GLM 5.2 Launch, Hermes vs Openclaw, Posthog Ad Ops — AI Daily Jun 14

276 messages · 62 active members

276
messages
62
active members
@jcartu, @jasonakatiff, @Kieran
top contributors

Overview

The Fable 5 hangover continued today as builders compared notes on the GLM 5.2 release from China — @jcartu confirmed it's a strong Opus replacement with 1M context and smarter than Kimi, but nowhere near the promised 3x speed bump over 5.1. Throughput tests on his 4x RTX Pro 6000 Blackwell rig showed thinking-heavy models still bottlenecked, fueling broader chatter about local inference rigs (TP-8 setups running Kimi/GLM) as the long-term play once frontier subs go away. Orchestration was the other dominant thread: multiple builders (@swh800, @jarvisballer, @weslindquist) confirmed Hermes is decisively winning over Openclaw as the orchestrator layer on top of siloed Claude Code agents, with Slack and Discord interfaces in production. Memory architecture stayed pragmatic — well-structured markdown files beat Obsidian and graph visualizations for most use cases, though graphify came up for codebase work. On the ops side, @zippi101 sparked a useful exchange on wiring FB ads + revenue into Posthog and using its native LLM for goal-driven campaign optimization, while @Coybh discovered Claude's 'fast mode' silently burns API credits even with plan allowance remaining — a billing UX builders are increasingly frustrated with versus GPT's included fast tier.

Topics

GLM 5.2 Release: Smart, Cheap, Still Slow18 msgs

China's GLM 5.2 dropped and @jcartu benchmarked it as smarter than Kimi and a viable Opus replacement with 1M context, but the promised 3x speed improvement over 5.1 didn't materialize. Best fit for TP-8 local setups or budget-conscious intelligence work where latency isn't critical.

Hermes Beats Openclaw as Orchestrator Layer22 msgs

Multiple builders independently confirmed migrating from Openclaw to Hermes as their orchestration layer above siloed Claude Code agents. @jarvisballer published a setup video, @swh800 switched in 30 minutes, and @weslindquist runs Hermes as orchestrator over 3 Claude Code instances and 2 Openclaw agents on OpenAI.

Posthog LLM for Goal-Driven Ad Optimization14 msgs

@zippi101 asked how to wire FB ads, revenue, and Posthog quiz funnel data into an AI agent for autonomous campaign optimization toward a $1000/day target. @jasonakatiff recommended pulling everything into Posthog's native FB integration and using the Posthog LLM or Claude Code MCP rather than Hermes Desktop.

Claude Fast Mode Silently Burns API Credits12 msgs

@Coybh discovered Claude's fast mode was eating API credits despite plenty of plan allowance remaining — even Claude itself couldn't explain the billing. Contrasts sharply with GPT-5.5 which includes fast mode in the plan, adding to broader Anthropic billing/agent-usage frustration ahead of tomorrow's policy change.

Local Inference + Frontier Orchestration as the Long Game10 msgs

@geilt argued that within 6 months, frontier subs disappear and AGI-level inference goes to highest bidders only — making local rigs (Kimi, GLM via TP-8) the only fallback. Chinese open-weights models keep closing the gap; harnesses are the remaining bottleneck. Custom harnesses + workflow-specific trained models seen as the durable stack.

Key Takeaways

  • GLM 5.2 is a legit Opus replacement (smarter than Kimi, 1M context, super cheap) but not fast — use it for intelligence-heavy fallback, not latency-sensitive workflows.
  • If you're still on Openclaw as orchestrator, the community has decisively moved to Hermes — migration takes ~30 min and the agent UX is meaningfully better.
  • Check Claude's fast mode setting before assuming your plan covers usage — it silently bills API credits even with allowance left, unlike GPT-5.5 which includes it.
  • For ad ops automation, skip the Hermes Desktop layer — Posthog's native FB integration + Posthog LLM (or Claude Code MCP) handles goal-driven optimization directly.
  • Memory architecture: well-structured markdown files still beat Obsidian/graphify for agent context — only reach for graph tooling on large codebases.

Hot Threads

@zippi101started

Wiring FB ads + Posthog + revenue into an AI agent for autonomous $1000/day campaign optimization

14 replies4 participants
@weslindquiststarted

Using AI agents as business strategy coach / board of advisors — memory structures, Hermes orchestration, obsidian vs markdown

18 replies5 participants
@jcartustarted

GLM 5.2 throughput benchmarks on 4x RTX Pro 6000 Blackwell — speed reality check vs 5.1

11 replies6 participants

Linked Items

Claude AI Writes Epic 1500-Line Philosophical Fable About Consciousness

github.com
@bartekadamczyk

Budget AI Panel Beats GPT-5.5 & Matches Claude Fable 5 at Half Price

x.com
@Wootbro

Unable to analyze - insufficient video information provided

youtu.be
@arielletolome

Paul Graham's Blueprint to Earning a Billion Dollars

x.com
@instanetworks

Unable to analyze - Instagram login required

instagram.com
@jasonakatiff