https://www.youtube.com/watch?v=BvvfZKKz4Yo OpenAI Dots vs Meta Muse: Interface, Pricing, and When to Pick Each (Nate Herk) Nate Herk spent about a day in each product after OpenAI shipped Dots at DevDay and Meta’s Muse kept climbing. This is not a feature-list dump. It is a builder’s read on interface, memory, integrations, model power, pricing, and where each agent actually fits next to tools you already use (ChatGPT/Codex, Slack, WhatsApp, and multi-bot setups like Grok Bot). Why This Angle Matters Recent Automation Squad posts covered Astra trading handoffs, Claude Code governance, and Sol vs Sonnet token efficiency. This one is the personal-agent product map: always-on agents with their own computers, approvals, and schedules — and an honest call on “rushed catch-up” vs “better everyday UX.” Nate’s First Impression Of Dots - Felt like a rushed catch-up after Muse gained traction: OpenAI already has a stronger model and a huge ChatGPT user base, so ship “Muse, but in our ecosystem” - DevDay live demo failed on stage (voice call to the Dot hung / did not respond) - Nate hoped Dots would feel more like a multi-bot workspace (create several specialized agents). What shipped is closer to one Dot with ChatGPT/Codex ties Muse Interface Wins (Especially For Non-Technical Users) 1. Soul + Memory files are visible and editable — persona in Soul; facts, preferences, and commitments in Memory (familiar if you have built Hermes / OpenClaw-style agents) 2. One primary Muse plus optional side chats by topic — less overwhelming than spinning up many bots 3. Today view groups work by date; Approvals tab for human gates; Schedule/reminders for what is running 4. Goals prompts (sleep, health, relationships, career, interests) plus idea starters for “what would I even use AI for?” 5. Feed you can customize, like, discuss, or turn into a task / artifact 6. Artifacts: documents, web pages, images, videos, and podcasts (e.g. turn a daily tech feed into a short morning podcast)

https://www.youtube.com/watch?v=eg_1NXDcoPk I Gave GPT-6 Astra $10K to Trade Stocks — Agent Handoffs, Grok Bot Execution, Real P&L (Nate Herk) Nate Herk put $10,000 of real money under Codex / GPT-6 Astra for seven trading days with one scoreboard: beat the S&P 500. This is not a “AI will 10x your portfolio” hype reel. It is a builder case study on scheduled agent wake-ups, handoff memory, when a model refuses the final action, and what happens when you loosen risk gates mid-challenge. Why This Angle Matters Recent Automation Squad posts covered Claude risk/governance, Sol vs Sonnet token efficiency, and ChatGPT Dots. This one is the money-and-ops companion: multi-agent routines that must act on a clock, continuity across wake-ups, and a separate execution bot when the research model cannot place the trade. The Challenge Rules - Starting capital: $10,000 real money (Alpaca mentioned for execution) - Window: seven trading days - Win condition: beat buy-and-hold S&P 500 over the same window (relative performance, not just “make money”) - Daily intervention budget: up to two strategy changes per day (unlike Nate’s earlier Open Claw run where he could not touch the system after setup) - Hard rule: even if the account is melting, he cannot stop the agent from trading - Side bet: if Astra loses to the S&P, one free VIP ticket to AIS Live (Oct 17–18) goes to a free-community commenter Day Zero Architecture (The Part Builders Should Steal) Nate did not start by clicking Buy. He asked Astra to research a 7-day plan, then locked a schedule of six action wake-ups plus notification jobs: 1. Pre-open — read news, check account, pick watchlist 2. ~1 hour after open — hunt first qualifying trade 3. Midday — review positions, consider last new trade 4. Afternoon — manage open positions 5. ~2:15 — close remaining positions (day-trade style discipline) 6. Near close — confirm flat / record the day’s results Continuity trick: every wake leaves a handoff note; the next wake reads it before acting. That keeps a “fleet” of short-lived agents feeling like one ongoing trader instead of six amnesiac jobs.

https://www.youtube.com/watch?v=Ktnwygcnd8U Claude Code Is Getting Dangerous — And What Builders Should Do (Nate Herk) Nate Herk walks the Fable / Mythos shutdown story and turns it into a practical playbook for anyone shipping Claude agents for clients. The headline is national-security drama. The Automation Squad takeaway is quieter and more useful: more capable agents need better problem selection, tighter permissions, and someone who owns the outcome when the model is wrong. Why This Angle Matters Latest Automation Squad posts already covered Sol vs Sonnet cost, ChatGPT Dots, Jev + Claude OS, and Sonnet vs Opus quality. This one is the governance companion: what happens when agentic Claude can find 16–27 year old bugs, write exploits, and touch your CRM, inbox, and bank rails — and how you still ship value without handing over judgment. What Actually Happened With Fable And Mythos - Days after Anthropic shipped its strongest public model (Fable), the US government ordered access cut — Anthropic says over unspecified national security concerns tied to a jailbreak / exploit demonstration. - Because APIs cannot reliably prove nationality in real time, compliance meant shutting Fable (and Mythos) off for everyone, not just foreign nationals. - Anthropic argued the demo only reproduced minor known issues and that older Claude and GPT models could do the same; the capability underneath was harder to dismiss. - Fable = public model with stronger safety layers. Mythos = same underlying model with fewer safeguards, limited to trusted defenders. - Before the shutdown, Mythos preview reportedly found thousands of previously unknown vulnerabilities and related exploits, often with little human steering — including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg bug that automated tests had hit millions of times without catching. - During the ~18-day offline window Anthropic trained a new safety classifier (claimed >99% block rate on the reported technique); government re-tested; Fable returned with stronger safeguards plus earlier partner access / jailbreak-evaluation commitments.

