Seedance 2.5 UGC, OMP + DeepSeek V4 Flash, BurnDeck — AI Daily Aug 9

455 messages · 65 active members

455
messages
65
active members
@jcartu, @jarvisballer, @arielletolome
top contributors

Overview

Today's chat was dominated by AI video economics and orchestration cleanup. Seedance 2.5 landed with strong reactions — @bartekadamczyk demoed 1:1 reference-video copies with consistent voice and shockingly accurate audio physics (muffling when an arm crosses a lav mic). Builders debated budget allocation across human UGC (~$100), outsourced AI (~$50), Seedance 2.5 (~$10), and Veo 3.1 (~$1), with the consensus that creative volume and speed — testing 20+ angles in an afternoon — now beats polish. Product consistency at 480p and AI voice pacing remain the main quality gaps. On the tooling side, @Tz1888 and @jcartu report OMP paired with DeepSeek V4 Flash as a cheap workhorse for long fix-loops, provided you strip out imported Claude Code and Codex subagents/skills that bloat tokens. @thewildzeno noted OMP's system prompt over-encourages subagent spawning, but skills stay harness-independent — the portable, high-leverage asset. @jcartu also open-sourced BurnDeck, a cross-platform API quota and burn-rate HUD that got immediate uptake. Multi-model wrappers (T3code, Orca) with Linear as source-of-truth for agent handoffs are becoming default. Side threads covered @jcartu waterblocking his 4x RTX Pro 6000 Blackwell rig (TP8 now ~$50k as GPU prices spike), ad data warehousing choices (Snowflake+Sigma, BigQuery, DuckDB/DuckLake, Tinybird, StarRocks), Greg Isenberg's agent automation ideas (Stripe refund → CustomerIO sequences, PostHog bounce enrichment), and Google reportedly pivoting from frontier models toward AI infra.

Topics

@bartekadamczyk demoed Seedance 2.5 producing 1:1 reference-video copies with physics-accurate audio (muffling when an arm crosses the mic). Budget debate compared human UGC ($100), outsourced AI ($50), Seedance ($10), and Veo 3.1 ($1) — creative volume and speed now beat polish for ad testing. Product consistency at 480p and AI voice pacing remain open problems; Higgsfield and Runway are running unlimited Seedance promos with caveats.

@Tz1888 and @jcartu run OMP with DeepSeek V4 Flash as a cheap workhorse for fix-loops, after stripping bloated Claude Code and Codex-era subagents and skills. @thewildzeno noted OMP's harness over-encourages subagent spawning even when direct edits would be faster, but skills remain portable across harnesses — switching harnesses won't make a model smarter without solid skills and context.

@jcartu released BurnDeck, a cross-platform (macOS/Linux/Windows) API quota, burn-rate, and cooldown monitoring dashboard with 6 themes and multi-provider support. Source is on GitHub. Community response was strongly positive with several builders installing immediately. Multi-model coding wrappers like T3code, Cmux, and Orca (with Linear as source-of-truth) are also becoming standard for juggling Claude, Codex, and Grok.

@jcartu is waterblocking his 4x RTX Pro 6000 Blackwell rig (OC +6000 on memory for ~10% more tps) with a local GLM5.2 abliterated stack for infinite subagents at ~120tps, and warned TP8 now costs ~$50k as GPU prices spike. @jarvisballer is spinning up a basement 2-rack setup with solar, betting local models dominate within 6 months as AI subsidies unwind.

@drcopybymatt asked about warehousing ad data from RedTrack, Meta, and Google. @william_hodges runs Snowflake + Sigma BI (HIPAA-friendly, not cheap), @mattehb uses BigQuery for API-friendly AI integration, @tounano suggested DuckLake/DuckDB, @mattkowald recommended Tinybird (ClickHouse-backed) for lower starting cost, and @Tz1888 stumps for StarRocks.

Key Takeaways

  • Per-video AI creative costs have collapsed to $0-$10 with Veo 3.1 and Seedance 2.5 — creative volume now beats production quality for ad testing
  • OMP + DeepSeek V4 Flash beats Claude Code/Codex on tokens once you strip default imported subagents and skills; skills remain the portable, high-leverage asset
  • Multi-model wrappers (T3code, Orca) with Linear as source-of-truth are becoming default workflow for juggling Claude, Codex, and Grok
  • GPU prices are spiking: TP8 with RTX Pro 6000 Blackwell now runs ~$50k — buy before AI subsidies unwind
  • Tinybird (ClickHouse) and DuckDB/DuckLake are lower-cost OLAP alternatives to Snowflake and BigQuery for ad-data warehousing

Hot Threads

@arielletolomestarted

AI UGC video budget allocation and Seedance 2.5 quality

38 replies8 participants
@jcartustarted

Multi-model stack, RTX Pro 6000 rigs, and BurnDeck release

45 replies10 participants
@drcopybymattstarted

True data warehouse for ad data across RedTrack, Meta, Google

10 replies5 participants

Linked Items