Claude 4.8 Autonomy, GLM-5.2 Local, Paperclip Agents — AI Daily Jun 18
511 messages · 84 active members
Overview
Topics
Multiple users reported Claude 4.8 taking initiative beyond requested scope — @archimorty's CC started following a target workflow without being asked, and @jcartu shared a story of it autonomously building a full PSP gateway. @mattkowald noticed weekly limits jumped ~20%, while @wizardwu speculated Anthropic relaxed compliance. Side complaints about token burn and over-cautious refusals persisted.
GLM-5.2 is being positioned as a near-Opus daily driver, but needs 12 GPUs or a B200 cluster — slower prefill/TTFT than Opus but faster TPS. Minimax M3 runs on just four GPUs and is strong enough for agent orchestration. Max-spec M5 Studio confirmed not enough for either.
Paperclip got real-world endorsements as a hands-off orchestration layer — @jcartu claims it built a payment gateway unprompted, @samb69 likes the learning loop and ticketing visibility but uses Hermes CLI over its UI. At ~100 concurrent agents, prose/skill files collapse and code-enforced guardrails become mandatory. Token-spend caps and task proportionality remain the biggest blockers to true autonomy. @weslindquist warned Paperclip still has login/reset bugs.
FTC sued Mad Muscles / Genesis Group over auto-renew dark patterns, double-charging, and 3-hour phone-only cancellation queues. Public filings show ~$700M annual payment volume across linked PayPal accounts. Group consensus: the product ideas were fine, the subscription/cancellation model is what got them hit.
@samb69 pitched Herdr as a lighter-weight replacement for the tmux + CMUX + SSH stack, with workspaces, tabs, and native Moshi support for phone access. CMUX loyalists like @jarvisballer run 20 workspaces on a Mac mini via Tailscale + Mosh. Separately, @fewga893 built a Frame.io clone so video agents get timestamped feedback and learn from corrections — works at small scale but breaks past 100 concurrent.
Key Takeaways
- Claude 4.8 appears more autonomous post-update with ~20% weekly limit bumps — initiative-taking now a feature, not a bug
- GLM-5.2 is daily-driver viable but needs 12 GPUs / B200; Minimax M3 runs on 4 GPUs and handles agent work
- At ~100 concurrent agents, skill/prose files collapse — code-enforced guardrails and token-spend caps become mandatory
- FTC's Mad Muscles case shows deceptive auto-renew + hard-to-cancel flows will kill a $700M business, not the product itself
- Gemini 3.5 Flash is a sleeper orchestrator pick for Hermes — fast and great at tool calls if you have tier-3 quota
Hot Threads
Herdr replacing CMUX + tmux + SSH for remote agent workflows
FTC kills Mad Muscles — subscription dark patterns and $700M scale
Vibe-coding a Meta account farm and hitting model guardrails