Mac Studio M5 Ultra, Codex Harness, Clickflare MCP — AI Daily Aug 25

690 messages · 76 active members

690
messages
76
active members
@jcartu, @samb69, @jasonakatiff
top contributors

Overview

Apple's new Mac Studio with M5 Max/Ultra dominated Sunday's conversation, with 512GB configs landing at ~$17–20K and 4-unit clusters hitting ~4.8TB/s aggregate memory bandwidth — genuine data-center-GPU territory. Builders debated 256GB now vs waiting for the 512GB variant in late October, and @jcartu argued free-weight models like Kimi, GLM 5.5, and Qwen are closing on frontier intelligence fast enough to justify local ownership over $1,300/mo RTX 6000 rentals. He also detailed a 3.5-bit EXL3 trellis quant of GLM 5.2 running across 4 GPUs with 500K KV headroom. Agent harness gripes ran hot: @samarh90 called GPT-sol-high inside the OMP harness 'over-engineered trash' after a 4-hour run he rebuilt in 20 minutes with Cursor, and @thewildzeno clarified that OAuth'ing a Codex sub into a third-party harness doesn't invoke Codex logic — the harness still owns prompts, tools, and API calls. @basant reported 10-minute Codex waits from Thailand while others on 60-hour sessions saw none, pointing to datacenter routing over aggregate load. Paseo (mobile streaming), Orca (full TUI), Herdr, and Manus were compared for daily driving. On the ad-ops and infra side, @expadz showed Clickflare's MCP automating offers, landers, postbacks, and campaigns end-to-end from one Everflow link, while builders confirmed Astro + GitHub + Cloudflare Pages has fully replaced LanderLab-class tools. @jasonakatiff shared a LeadsPedia → in-house migration playbook (Verticals → Offers → Affiliates → Advertisers → Contracts → Campaigns) and a pen-test loop using Kimi K3 to pull top GitHub libs and chain attacks across models. @tounano's RLVR thread explained why agents code well but converse awkwardly — they're optimized for verifiable pass/fail — and the group agreed Chinese labs should openly acknowledge OpenAI distillation so efficiency claims can be judged fairly.

Topics

Apple's M5 Max/Ultra Mac Studio launched with up to 512GB memory (late October) and clustering to ~4.8TB/s across 4 units, at ~$10K for 256GB and ~$17–20K for 512GB. @jcartu pushed a 3.5-bit EXL3 trellis quant of GLM 5.2 across 4 GPUs with 500K KV free, and argued free-weight models (Kimi, GLM 5.5, Qwen) make local abliterated inference a real alternative to Nvidia Spark or renting RTX 6000s.

@samarh90 said GPT-sol-high in the OMP harness took 4 hours to produce work he redid in Cursor auto in 20 minutes; @thewildzeno clarified that OAuth'ing a Codex sub doesn't mean Codex logic runs — the harness owns system prompts, tool schemas, and API calls. @basant's Codex slowdowns from Thailand vs @jcartu's clean 60-hour US session pointed to datacenter routing over infra saturation. Paseo (mobile streaming), Orca (full TUI), Herdr, and Manus were compared as daily drivers.

Builders confirmed Astro JS + GitHub → Cloudflare Pages has fully replaced LanderLab and similar paid landing tools — free if you already have a Claude sub. @expadz showed Clickflare's MCP on the $89 starter plan automating offers, landers, postbacks, and campaigns from one Everflow link, while @tounano argued CLI scripts beat MCP unless you're shipping an agentic product for others to extend.

@tounano explained RLVR (Reinforcement Learning Verification Reward) as why modern agents code well but communicate strangely — they're optimized for verifiable pass/fail, not prose. The thread agreed distillation of OpenAI outputs by Chinese labs is inevitable and useful, but labs should openly acknowledge it so efficiency claims can be evaluated; without US labs pushing the frontier, a distillation-only path stalls.

@jasonakatiff shared a LeadsPedia → in-house migration order (Verticals → Offers → Affiliates → Advertisers → Contracts → Campaigns), noted the APIs are usable, and open-sourced packet-sniffed LeadsProsper internal APIs on BuiltWithAI. He also detailed a pen-test loop using Kimi K3 to pull top GitHub pen-test libs, run multi-model attacks, then chain exploits — with frequent git pushes as the real safety net after a Claude run wiped a codebase.

Key Takeaways

  • M5 Ultra Mac Studio clusters hit ~4.8TB/s aggregate bandwidth; combined with free-weight models like GLM 5.5 and Kimi, local frontier-class inference is finally practical at ~$17–20K.
  • OAuth'ing a Codex subscription into a third-party harness doesn't invoke Codex logic — the harness still owns prompts, tools, and API behavior, so blame the planning layer, not auth.
  • Astro + GitHub + Cloudflare Pages fully replaces paid landing platforms, and Clickflare's MCP on the $89 tier automates the entire ad ops setup from one Everflow link.
  • RLVR training explains why agents write great code but converse awkwardly — they're optimized for verifiable pass/fail, not natural prose.
  • Frequent git pushes beat trying to bulletproof agents; for lead-platform migrations, port in dependency order (Verticals → Offers → Affiliates → Advertisers → Contracts → Campaigns).

Hot Threads

@jcartustarted

Paseo vs Orca for mobile agent steering and M5 Ultra local rigs

25 replies8 participants
@basantstarted

Codex resets and infra saturation vs datacenter routing

32 replies5 participants
@jasonakatiffstarted

LeadsPedia migration order and reverse-engineered LeadsProsper APIs

12 replies3 participants

Linked Items