GLM-5.2 Drops, Claude Outage, Patter AI Voice — AI Daily Jun 16
705 messages · 86 active members
Overview
Topics
Zhipu's GLM-5.2 dropped on Hugging Face at 750B params, reportedly approaching Mythos quality at 150 TPS but requiring 6 GPUs (MiniMax M3 NVFP4 fits on 4). @jcartu detailed his stack: Kimi 2.6 turbo on Fireworks at 150-170 TPS, DS4f local at 270 TPS with 1M context, Opus 4.8 for scaffolding via OpenCode + Sisyphus. Kimi 2.7's 6x speed tier could hit 600-1000 TPS on Cerebras.
Claude went down mid-day, hitting users who'd maxed out Codex as backup. Builders pushed back on Anthropic's updated privacy policy and the June 22 API-only transition, with several declaring they'll switch to GLM-5.2 or Zen Black. OpenAI rolled out free Codex rate limit resets with a referral program through June 24.
@arielletolome shared Patter AI at $0.025/min — roughly $1.50/hour call center agents, undercutting Retell and VAPI — plus a $200k/mo ACA transfer playbook that raised TCPA concerns. She also shared a Claude-driven trading bot doing top-down analysis across S&P 500, Nasdaq, gold, silver, and FX with a 38% papertrading win rate; @fewga893 flagged liquidity/slippage gaps in live trading.
@justingacina kicked off a thread on pulling Chase transactions for invoice tracking, with the group converging on QuickBooks API as easier than Plaid direct. @weslindquist shared a production setup running continuous AI bookkeeping for 10 businesses, recommending Ramp and Bill.com for vendor payments plus the $10/mo QB ledger tier for accounting firms.
Hermes Atlas launched as a curated registry of 169+ open-source tools across 12 categories for the Hermes Agent. @sibunting raised SOTA coding agents passing QA but piling on defensive bloat — ponytail floated as mitigation. Builders are using Seedance for single-image animation and stacking prop libraries, routing around OpenAI's tightening similarity guardrails with Stable Diffusion pipelines.
Key Takeaways
- GLM-5.2 (750B) is live on Hugging Face and reportedly approaches Mythos/Fable quality at 150 TPS — needs 6 GPUs locally; MiniMax M3 NVFP4 quant is the 4-GPU fallback.
- Anthropic's June 22 API-only transition is pushing power users to Chinese models; Codex now banks free rate limit resets and runs a referral program through June 24.
- Kimi 2.6 turbo delivers 150-170 TPS at ~10x cheaper than Opus; pair with DS4f local (270 TPS, 1M context) and Opus 4.8 for planning in OpenCode + Sisyphus.
- Patter AI at $0.025/min makes sub-$2/hour voice call centers viable — but TCPA exposure on outbound AI dialing is the real constraint.
- For bookkeeping automation, QuickBooks API approval is easier than direct Plaid; Ramp/Bill.com handle vendor payments cleanly, and the $10/mo QB ledger tier works for firms managing multiple clients.
Hot Threads
GLM-5.2 release plus Kimi/DS4f local stack and OpenCode workflow
Automating Chase transaction pulls and continuous AI bookkeeping
Patter AI voice agent economics and $200k/mo ACA transfer playbook