Hermes Fleets, GLM-5.2, Meta Ban Wave, Midjourney MRI — AI Daily Jun 17

653 messages · 87 active members

653
messages
87
active members
@arielletolome, @jarvisballer, @Coybh
top contributors

Overview

Today's chat split between agent architecture, a wave of Meta account bans, and Midjourney's surprise pivot into MRI hardware. @jarvisballer, @rmktg, and @sibunting went deep on multi-profile Hermes/OpenCode fleets — one orchestrator plus department-specific sub-agents, locked-down soul.md/fleet.md edits behind a single fixer agent, GitHub backups for rollback, and per-team-member assistants. Kieran walked through his Next.js + Supabase client dashboards, fixing slow renders with virtualization and pagination, with debate on whether dashboards or 9am Slack bot reports actually get used by clients. On models, Z.ai's GLM-5.2 (open weights) reportedly beats GPT-5.5 on long-horizon coding benchmarks at 1/6th the cost, sparking plans to test it against Opus and Kimi. Anthropic quietly shipped Claude Design, and members are running omo/Sisyphus harnesses on VMs to drive coding agents remotely. The recommended workflow remains Opus 4.8 in plan mode for architecture, Codex for execution, and avoid Haiku for complex planning. Coybh made the case that Fable is the only model with real design taste, while Opus/GPT default to AI slop. A Meta ban wave hit multiple agencies — random BMs and personal profiles disabled with no appeal — reinforcing the rule to never connect personal profiles to BMs. On creative, @jonmacofficial argued Kling 3 wins on cost/quality/speed for AI UGC, with Seedance 2 reserved for VSLs; Higgsfield's 'unlimited' Seedance turned out to have hidden caps. Side threads covered Seedance face-detection bypass via grid overlays, free Nano Banana spam via Chrome MCP, AI-driven affiliate payout reconciliation, forex signal bots, and Midjourney's stunning pivot to medical imaging hardware after losing the image-gen race.

Topics

Deep dive on running Hermes/OpenCode with one orchestrator plus department-specific sub-profiles, each with their own skills and 1Password vaults. Lock soul.md/fleet.md edits behind a single fixer agent, back to GitHub, and manage compactions at 80% context. Kieran shared his Next.js + Supabase client dashboard stack, fixing slow renders with virtualization/pagination, with consensus that dashboards alone get ignored — pair them with a chat bot that explains the data.

Z.ai's open-weights GLM-5.2 reportedly beats GPT-5.5 on long-horizon coding at 1/6th the cost, prompting builders to test it against Opus, Kimi, and DS4f. Anthropic shipped Claude Design, and members are running omo/oh-my-openagent and Sisyphus harnesses on VMs. Recommended flow: Opus 4.8 in plan mode (shift+tab) for architecture, then Codex for execution. Coybh called Fable the only model with real design taste vs Opus/GPT slop.

Multiple agencies reported personal FB/IG accounts and connected BMs disabled with no clear cause or appeal path. Consensus: never connect personal profiles to BMs, expect false positives, and ban batches follow same-device/IP fingerprints. @RayTGcontact's thread on AI reconciling network payouts across bank statements drew working setups from @justingacina. Side discussion on affiliate verticals — igaming strong, sweeps/bizop dying, telehealth viable.

@jonmacofficial showed ViralView examples arguing Kling 3 is the best cost/quality/speed tradeoff for AI UGC; Seedance 2 wins for VSLs because start/end frame syncing reduces editing; Omni has marginally better mouth sync but worse economics. Higgsfield's 'unlimited' Seedance has hidden rate limits — direct API + Claude orchestration is the power-user move. Seedance face-rejection bypassed via 4x4/6x6 grid overlays (lumiying.com); @victorbrss shared free Nano Banana + Chrome MCP skill for static/video at scale.

Midjourney announced a medical imaging / MRI alternative — a bold hardware moat play after losing image-gen ground to Google and OpenAI, especially notable as a community-funded 9-figure-revenue company with no VC. Separately, @arielletolome's AI forex signal bot (5 signals, multi-timeframe, 40% win rate at 1:3-1:4 R:R, manual execution only) drew skepticism that AI edge is limited in bull markets. Image-to-3D meshes are solved via a Chinese tool; rigging/animation remain the next frontier.

Key Takeaways

  • Run Hermes as one orchestrator + department profiles; gate soul.md/fleet.md edits behind a single fixer agent and back to GitHub for rollback safety.
  • GLM-5.2 (open weights) reportedly beats GPT-5.5 on long-horizon coding at 1/6th the cost — test it against Opus and Kimi this week.
  • Never connect personal Meta profiles to BMs — ban waves are random with no appeal path; pair client dashboards with a chat bot or they get ignored.
  • For AI UGC: Kling 3 wins on cost/quality/speed, Seedance 2 for VSLs only; Higgsfield 'unlimited' has hidden caps — go direct to APIs via Claude.
  • Use Opus 4.8 in plan mode for architecture, hand off to Codex for execution; Fable is the rare model with actual design taste vs Opus/GPT slop.

Hot Threads

@arielletolomestarted

Facebook and Instagram accounts disabled, 15 BMs affected in Meta ban wave

25 replies11 participants
Kieranstarted

Slow Next.js client dashboard — caching, virtualization, and dashboard vs. agent debate

30 replies8 participants
@jonmacofficialstarted

Kling 3 vs Seedance 2 vs Omni for AI UGC with ViralView examples

22 replies8 participants

Linked Items