Fable 5 Deadline Rush, GLM 5.2 Rigs, Seedance Ads — AI Daily Jul 04

633 messages · 63 active members

633

messages

active members

@jasonakatiff, @jcartu, @fuckyesiwannatalkbusiness

top contributors

Overview

Independence Day found builders at the keyboard, not the grill, racing to squeeze every drop out of Fable 5 before its July 7 Max-plan cutoff. The dominant workflow crystallized: use Fable for planning, skill audits, and sub-agent orchestration (Hermes, OpenClaw, Codex), then hand execution to Sonnet, Opus, or Codex to conserve tokens. Complaints piled up about Fable executing code when only asked questions, mid-session auto-downgrades to Opus 4.8 torching expensive multi-agent runs, and tighter copy guardrails silently rewriting marketing text. @c_1media surfaced the fix: `/config` and disable 'Switch models when a message is flagged.' Members are stacking 3-6 Max plans and using gmail dot/plus tricks to stretch quota through the deadline. Hardware talk went big. @jcartu made a sustained case for a ~$150k EUR 8x 6000-class node (or ~$30k/mo B200 rental) running FP8 GLM 5.2 at 120+ tps for 8 concurrent devs, with GLM 5.5 rumored for August at Opus 4.8 quality plus vision. Draft models (dflash) are roughly doubling throughput — DeepSeek v4 flash jumped from ~180 to 360-400 tps. @robinroy floated splitting the rig cost across a small group. On the creative side, @arielletolome shared a production ad stack: Higgsfield CLI for Seedance, a Hetzner RTX 6000 for 480p→4K upscaling, reusable body+CTA clips, and OpenClaw agents callable from Slack. Wan 2.2 was ruled out for weak lip sync. Side threads rounded out the day: @thewildzeno pushed back on hype that AI can design side-effect-free drugs (side effects are mechanism-linked) while granting massive gains in prototyping and specificity. @jasonakatiff detailed a telephony/CRM/AI-dialer stack and is hiring around it. Hermes v0.18.0 shipped with /journey timelines, MoA swarms, Vertex AI, and a data-broker opt-out skill. And Twitter chatter pegged July 7 for a possible new GPT drop and July 17 for Gemini 3.5 Pro — builders are timing experiments accordingly, with zero loyalty to whichever model wins the week.

Topics

Fable 5 Deadline Rush, Orchestration & Plan Mode115 msgs

With Fable 5 leaving Max plans on July 7, builders are stacking 3-6 subscriptions, using gmail dot/plus tricks, and burning weekly resets. The settled pattern: Fable for planning, skill audits, and orchestrating sub-agents (Hermes, OpenClaw, Codex), then hand execution to Sonnet/Opus/Codex. Plan mode and explicit 'answer only' prompts are the fix for Fable silently executing code when only asked questions.

Claude Opus 4.8 Fallbacks & Guardrail Frustrations33 msgs

Mid-session auto-downgrades from Fable 5 to Opus 4.8 are torching expensive multi-agent runs. @c_1media's fix: `/config` and set 'Switch models when a message is flagged' to false. Separately, Fable 5's copy guardrails are stricter than Opus 4.8, silently rewriting or blocking marketing text — pushing performance marketers toward Opus + orchestrator patterns and Nano Banana for text-in-image work.

Local GPU Rigs for GLM 5.2 / 5.5 Unlimited Inference30 msgs

@jcartu argued a ~$150k EUR 8x 6000-card node (or ~$30k/mo B200 rental) runs FP8 GLM 5.2 at 120+ tps for 8 concurrent devs, with GLM 5.5 in August rumored at Opus 4.8 quality plus vision. Draft models (dflash) roughly double throughput — DeepSeek v4 flash jumped from ~180 to 360-400 tps. @robinroy proposed splitting hardware costs across a small operator group.

Seedance Ad Pipeline & Video Insight Ingestion33 msgs

@arielletolome shared a full ad-creative stack: Higgsfield CLI for Seedance generation, Hetzner RTX 6000 for 480p→4K upscaling, reusable body+CTA clips with swappable hooks, and OpenClaw agents triggered via Slack. Wan 2.2 was ruled out for weak lip sync and 8s duration. @artmalk opened a parallel thread on ingesting 3 months of ad videos, transcripts, and $1.5M+ spend data into a searchable insight system — TwelveLabs vs vector DB vs Obsidian.

Model Release Windows, Drug Discovery & Hermes v0.18.027 msgs

Twitter chatter pegs July 7 for a possible new GPT drop (GPT 5.6 rumored) and July 17 for Gemini 3.5 Pro. GPT-5.5 Extra High is the practical default; Cyber's KYC requirement is friction. @thewildzeno pushed back on AI-designs-side-effect-free-drugs hype — side effects are mechanism-linked — while granting big prototyping gains. Hermes v0.18.0 shipped with /journey timelines, MoA swarms, Vertex AI, per-channel Telegram overrides, and an Unbroker data-broker opt-out skill.

Key Takeaways

Fable 5 leaves Max plans July 7 — stack plans now and use Fable for planning while offloading execution to Sonnet/Opus/Codex to conserve tokens.
If Claude Code auto-downgrades to Opus 4.8 mid-run, `/config` and disable 'Switch models when a message is flagged' to stop expensive tangents.
A ~$150k EUR 8x 6000-card rig runs FP8 GLM 5.2 at 120+ tps for 8 devs; draft models (dflash) are roughly doubling local inference throughput.
Higgsfield CLI + Seedance + a Hetzner RTX 6000 for upscaling is a viable production ad-video stack; Wan 2.2 is not yet competitive on lip sync.
Watch July 7 for a rumored new GPT release and July 17 for Gemini 3.5 Pro — builders are timing experiments and switching with zero loyalty.