AI UGC Factory, Docker Hijack, Codex vs Claude — AI Daily May 18

584 messages · 83 active members

584
messages
83
active members
@launchmanmathias, @jasonakatiff, @Wootbro
top contributors

Overview

Today's standout was @launchmanmathias breaking down a 2,000-AI-videos-per-day ad operation: Kling for generation, ~$4k/day compute, a 24-person team, and a static-image → AI video → real actor testing funnel. Counter-data from @startropics and @fewga893 showed real UGC still beats AI by 20-30% on identical scripts, but AI crushes 4-second hooks and b-roll iterations. The wildest win: making ads obviously AI (everyone as a giraffe) printed for 5 months. He also revealed his full Claude pipeline (hooks/CTAs → VSL scripts → AI video prompts) and @Coybh shared a still-working trick to bypass Facebook's 250-ad-per-page cap by scheduling future ads. The day's most important PSA came from @sibunting, who caught a Russian XMR miner exploiting an agent-deployed pgvector container with default settings (POSTGRES_HOST_AUTH_METHOD: trust + 0.0.0.0 binding). Full IOCs, audit commands, and the critical warning that Docker bypasses ufw by default were shared. On tooling, builders kept comparing Claude Code vs Codex — many report Claude has felt 'spiritually lazy' while Codex (GPT-5.5 xhigh) is becoming the implementer of choice, often run in tandem with one auditing the other. Hermes continues its rise as orchestrator, with Nous Research's new Codex App Server runtime enabling proper integration; compaction sweet spot is ~150k tokens. Elsewhere: Musk lost his OpenAI lawsuit and is reportedly handing Anthropic access to Colossus while building Colossus 2 for SpaceX/xAI. Seedance 2 (via Dreamina direct, cheaper than Runway) emerged as the video gen favorite over Kling 3 and Veo 3 — keep clips under 8s to avoid lip sync issues. @geilt closed the day with a philosophical take on the AI hype cooling: rising costs, lobotomized models, and the looming AI-to-AI economy question.

Topics

@launchmanmathias detailed producing 2k AI ads/day for home/auto/medicare/loans using Kling (~$4k/day), a 24-person team, and a static-image → AI video → real actor testing funnel. Counter-data: real UGC beats AI by 20-30% on identical scripts, but AI excels at hooks and b-roll. He also revealed his full Claude pipeline turning hooks/CTAs into VSL scripts and AI video prompts, while @Coybh shared an FB trick to bypass the 250-ad page limit by scheduling future ads.

@sibunting's dev box was compromised when an agent spun up pgvector with default settings (port bound to 0.0.0.0 + POSTGRES_HOST_AUTH_METHOD: trust). Attacker used COPY FROM PROGRAM to drop a Monero miner for ~10 hours. Full IOCs, audit grep commands, and the warning that Docker bypasses ufw by default — use 127.0.0.1: prefix and DOCKER-USER iptables chain.

Multiple builders reported Claude Code has been degrading — @rmaigue feels like 'GPT-4', @Max_Bernstein saw crashes. @GuruTime is now running Codex (GPT-5.5 xhigh) as implementer with Claude as auditor, saying 'at this level they're basically the same.' @Wootbro also walked @danfeldman through Claude's /remote-control flow for running sessions from iPhone.

@filiuser confirmed Hermes can run GPT-5.5-codex via OAuth using Nous Research's new Codex App Server runtime. Builders debated compaction thresholds (150k tokens emerging as default) and skill bloat audits. @jcartu open-sourced 'rasputin-mantle,' a self-hostable Manus-style agent with sandboxed Chromium and CodeAct loop. Security-conscious builders lean toward isolated VPS + scoped tokens.

Seedance 2 via Bytedance's Dreamina (cheaper than Runway's wrapper) is winning for ad creative; keep clips under 8s to avoid lip sync issues. Musk lost his OpenAI lawsuit (his own texts pushing for-profit conversion hurt the case) and is reportedly giving Anthropic Colossus access while building Colossus 2. @geilt argued AI hype is cooling — rising API costs, lobotomized models, no true autonomy — though others compared it to the dot-com bust: overhyped short-term, transformative long-term.

Key Takeaways

  • AI UGC wins as a cheap testing layer (static image → AI video → real actor), but humans still beat AI by 20-30% on identical scripts — except for 4s hooks + b-roll where AI prints. Counterintuitive: making ads obviously AI (giraffe characters) outperformed photoreal for 5+ months.
  • If you let agents run with --skip-permissions, audit docker-compose NOW: grep for 'POSTGRES_HOST_AUTH_METHOD.*trust' and port bindings without '127.0.0.1:' prefix. Docker bypasses ufw by default — use the DOCKER-USER iptables chain.
  • Codex (GPT-5.5 xhigh) is overtaking Claude Code as the implementer of choice; many builders now run them in tandem with one auditing the other rather than picking a side.
  • Hermes + Nous Research's Codex App Server runtime is the emerging autonomous agent stack — compaction sweet spot ~150k tokens, run on isolated VPS with scoped tokens, not your workhorse droplet.
  • Seedance 2 via Dreamina (Bytedance direct) is cheaper than Runway's wrapper for the same model; keep clips under 8s and skip on-screen text to dodge lip sync issues. FB's 250-ad page cap still bypassable via scheduled future ads.

Hot Threads

@launchmanmathiasstarted

2k AI videos/day ad factory + full Claude pipeline + FB 250-ad page limit bypass

83 replies12 participants
@robinroystarted

Elon Musk loses OpenAI lawsuit — debate on Altman, SF juries, and Colossus going to Anthropic

35 replies10 participants
@sibuntingstarted

Russian XMR miner exploited agent-deployed Postgres container — full IOCs + audit guide

3 replies3 participants

Linked Items