Gemini 4 Argon, Hermes Tuning, Seedance 2.5 — AI Daily Oct 19
530 messages · 72 active members
Overview
Topics
@geilt ran Gemini 4 Argon (Medium mode) through AGY CLI overnight to migrate apps off multiple macOS servers onto Linux, configure AWS/Cloudflare/GoDaddy, stand up a Grafana+Loki+Mimir+Tempo observability stack bridging bare metal and AWS, handle KMS secrets, SAML, and security scans. Slower than Flash but very impressive end-to-end. Raised questions on auth risk when giving CLI agents that much surface area.
@GuruTime shared a four-week Hermes tuning methodology: cache fixes to 91-98% hit rate, compaction at 180K tokens, memory hygiene (facts to mem0, procedures to skills, no stale ROAS snapshots), promise-guard plugins, and daily tuning guards to survive updates. Separately, builders debated migrating from Hermes/Orca to OpenMausBot on cua.ai VMs for cleaner UI and per-agent isolation vs. defending a heavily tuned lean Hermes.
New providers are serving DeepSeek 4.1 Flash at 280-400+ tps for coding workloads at ~$0.60/M tokens, well above OpenRouter's advertised numbers, with fewer refusals than OpenAI. Inception's Mercury 3 is clocking ~1000 tok/s though early CRO landing page tests missed the brief. A new federal register notice also bans DeepSeek in software sold to the US government.
Seedance 2.5 at ~$2.34 per 30 seconds via fal (and ~23% below Kie/Fal on BytePlus) is being called category-shifting for AI video if it holds up — @iammikevineyard is swapping Blender for it in an Opus 5 + Codex funnel. Sunra.ai was flagged as a potential new Kie competitor. Rumors of OpenAI reviving Sora drew nostalgia from builders who called the shutdown a mistake.
@rmktg, @scalingfrog and others debated delivery channels for team-wide AI access. Grokbot removes compute/storage headaches but can'anonymous be shared; dedicated Hermes servers piped into scoped Telegram or Slack bots drove the highest sustained usage, with 4 power users generating phenomenal output. Alternative: ship an MCP + knowledge base for each teammate's own Grokbot. @rmktg also pushed stripping bot reports to Spend/ROAS/sales only.
Words worth knowing
Three terms from the glossary.
Key Takeaways
- Hermes cache hit rate jumps to 91-98% by stabilizing per-message details; move compaction trigger to ~180K tokens and separate memory (facts to mem0, procedures to skills, never store stale ROAS/spend).
- Gemini 4 Argon in Medium mode via AGY CLI can own an end-to-end infra migration — servers, DNS, observability, SAML, security audits — in a single overnight session.
- Speed is becoming the dominant model-selection axis: DeepSeek 4.1 Flash at 400+ tps and Mercury 3 at ~1000 tok/s are reframing 'good enough intelligence, fast' as the default.
- Seedance 2.5 at ~$2.34 per 30 seconds via fal (and ~23% below Kie on BytePlus) is worth testing as a Blender replacement for AI video pipelines.
- Claude paid plans include $200/mo in API credits tracked separately from subscription limits — easy to miss and effectively free compute.
Hot Threads
Four-week Hermes optimization breakdown and tuning methodology
Gemini 4 Argon overnight infra migration and full Grafana observability stack
Is AI video editing as good as a human yet?