Jev Routing, Meta Verification Crackdown, Codex SSH — AI Daily Sep 18
656 messages · 69 active members
Overview
Topics
@jasonakatiff found headless browser UI testing as a killer Jev use case, clocking 0.4s vs 8s for base LLM calls, with plans to layer probabilistic testing on 27k existing fixtures. @Anonymoushat proposed Jev for graph-maintainer routing (probability distribution over sub-team buckets) and git-diff blast-radius scoring. Consensus: Jev shines as a post-planner dispatcher and confidence-scoring classifier for enumerable options, but flops when forced between humans and other models — Astra/Fable still own reasoning-heavy orchestration.
Meta is enforcing hard ID verification across business managers and entire infra setups, with agencies requesting LLC info. Finance and aggressive-angle advertisers are running 75-80% reject rates as baseline — @rockdm submits 200+ ads expecting only 20-25% acceptance. System-user API access stays ban-safe at 3000+ calls/day while MCP triggers bans. Multiple builders are replicating Whop's cleaner UI or wiring Hermes/Slack agents to bypass Ads Manager lag.
Grok Bot's separate usage meter burns 18-20% per day of direct chat; @rmktg and others are routing it through Hermes for background work (image gen, campaign changes) instead. Opus 5 is strong at coding/QA but a nightmare orchestrating other agents — reportedly worse than 4.6 in that role, pushing members back to Fable-as-orchestrator with Opus as subagent. Claude is now blocking plus-addressed aliases; Simplelogin's 10 free aliases work as a bypass, and new accounts increasingly require ID.
Members reported strong impressions of Codex autonomously learning codebases when given SSH access, then recommending and implementing solutions. This sparked a broader take framing the current moment as a shift from instruction-based to intent-based computing — describing outcomes rather than steps, with builders acting as directors of specialized agent crews. Humans remain essential mainly for undocumented tacit knowledge.
For freshly launched apps with 0 followers, run engagement-objective campaigns targeting low-CPM geos (Bangladesh, Colombia, Egypt, India, Indonesia, Nigeria, Peru, PH, PK, VN) at ~$20-30/day for a few weeks before scaling to CPA. Cute/funny/religious static creatives with app name embedded work best. Separately, builders debated adding PayPal to Stripe checkouts — PayPal reportedly sides better with merchants on chargebacks; Paddle and Checkout.com surfaced as alternatives worth a payment orchestration layer.
Words worth knowing
Three terms from the glossary.
- Max PlanThe premium subscription tier for Claude Code that costs ~$200/month with higher usage limits.
- Media BuyingRunning paid advertising campaigns, primarily on Facebook and Google, to acquire customers or leads.
- Context WindowThe amount of text and conversation history a model can remember and work with at one time.
Key Takeaways
- Jev's sweet spot is enumerable-option decisions with probability outputs — headless QA at 0.4s/call and post-planner routing are the clearest wins; don'anonymous force it into top-level orchestration.
- Meta system-user API access stays safe at 3000+ calls/day; MCP and browser automation trigger bans, and full-infra ID verification is now escalating — have LLC and personal ID ready.
- Grok Bot burns 18-20% of its weekly meter per day of direct chat; route it through Hermes for background work to preserve capacity.
- Warm up new FB/IG pages with $20-30/day engagement campaigns in low-CPM countries before running CPA, and expect 75-80% ad reject rates in finance verticals.
- Codex with SSH access can meaningfully self-learn codebases and propose fixes — reinforcing the shift from writing instructions to directing agent crews.
Hot Threads
Jev orchestration: real router/classifier use cases vs forced fit
Managing 200+ ad uploads with 75% reject rate and account hygiene
GTM warmup for newly launched apps with zero followers