on RTX 6000s — NVFP4's cutlass kernel hardware compression makes prosumer inference viable without naive 4-bit quality loss.

Retell + n8n is the working stack for AI voice qualification and appointment setting, even on 60k cold lists for life insurance — but closing and high-ticket calls remain human territory (@rmaigue: closer to money = less AI). For multi-agent rigs, cmux + tmux + autossh + mosh is the emerging remote setup. Hermes is slow because it self-upgrades skills into context — orchestrate with a cheap frontier or free local model instead.

Key Takeaways

  • Meta Ads CLI is live — first-mover automation opportunity, but expect rate limits and ban risk; buy profiles not pages, and submit a public app to unlock real API throughput.
  • Rip winning ads → feed Claude 50 variations of the same trigger → ~10 will print. Volume of iterations on a proven concept beats brainstorming new ones.
  • Codex/GPT-5.5 limits reset and remain nearly unhittable — a real alternative as Claude struggles with outages, slow 4.7 responses, and aggressive refusals. Opus still wins on humanized copy but the gap is closing.
  • NVFP4 with cutlass kernels is hardware compression, not naive 4-bit — the only realistic way to run serious abliterated models on prosumer GPUs without major quality loss.
  • Rule of thumb: the closer a workflow is to money (closing, high-ticket sales), the less AI should be involved. The farther from money, the more it should be automated.

Hot Threads

@mb29266started

200x ad creative output with Claude variations on ripped winners

30 replies8 participants
@cdxxxstarted

AI voice agents for lead qualification — Retell, Hume, Twilio compared

22 replies7 participants
@tounanostarted

14" vs 16" MacBook for AI dev work — and the local iron counter-argument

30 replies9 participants

Linked Items