Grok Bot, Ox Alpha GLM 5.3, M5 Ultra Mac Studio — AI Daily Aug 26
577 messages · 76 active members
Overview
Topics
A reconstructed open-source Grok Bot 0.18 on GitHub lets builders swap Claude Code, Codex, or local models as the engine while keeping the polished UX. Separately, @jasonakatiff recommends running Grok alongside Codex on every code review — Grok consistently finds more real issues, even if you build with Claude Code or Fable. Consensus: Grok's coding quality has jumped noticeably.
Bloomberg reporting ties the stealth Ox Alpha model to Z.AI, with weights to be released free. @jcartu says it's GLM 5.3 with image and video towers added — he's already running text-only GLM 5.3 direct from z.ai and finds it faster. GLM 5.3 Flash pricing landed at $0.15/$0.50 per M tokens, with resale via Surplus Intelligence even cheaper.
@dragodimitrov bought a maxed M5 Ultra 256GB Mac Studio with 4TB storage. Discussion centered on running GLM 5.3 Flash at 256k context and ~50 tok/sec reaching Opus 4.5-level quality, plus the trade-off of buying now vs waiting for the 512GB variant in late October or M6 MacBooks.
@jonmacofficial shared his Gemini 3.7 Flash orchestrator (Jarvis personality) assigning tasks to Grok sub-agents inside Orca with OMP. He released three Grok-authored OMP skills — Bug Fix, QA Test, and New Feature. Others reported strong results using GLM in OMP as well.
@scalingfrog complained about Codex adding absurd scope like '5-minute rechecks for 24 hours' unprompted. @jcartu's fix: have Fable spec an ADR with gates and rubrics first, then hand it to the build agent, which cannot exceed scope. @thewildzeno added that auditing peer-review logs and building a failure-pattern tally sharpens both planning and review determinism.
Key Takeaways
- Grok Bot 0.18 has been reconstructed on GitHub — run it with Claude Code, Codex, or local models as the engine while keeping the polished UI
- Run Grok as a second-pass code reviewer even if you build with Claude Code or Fable — it catches issues Codex misses
- Ox Alpha is Z.AI's GLM 5.3 with multimodal towers; weights will be free, and API is $0.15/$0.50 per M tokens ($0.05/$0.15 via Surplus Intelligence resale)
- M5 Ultra 256GB Mac Studio runs GLM 5.3 Flash at 50 tok/sec with 256k context, approaching Opus 4.5 quality for many local tasks
- Force agents to work from a Fable-authored ADR with explicit gates and rubrics to eliminate scope creep — and don't vibe-code trackers to save $49/mo
Hot Threads
Is Grok Bot actually useful, or just for the masses?
Pulling the trigger on maxed M5 Ultra Mac Studio
Fable ADRs to constrain build agents and stop Codex overengineering