Daily Digest — Saturday, April 25, 2026
569 messages · 70 active members
Overview
Topics
RASPUTIN Shutdown & Memory SOTA Pivot
18 msgs@jcartu archived RASPUTIN (72.4% LoCoMo) because Hindsight (89.6%) and Mem0's April rewrite (91.6%, under 7K tokens/call) moved the field 18-20 points in 60 days. He recommends combining Hindsight (fact retrieval) with Honcho (theory-of-mind/psychological modeling) as the new stack, with existing qdrant users pivoting directly to mem0.
Opus 4.7 Performance Saga & GPT-5.5 Migration
43 msgsAfter 2-3 rough days with widespread complaints about slowness and bugs (theorized as Anthropic compute rationing), 4.7 on xHigh/max effort is now reportedly matching or surpassing peak 4.6 for complex code. Mid-week workarounds included flipping to 4.6 mid-session and migrating orchestration to GPT-5.5 via Codex despite 2x cost and more hallucinations. Anthropic published a postmortem on the April 23 outages.
cmux + Hermes: New Terminal Stack for Multi-Agent Workflows
49 msgs@jasonakatiff championed cmux as a low-memory alternative to running CC inside VSCode/Cursor, with users managing 30-40 concurrent sessions across project tabs. /rename + /resume from session start saves you from crashes and SSH drops. @Wootbro reported Hermes ran a 5-6 hour autonomous ULTRAWORK PRD overnight where OpenCode would have halted for permissions. Note: roll back CC 2.1.120 → 2.1.119 to fix the broken session resume bug.
iamgalba's 1M-Line Google Ads Agentic System
25 msgsDetailed breakdown of a 114-skill system covering shopping feeds, RSAs, PMax, presell/offer pages, Demand Gen video pipelines, competitive intel, and self-improving prompt evals. Built over 2 years with $100k+ in token spend, runs on 5x Claude Code Max + 1x Codex Pro, manages ~$10M in ad spend, and includes a refinement learner that auto-patches prompts from human edits. Available at tegra.co.
$5k/Mo AI Courses vs. Cheaper Alternatives & LLM Reliability for Finance
44 msgsStefanBrain ($5k/mo) and Evolve ($1.5k/mo) sparked debate, with @Coybh recommending Franky (sub-$100) and 0xRoas as better-value alternatives—consensus: AI moves too fast for static courses, foundational DR principles get recycled across all of them. Separately, @ericshaf argued LLMs are borderline useless for financial data due to silent hallucination corruption; @jcartu countered with deterministic tool calls, thinking-disabled modes, and variance-detection guardrails.
Key Takeaways
- Memory SOTA jumped 18-20 points in 60 days—Hindsight (89.6%) + Honcho is the new recommended stack, with Mem0's April rewrite hitting 91.6% LoCoMo under 7K tokens/call.
- cmux is the new go-to terminal for AI coding—dramatically lower memory than VSCode/Cursor when running multiple concurrent agents; use /rename + /resume from session start to survive crashes.
- Claude 4.7 on xHigh/max effort has caught up to peak 4.6 for complex code after a rocky launch week—give it another shot if you bounced early. Roll back CC 2.1.120 → 2.1.119 for session resume.
- For hallucination-sensitive domains, don't let the LLM BE the solution—have it BUILD the solution: deterministic tool calls with thinking disabled, single source of truth, pandas/numpy scripts.
- Hermes is emerging as the preferred harness for long-running autonomous work because it doesn't halt for permissions like OpenCode does—ideal for overnight ULTRAWORK PRDs.
Hot Threads
Full breakdown of 1M-line Google/Bing Ads agentic system with 114 skills
cmux vs VSCode/Cursor for managing Claude Code sessions
RASPUTIN shutdown and pivot to Hindsight + Honcho memory stack