Smart Orchestrator Stacks, Jev Reality Check, Gemini 3.8 Flash — AI Daily Sep 20
402 messages · 59 active members
Overview
Topics
@jcartu, @tounano, and @thewildzeno converged on using a fast/cheap model (Gemini 3.8 Flash, GLM Flash) as the main orchestrator while reserving Fable/Astra for ADR planning and critical gates. The key discipline: tell the planner exactly which dumb model will execute so it builds rubrics, gates, and must-pass checkpoints into the plan. @seekersight and Kieran shared full role configs mapping models to orchestrate/plan/build/review/scout.
Multiple builders tested Jev on funnel analysis, transcripts, and campaign interpretation. Consensus: it's a classifier that needs the official skill installed and heavy pre-mapping, and often loses to tuned local models like Llama or GLM 5.3 Flash. Real wins are narrow — transcript risk scoring, E2E testing, and B2B lead scoring — not general workflow intelligence.
@samb69 clocked Gemini 3.8 Flash at 350 tok/s and @jcartu called its vision tower 'sick,' making it the current favorite for main orchestration. Detractors like @seekersight found it wasn'anonymous passing full context to Fable during planning handoffs. Gemini 4 is reportedly being tested for release this week.
@scalingfrog wired Grok to ad accounts to kill underperformers and iterate winners, reporting first sales and sparking a CBO-vs-micromanage debate. @jcartu shared the Grok API playbook: $1k spend + 30 days to reach Tier 3, where it becomes usable and excels at browser automation. Anthropic paused new $200 accounts but old ones can reactivate, with a banked reset expected Tuesday.
@watchmedropship hit ~300 tok/s on Qwen 3.8 Next Flash via tensor parallelism, with @samb69 explaining the tradeoff vs. pipeline parallel. Qwen3.8-LiveTranslate (2.3s lag, 60 languages) and PrismML's Ternary Bonsai 2 27B (5.9GB, 98.2% of Qwen3.8 27B) dropped. Separately, @jasonakatiff warned that overloaded terms like 'workflow' silently degrade agent output — treat naming as first-class architecture.
Words worth knowing
Three terms from the glossary.
- HermesAn AI agent framework members use to orchestrate tasks, delegate to models, and automate workflows.
- Browser AutomationUsing AI or scripts to control a web browser programmatically for scraping, testing, or workflows.
- FB AdsFacebook advertising campaigns, a core channel for members doing paid acquisition.
Key Takeaways
- Use Gemini 3.8 Flash or GLM Flash as your main orchestrator and reserve Fable/Astra for ADR planning and critical gates — speed and cost improve dramatically without sacrificing plan quality.
- When planning with a frontier model, explicitly tell it which dumb model will execute the work so it builds rubrics, gates, and must-pass checkpoints into the plan.
- Jev is a classifier, not an intelligence layer — install the official skill, pre-map examples, and scope to transcript risk, E2E testing, or lead scoring.
- Tensor parallelism unlocks ~300 tok/s on Qwen 3.8 Next Flash, and PrismML's Ternary Bonsai 2 27B retains 98.2% of Qwen3.8 27B performance at just 5.9GB.
- Define a strict platform vocabulary early; overloaded words like 'workflow' will silently degrade agent output as your codebase grows.
Hot Threads
Smart orchestrator with dumb workers — why Fable/Astra for orch is backwards
Testing Jev on overnight data analysis for lead reverse-engineering
Grok bot managing ad accounts — kill losers, iterate winners