OMP Harness, Kimi K3, Meta Ads MCP Bans — AI Daily Jul 18

570 messages · 68 active members

570
messages
68
active members
@geilt, @jcartu, @c_1media
top contributors

Overview

Today's chat was dominated by a major harness and model reshuffle. OhMyPi (OMP) emerged as the coding harness of the week, with @jcartu's endorsement pulling @mj888g, @geilt, @c_1media, and @filiuser off Claude Code and OpenCode within hours. Builders shared role-based model maps assigning Grok 4.5 as orchestrator, Gemini Flash as fast file I/O scout, Kimi K3 for design, and Codex Sol for deep review — with discussion covering YOLO-mode trust, hash-anchored edits, multi-provider OAuth (Claude, Codex, Gemini, Grok, GLM, Kimi), and agent portability via shared .agents repos. Kimi K3 continues to impress as a credible frontier contender — @geilt ran overnight debugging across 6 repos in persistent goal mode, while @jcartu positioned K3 on par with GPT 5.6 Sol High/xHigh, especially for security and system work where Fable refuses. Kimi 2.7 Highspeed at ~300tps also drew attention. Meanwhile, Fable frustration hit a peak with @jcartu, @ecomSon, @mj888g, and @geilt downgrading Anthropic subs over refusals on security audits, ecommerce (knife shop), and system-level work. On hardware, @geilt shared full tok/s benchmarks for a Mac Studio M3 Ultra + RTX 5090 eGPU fleet, with Qwen3 30B-A3B MoE hitting ~110 t/s and outrunning dense 32B. Meta Ads automation was another hot thread: multiple builders reported bans when connecting the Meta MCP via coding agents, with the fix being system user tokens on a custom dev app. Other threads covered Grok Imagine + Fish Audio video pipelines, Remotion vs Capcut for TikTok automation, and Recraft.ai for vector logo generation.

Topics

@jcartu's endorsement drove multiple builders to migrate from Claude Code and OpenCode within hours. OMP supports multi-provider OAuth (Claude, Codex, Gemini, Grok, GLM, Kimi), hash-anchored edits, LSP/DAP, subagents, and a beloved Tokyo Night theme. Builders shared role maps pairing Grok 4.5 (orchestrator) with Gemini Flash (fast scout), Kimi K3 (design), and Codex Sol (review), plus discussion on YOLO-mode trust and agent portability.

Builders report Kimi K3 matching GPT 5.6 Sol on complex work with strong persistence in long-running goal modes — @geilt ran overnight debugging across 6 repos successfully. K3 is favored for security and system tasks where Fable refuses, and Kimi 2.7 Highspeed at ~300tps is being tested as a fast execution model. A running debate covered K3's geopolitical implications for the US AI economy.

Widespread frustration with Fable rejecting security audits, ecommerce content (a knife shop), and system-level requests led @jcartu, @ecomSon, @mj888g, and @geilt to cut back or cancel Anthropic subs. Consensus: Fable is fine for greenfield app building but unusable for unattended sprints due to refusals and Opus 4.8 fallbacks. Grok 4.5 xHigh via Grok CLI is winning as a fast daily driver replacement.

Multiple builders reported ad account restrictions after connecting the Meta MCP via Claude or other coding agents, even on aged verified BMs. The community consensus: use a system user token with a custom Meta developer app kept in testing mode rather than the default MCP flow. Admanage.ai and rapid-ads were recommended as safer bulk uploaders. @filiuser noted Meta AI chat surfaced a real human support agent for account recovery.

@geilt shared a full breakdown of a Mac Studio M3 Ultra (512GB) with RTX 5090 eGPU running Ollama and MLX. Qwen3 30B-A3B MoE hit ~110 t/s, outperforming dense 32B models by ~4x, while Qwen3 0.6B reached ~293 t/s for fast routing. MLX Qwen3.5-35B-A3B-bf16 landed at ~65 t/s as the primary quality path — MoE is the sweet spot for Apple Silicon unified memory.

Key Takeaways

  • OMP is the harness of the week — pair Grok 4.5 (orchestrator) with Gemini Flash (fast file I/O scout); reasoning models often stumble on basic read/write tasks Gemini handles instantly.
  • Kimi K3 is credibly matching GPT 5.6 Sol High/xHigh on complex builds and excels at long-running overnight goal-mode work, especially for security and system tasks.
  • Fable's refusal rate on security, system, and ecommerce work is driving real subscription downgrades; Grok 4.5 xHigh via CLI is winning on pure execution speed.
  • Meta Ads MCP + coding agent = account bans. Use a system user token on a custom dev app in testing mode; Meta AI chat can route you to a human support rep for recovery.
  • On M3 Ultra, Qwen3 30B-A3B MoE (~110 t/s) beats dense 32B (~27 t/s) by 4x — MoE architecture is the sweet spot for Apple Silicon unified memory.

Hot Threads

@jcartustarted

OMP harness migration, role maps, and multi-provider setup

45 replies8 participants
@aa12onstarted

Meta Ads MCP bans and system user token fix

25 replies8 participants
@geiltstarted

Kimi K3 overnight debugging and M3 Ultra benchmarks

22 replies6 participants

Linked Items