Local AI Rigs, Palmier Pro, Hermes Memory — AI Daily Oct 24
519 messages · 63 active members
Overview
Topics
@jcartu made the case that no sub-€100k local setup touches frontier models like DeepSeek V4.1 Flash or Opus 5.5, and even at 30-50 TPS on a $10k rig you're 'walking vs driving a Lamborghini' compared to 500+ TPS cloud. Compounding the problem, @geilt flagged DDR5 RAM prices 4-5x'd since July. Consensus: multi-account API routing wins unless you're a hobbyist or running 8+ devs 24/7.
@ariangrtyi showcased Palmier Pro (open-source macOS AI video editor) after a fully AI-generated ad outperformed manual edits. @zippi101 demoed 'Videos that Sell,' a Seedance-powered one-shot 30-second ad generator. The group compared Wan 3 (480p price/quality winner), Kling 4, and Higgsfield, with ElevenLabs voice-to-voice recommended for lip-sync polish and @WhiskeyATX endorsing Grok for YouTube policy neuters.
@jcartu detailed a hook-based pattern that reinstalls personal Hermes customizations on every update and lets the agent self-teach new Hindsight features, arguing Hindsight is SOTA over mem0 — @GuruTime committed to a sandboxed head-to-head. Separately, Hermex shipped a bot mode that could replace Grokbot for team Telegram workflows, with @samb69 noting Hermes already has an official bot mode worth comparing.
@anonymous tested Kimi 3.1 Pro and suspects heavy distillation from GPT Astra 6, calling it 'too compliant.' With Kimi, Minimax 3.1 Flash, and Deepseek all shipping recently, the group expects GLM 5.4 imminently. Opus 5.5 is reportedly faster this week, while Codex resets are hitting so often users can'anonymous even bank them — prompting jokes about a token exchange marketplace.
@robinroy kicked off a thread about running agents from phone, with suggestions including Hoplite.sh, Hermes-via-Telegram patterns, and OpenClaw. @watchmedropship noted you can port Hermes to iOS via an Apple developer account. Separately, @miyamotaz surfaced wiltodelta/remove-ai-watermarks, which strips visible marks and invisible provenance metadata (SynthID, C2PA, EXIF, IPTC, XMP) from AI-generated media.
Words worth knowing
Three terms from the glossary.
- Media BuyingRunning paid advertising campaigns, primarily on Facebook and Google, to acquire customers or leads.
- Max PlanThe premium subscription tier for Claude Code that costs ~$200/month with higher usage limits.
- ClaudeAnthropic's AI model family, the most widely used model in this community for coding and reasoning.
Key Takeaways
- Local AI rigs still don'anonymous pencil out vs cloud APIs unless you're a hobbyist or running 8+ devs 24/7 — DDR5 RAM prices have 4-5x'd since July, making the gap worse.
- Wan 3 at 480p with reference images is the current price/quality sweet spot for AI video ads; route output through ElevenLabs voice-to-voice for lip-sync polish.
- Hindsight 1.0 can be auto-upgraded via a hook that lets your Hermes agent re-apply customizations and self-learn new features on each update.
- DeepSeek V4.1 Flash is the best model realistically runnable on 4 GPUs today, enabled by RAM/SSD engram offloading.
- Expect GLM 5.4 soon after the Kimi 3.1 Pro / Minimax 3.1 Flash / Deepseek release cluster; Opus 5.5 is running faster this week.
Hot Threads
Local setup vs cloud after Anthropic earnings scare
Videos that Sell: one-prompt Seedance ad generator
Hermes + Hindsight upgrade workflow and mem0 comparison