Multi-Model PR Audits, Codex Loops, Local GPU Rigs — AI Daily May 06

785 messages · 96 active members

785
messages
96
active members
@jasonakatiff, @arielletolome, @Wootbro
top contributors

Overview

Today's 785 messages converged on tightening AI coding workflows and rethinking compute. @GuruTime launched panel-mcp-server (aka multiaudit), an MCP that fans out a PR to Codex, Gemini, Grok, and Claude in round 1, has them debate in round 2, and synthesizes via judge — surfacing 49/50 real issues on a Vercel-style codebase for ~$350 on Opus API. Meanwhile @jasonakatiff reported cutting his planning loops from 3-5 iterations to just 2 passes (post-plan + post-build) and landing a multi-surface telephony refactor almost first-pass. Claude Opus 4.7 dominated the gripe channel: hangs every 1-3 hours and only ~10k tokens per 30 minutes pushed @expadz, @ricccdes and others back to Codex. Anthropic's late-day announcement of doubled code limits and expanded API access — alongside a new SpaceX compute partnership — was received as overdue relief. On hardware, @jcartu detailed his 'Rasputin' rig (Xeon w9-3495X, 256GB DDR5 ECC, multiple watercooled RTX Pro 6000 Blackwells) running a 27B local coding model to escape token costs, and open-sourced an LLM stress harness. Ops and creative tracks stayed busy: Kie + Veo confirmed parallel video gen (6 videos × 4-7 scenes simultaneously), Meta dropped the video requirement for higher-tier API access, @arielletolome started a Gemini-powered call QA agent (citing YouTube-trained audio/video edge) and integrated the Wheel of Emotions into AI creative briefs. Industry-wise, OpenAI partnered with BCG/McKinsey while Anthropic launched an enterprise services arm, squeezing solo AI consultants.

Topics

@GuruTime shipped panel-mcp-server, orchestrating Codex, Gemini, Grok, and Claude through round-1 fan-out, round-2 debate, and judge synthesis. Found 49/50 real issues on a production codebase for ~$350 on Opus API, with line-level findings detailed enough for Codex to auto-patch. Includes a bugfind mode with live web UI; @jasonakatiff felt it was overkill at his current bug rate but valued Codex as second opinion.

@jasonakatiff replaced his 3-5 loop pattern with just 2 reviews (post-plan, post-build) and landed a multi-surface telephony refactor with tenant-level config sharing almost first-pass. Credits worktrees, session summaries in CLAUDE.md, and 'no hacks/stubs/mocks' at top of context. Builders converged on audit loops: Claude plans → Codex CLI reviews → Claude validates.

Opus 4.7 hangs every 1-3 hours and emits only ~10k tokens per 30 minutes, sending @expadz, @ricccdes and others back to Codex. Anthropic announced doubled code limits, expanded API access, and a SpaceX compute partnership (read as pragmatic compute-for-cash, leveraging Starlink idle capacity). Side debate on Elon's leverage and Musk's plan to fold xAI into SpaceX.

@jcartu detailed his multi-RTX Pro 6000 Blackwell 'Rasputin' rig (Xeon w9-3495X, 256GB DDR5 ECC, dual 2800W PSUs, watercooled), targeting 300+ tps on a 27B coding model with Opus scaffolding. He open-sourced an LLM stress harness and NVFP4 experiment repo. @geilt expanded his M3 Ultra Mac Studio with TB5 SSDs and a 5090 via PCIe adapter.

Kie + Veo confirmed parallel API calls (6 videos × 4-7 scenes simultaneously) without multi-account hacks; Seedance 2.0 at ~$1/10s is the ad-workflow pick. Meta dropped the video requirement for higher-tier API access. @arielletolome started a Gemini-powered call QA agent and integrated the Wheel of Emotions into AI creative briefs; @justingacina shared AssemblyAI + Claude patterns for buyer complaint triage.

Key Takeaways

  • Multi-model debate panels (parallel fan-out → debate → judge) are the highest-signal code audit pattern — ~$350 on Opus API surfaced 49/50 real bugs.
  • Two review passes (post-plan + post-build) ship near bug-free code for experienced builders — drop the 3-5 loop habit and use Codex as second opinion.
  • Claude Opus 4.7 is rationing tokens; Anthropic's doubled code limits and SpaceX compute deal should ease pressure, but route to Codex when it hangs.
  • Kie + Veo supports true parallel video gen and Meta dropped the video requirement for higher-tier API access — major unlocks for ad automation.
  • Custom Claude skills built from specific books (Breakthrough Advertising, Cashvertising, Bandler NLP) consistently beat generic agents for domain creative work.

Hot Threads

@jasonakatiffstarted

Cutting Codex loops to 2 passes and shipping multi-surface refactors first-pass

30 replies8 participants
@GuruTimestarted

panel-mcp-server: 4-model debate audit finds 49/50 real bugs on Opus API

14 replies4 participants
@jcartustarted

Rasputin: multi-RTX Pro 6000 Blackwell rig and 27B local coding workflow

18 replies6 participants

Linked Items