Autonomous Ad Agents, DeepSeek Flash, Anthropic Git — AI Daily Oct 15

578 messages · 68 active members

578
messages
68
active members
@anonymous, @rmktg, @greglavidaloca
top contributors

Overview

October 15 was dominated by @rmktg's walkthrough of a near-autonomous ad agent running at 80-85% autopilot: image and copy generation end-to-end, batch-QA'd video clips auto-assembled, ROAS as sole metric, $300-800 CBOs with tiered pause thresholds, and dead ads above 1.0 ROAS revived into fresh campaigns to counter attribution lag. Grok 4.7 remains the primary copy model with Claude in testing. In parallel, a long thread on ad comment seeding reached rough consensus: minimal lift for affiliate, still valuable for ecom trust, and a clear gap for an API-driven provider — @seekersight surfaced comdrop.io and @anonymous is building on top of a scraper + profile warmup stack. On models and orchestration, @miyamotaz reported DeepSeek Flash via OpenRouter (sorted by throughput) running a 16k-word multi-agent research report in 15 minutes for ~$0.15. @jcartu is using Astra as a fast Hermes orchestrator across rotating accounts, @Anonymoushat moved back to Claude for long-horizon intent, and @samb69 confirmed DeepSeek Flash holds up on long creative turns. OMP emerged as a key unlock, routing across Claude, Codex, and Grok plans with auto-switching when accounts anonymous out — essential for overnight runs. Infra pain hit when @jrizzolo's Orca killed worker sessions on a 32GB VPS; consensus is 96-128GB minimum for serious agent fleets. Anthropic's Git rebuild sparked debate — @thewildzeno pushed back on the 300 merges/day flex while @Uziarbiv read it as a signal of bigger infra ambitions. Jev shipped multimodal (@justingacina clocked 98% vs Claude's 99% on 150 images for 3 cents), @Anonymoushat hyped Subquadratic as '10x Jev' with 384K context, and Meta dropped a new open-weight model @samb69 pegs at Sol 6.1 level. Business threads covered Bending Spoons' Airtable acquisition as the Elon-at-Twitter playbook, plus new builder intros from Ronald (Shopify attribution dashboard, $500k/mo spend) and Soheil (trade journaling site).

Topics

@rmktg detailed a production ad system generating images/copy autonomously and assembling video from pre-QA'd clip banks. ROAS is the only metric, CBOs run $300-800 with tiered pause thresholds, and killed ads above 1.0 ROAS get rolled into fresh campaigns to recover from attribution lag. Grok 4.7 is primary for copy, Claude in testing.

@miyamotaz reported DeepSeek Flash via OpenRouter handling a 16k-word multi-agent research report in 15 min for ~$0.15. @jcartu is running Astra as an ultra-fast Hermes orchestrator across rotating seats, while @Anonymoushat moved back to Claude for long-horizon intent. @samb69 confirmed Flash holds up on long multi-turn creative tasks.

Anthropic published their rebuild of Git to handle Claude's merge volume. @thewildzeno pushed back on the 300 merges/day flex — asking where the code actually ships — and proposed 'meaningful fix/feature per day' as a better metric. @Uziarbiv and @ConnectCallPartners read it as a signal of much bigger infra ambitions.

@jrizzolo hit Orca killing worker sessions overnight on a 32GB Linux VPS. Consensus: 32GB is bare minimum, 96-128GB recommended for heavy agent loads. OMP surfaced as a must-have unlock, routing across Claude, Codex, and Grok plans concurrently with auto-switching when accounts anonymous out — critical for 24/7 agent runs.

Jev shipped image input — @justingacina benchmarked it at 98% vs Claude's 99% across 150 swipe-file images for 3 cents total. @Anonymoushat hyped Subquadratic as '10x Jev' with 384K context, ~1s decisions over 100K tokens, and $0.05/M input tokens. Meta also dropped a new open-weight model @samb69 claims is Sol 6.1 level.

Words worth knowing

Three terms from the glossary.

Key Takeaways

  • ROAS-only optimization with aggressive early kills and dead-ad revival beats over-engineered attribution — let CBOs sort winners.
  • DeepSeek Flash on OpenRouter (sort by throughput) is a serious cost/speed unlock — 16k-word multi-agent runs for ~$0.15.
  • Running many agents in Orca needs 96-128GB RAM minimum; 32GB will silently kill worker sessions overnight.
  • OMP enables parallel Claude + Codex + Grok plan use with auto-switching — essential for overnight agent runs.
  • Jev's multimodal is nearly Claude-grade for image classification at a fraction of the cost (~$0.02/100 images).

Hot Threads

@rmktgstarted

Autonomous ad agent architecture, CBO thresholds, and ROAS-driven workflows

30 replies8 participants
@miyamotazstarted

DeepSeek Flash speed and cost for multi-agent flows

11 replies4 participants
@jrizzolostarted

Orca killing worker sessions — RAM requirements for agent fleets

11 replies4 participants

Linked Items