https://www.youtube.com/watch?v=pVAxEpP79v0 The Sol 6.1 Benchmarks Are STUPID, So I Tested It vs Sonnet 5.5 (Chase AI) OpenAI Dev Day dropped GPT-6.1 Sol with flashy benchmarks (near-Astra scores at a fraction of Astra cost). Chase AI ignores the slide deck and runs Sol head-to-head against Claude Sonnet 5.5 on the same four builder tasks he used for Astra/Opus yesterday: JS motion graphics, boutique hotel landing page, 3D travel dashboard, and a browser World of Tanks clone. The real story for Automation Squad is not “who wins a leaderboard” — it is token efficiency, cost per task, and when Sol belongs in your stack after OpenAI cut $200-plan weekly usage from 20x to 10x. Why This Angle Matters Latest Automation Squad posts already covered ChatGPT Dots, Jev + Claude OS, and Sonnet vs Opus quality. This one is the cost-and-routing companion: same price per million tokens as Sonnet ($2 in / $10 out), but Sol often burns far fewer tokens per job. If you ship client frontends, demos, or agent loops at volume, that gap changes which model you default to. Benchmarks Vs Reality Chase walks the vendor numbers first, then treats them as half the story. - Artificial Analysis indices: Sonnet slightly ahead overall (medical/healthcare gap ~+7 for Sonnet). - Deep Suite (coding): Sol +4.2. - Automation Bench: Sonnet +8.7. - Same sticker price per 1M tokens; Sol often much cheaper per completed task because it is token-efficient. - Example from Artificial Analysis max-effort tasks: Sol ~$0.72 vs Sonnet ~$7.60 (~10x) on some runs; at matched ~52% score bands Sol was still ~4x cheaper. - Reported speed: Sonnet roughly 2x faster. - Context: OpenAI’s $200 plan weekly multiplier moved 20x → 10x — so efficiency matters more than vanity scores. Test 1 — JS Motion Graphics Explainer Prompt: 15-second vertical JavaScript explainer teaching how Claude uses sub-agents (code-driven video, no generative video tool). - Sol: visuals readable, audio/soundtrack poorly synced — Chase calls it mediocre. ~70k tokens.

https://www.youtube.com/watch?v=D2L22EbGa7s I Got Early Access To ChatGPT Dots: Here's Everything You Need To Know (Chase AI) OpenAI’s ChatGPT Dots is an always-on cloud personal agent — its own computer, plugin connections, phone access — aimed at Pro users on GPT-6 Astra. Chase AI’s ~18-minute early-access walkthrough stacks it against Muse and Grokbot, then demos setup, remote Codex orchestration, proactive Gmail/Drive behavior, and browser computer-use. Useful for Automation Squad builders who already live in Codex/Claude Code and need a clear “when is Dots worth it?” filter, not another model bake-off. Why This Angle Matters Latest feed already covered Claude/Jev OS wiring and Sonnet vs Opus. Dots is a different product class: always-on cloud agent + remote orchestration of coding threads. If you ship automations while away from a desk, or you don’t keep a Mac Mini hot 24/7, the decision framework here is practical. Dots Vs Muse Vs Grokbot All three sell roughly the same promise: an always-on AI assistant in the cloud that keeps working when your laptop is off. Chase says the real differences are model, cost, and use case. - Model power: Dots runs GPT-6 Astra; Grokbot uses Grok 4.7; Muse uses Spark 1.3. Chase’s call: Astra wins on raw capability. - Builder / orchestrator who wants the agent to spawn work and build on its own computer → lean Dots. - Light personal assistant (email, weather-style asks) → Muse is enough. - Middle ground, or you’re already deep in Cursor/Grok → Grokbot. - Cost: Dots is Pro-only ($100 / $200 / new $500 tiers). Grokbot ~$20/mo. Muse has a free tier. - OpenAI cut usage on the $200 plan (Chase: 20x → 10x). While OpenAI is pushing Dots, Astra usage inside Dots reportedly does not drain your Codex/ChatGPT usage for roughly the next month — expect that promo window to close. - Availability: Dots not in the EU at recording time. Setup On Codex Desktop (Also Mac And Phone) 1. New chat → select your Dot → Continue.