Grok Bot Orchestration, GLM 5.3, CMP 170HX Unlocked — AI Daily Aug 22
304 messages · 58 active members
Overview
Topics
anonymous documented a full Grok bot setup: Fable wrote a handoff spec, Grok bot ran an unbiased self-audit, then a chief-of-staff bot merged both plans to build 8 sub-bots. After ~15% of weekly usage he says Grok bot feels smoother than Hermes, close to local model speeds, and delegates cleanly to CC/Codex/local models. Open questions remain on jailbreaking, browser task token burn, and Grok-only model lock-in.
The CMP 170HX has been unlocked and measured at 164 TFLOPS FP16 with 64GB verified for ~$1.2–1.5k, with VLLM 0.26 already running DeepSeek across three cards. Trade-offs remain: PCIe Gen 2, no NVLink, quirky drivers. @jcartu emphasized GLM 5.3 is a tuned 5.2 (~20% smarter, same weights) dropping in a week and fits on four cards — no need to chase Kimi K3's 2.8T params.
@tounano shared his refined Fable pipeline: Opus implementer agents write pragmatic-decision deviation logs, a Fable High orchestrator spawns new agents to resolve them, and a second Fable pass surfaces only what needs human judgment. Result: 20-minute human review of multi-hour runs. Bonus insight — agents reliably self-report when they wrote shitty code.
Vercel's Is Agentic launched an AI Agent Readiness Score for benchmarking stack preparedness for agent workloads. Dreadnode shared research on prompt-level mitigation of model cheating on offensive cyber tasks. Alibaba's Tongyi team published a Qwen-UI-Agent technical report on UI-grounded agent capabilities, and Nile positioned itself as a backend purpose-built for agentic commerce.
@leewardbound is building an Orca skill that renders HTML one-pagers explaining feature branches, bugfixes, or perf test results — replacing terminal walls-of-text. Thesis: at a PC you can absorb dense output, but mobile/away requires visual synthesis. @iannagy independently converged on the same pattern using R2-hosted HTML pages for storyboarding and iteration.
Key Takeaways
- Grok bot's polish and native task-learning from screen recordings make it a serious Hermes challenger, but Grok-only model lock-in is a real constraint for power users.
- GLM 5.3 (dropping in ~1 week) is same-size weights as 5.2 with ~20% smarter tuning — fits on 4 cards, no need to chase multi-trillion-param models like Kimi K3.
- CMP 170HX is unlocked at 164 TFLOPS FP16 / 64GB verified for ~$1.2–1.5k, with VLLM 0.26 already running DeepSeek across 3 cards — a cheap large-VRAM path for inference.
- Deviation-log pattern (agents write pragmatic-decision logs, orchestrator resolves, human reviews remainder) compresses multi-hour agent runs into 20-min reviews.
- For mobile/away workflows, render agent output as HTML one-pagers instead of terminal wall-of-text — dramatically better absorption of complex state.
Hot Threads
Migrating from Hermes to Grok bot as primary orchestrator
CMP 170HX 64GB rigs unlocked for running GLM 5.3 locally
Making AI media-buying agents follow kill rules reliably