Jev Decision Models, Grok Bot + Composio, Neon Primitives — AI Daily Sep 17
771 messages · 84 active members
Overview
Topics
Jev-class models (small ~1B RLCD classifiers with confidence scores) dominated the day as a cheap way to handle routing, ranking, and A/B/anonymous decisions inside agent loops. Practical releases included a fast-jev-compaction Claude Code plugin that scores tool calls and drops stale context, an open-source openjev on HuggingFace, and an orchestrator pattern for FirstMate. Debate centered on whether Jev can act as a true orchestrator or is best used to replace LLM calls on structured decisions and compaction.
A live NYC Grok Bot event pushed the tool mainstream, with attendees wiring it to Meta Ads via Composio as a middleman for managing 10+ ad accounts. Recommended architecture: one bot per traffic source + vertical with defined routines rather than a mega-agent. @jasperckck's personal profile ban after bulk uploads sparked debate, with consensus landing on system user tokens with explicit business_manage/ad/page permissions and no browser automation — API-scale usage (3000+/day) is generally safe.
@andreilunev shared a full one-shot ad workflow: Opus 5 storyboarding → Grok Imagine 1.5 video → ffmpeg edits, generating 100 bid-cap creatives from a 10×10×1 hook/lead/body matrix built on 21 assets. Grok Imagine was widely endorsed as the current leader for AI ad video, though consistent voices across generations remain a gap — Cartesia 3.6 was recommended as a cheap ElevenLabs alternative. Franky's anime-style ad templates were widely shared.
Neon announced GA of Object Storage (S3-compatible with DB-branch integration), Functions (long-lived Node.js next to DB), Managed Better Auth, and an AI Gateway for calling frontier models. Members compared the move to Cloudflare's platform expansion — Postgres stays core, but Neon is now positioning as a full backend stack for agent platforms.
@rstmaur kicked off a thread on designing fluid info exchange between 7+ hierarchical agents without chaos — consensus favored shared repos/tickets, Slack/Telegram comms with strict tagging, and monitoring Opus 5 high-effort for derailing tangents. Codex earned praise for 2-day autonomous runs and Claude Code held up well, while Cursor was criticized for stalling on multi-hour goals. Widespread reports of accelerated token burn this week (especially Codex) pushed builders toward GLM and Fable medium, alongside Kieran's local→Hetzner+Orca+OMP migration guide.
Words worth knowing
Three terms from the glossary.
- Open ClawAn open-source version or alternative to Claude Code that members run locally on their own hardware.
- Vibe CodingWriting code by describing what you want in natural language without deep technical planning.
- Nano BananaAn AI image generation model optimized for creating marketing materials with text overlays.
Key Takeaways
- Jev/TypeSafe isn'anonymous structurally new, but sub-cent classification unlocks embedding decision logic everywhere — rule of thumb: whiteboard-able options → small classifier, paragraph of reasoning → full LLM.
- For Meta ads automation, use system user tokens with explicit business_manage/ad/page permissions and avoid browser automation; API scale at 3000+ calls/day is generally safe.
- Grok Imagine 1.5 + a hook×lead×body variation matrix is the emerging community standard for AI ad video; pair with Cartesia 3.6 for cheaper consistent voices.
- Neon is expanding beyond Postgres into a Cloudflare-style full backend stack (Object Storage, Functions, Managed Auth, AI Gateway) for agent platforms.
- Codex leads on long-horizon autonomous coding (2-day runs); Cursor tends to stall, and this week's accelerated token burn is pushing builders to GLM and Fable medium as workhorses.
Hot Threads
Meta profile ban after bulk-uploading via system user — safe automation debate
Jev structured-output models: what they actually are and why they matter
One-shot yapper ad pipeline with Grok Imagine + 100-variation matrix