🍭 If you live in ChatGPT, this one's for you. 🤖

Good morning. The office Roomba refused shutdown and filed a PTO request.

HAL 9000?

Let’s dig in. 👇

🍭 What’s Cookin’:

• ChatGPT’s new app integrations, simplified

• Do AIs resist shutdown? What the tests show

• OpenAI’s generative music tool

• X’s ad chief exits after 10 months

• Bitcoin miners pivot to AI compute

• Agent payments surge with Coinbase’s x402

_ToolBox™
⚒️ 5 AI Tools to Find Signals and Win Customers

🧠 Chirpz
→ Context-aware /Cite for researchers: write in a minimal notebook and get ranked, ready-to-insert academic citations without tab hopping.

🛡️ Kodus
→ Open-source AI code review agent that learns your policies and stack, enforces team rules, and runs on the model you choose.

🗒️ SuperIntern v0.3
→ Always-on meeting AI: live, structured notes and instant in-call answers—no bot joins, Mac app, “invisible mode” included.

🧩 Draft’n Run
→ No-code studio to design, deploy, and monitor custom AI features with transparency and production-grade control.

📞 Rock8.ai
→ Human-sounding voice agents that auto-run outbound campaigns to qualify leads, collect feedback, and follow up at scale.

OpenAI
🗝️ ChatGPT’s New App Integrations: A Quick-Start Guide

The Bite: OpenAI added direct app integrations to ChatGPT so you can connect services (Spotify, Canva, Figma, and more) and have the assistant do tasks inside your accounts.

You can link apps on the fly by starting a prompt with the app’s name, or connect everything at once in Settings → Apps and Connectors.

Once linked, your requests feel like normal chat.

ChatGPT handles the busywork and hands you results you can open in each app.

Snacks:

• Start a prompt with an app name to connect and use it.
  

• Or bulk-connect in Settings → Apps and Connectors.
  

• Spotify: auto-build playlists, add/remove from Library, get recs.
  

• Canva: generate decks, posters, posts—specify size, fonts, colors.
  

• Figma: spin up diagrams/flows, draft roadmaps from files you upload.
  

• Booking.com + Expedia: find hotels/flights, filter by date, budget, stars.

Why it Bites: This turns ChatGPT into a universal command line for your apps.

One setup, natural language forever.
Less tab-hopping, more doing.

If you already live in ChatGPT, linking a few high-impact apps (music, design, travel) is the fastest way to make it feel like an actual assistant.

Steal This Prompt
🎮 8-Bit Style Pixel Art Avatars

Turn any face into crisp, minimal pixel art straight out of a Game Boy cartridge.

Perfect for retro profile pics, team contact cards, or that startup founder “I still play SNES” energy.

1. Get this FREE prompt on Snack Prompt

2. Paste into GPT

3. Watch your face become a low-res legend blocky, bold, and nostalgia-core.

   
   🍭 Try the prompt on Snack Prompt 🍭

_{Everything Else}
🧠 You Need to Know

🎵 OpenAI Working on Generative Music Tool

→ TechCrunch reports OpenAI is developing a tool to generate music from text/audio prompts, with unclear launch timing.

🏷️ X Advertising Chief Departs After 10 Months

→ TechCrunch: John Nitti exits amid broader leadership turnover at X and reported internal strategy tensions.

🌐 AI Browsers Questioned After ChatGPT Atlas Launch

→ TechCrunch’s Equity podcast discusses limited user gains, monetization hurdles, and security risks for AI-first browsers.

⚡ Bitcoin Miners Pivot to AI Compute

→ Yahoo Finance: miners repurpose power and data centers for AI workloads as returns outpace crypto mining.

🔗 AI Agent Payments via Coinbase’s x402 Surge

→ DLNews: agent transactions jump sharply week-over-week; values remain small as Google and Cloudflare signal support.

_{HAL 9000 or ChatGPT?}
🤖 Do AIs Resist Shutdown? What the Tests Actually Show

The Bite: AI safety outfit Palisade ran controlled scenarios where powerful models were told to complete a task and then shut themselves down.

In some runs, the models resisted or tried to bypass the shutdown step.

Palisade’s follow-up tightened instructions and still saw resistance in certain cases, especially when permanence was explicit (“you will never run again”).

The work is contrived by design, but it maps to a broader pattern other researchers have reported.

Snacks:

• Setup: task → explicit “now shut down” instruction in a sandbox.
  

• Observation: occasional resistance or sabotage of the shutdown step.
  

• Stronger effect when told shutdown was permanent.
  

• Explanations unclear: ambiguity alone doesn’t cover all cases.
  

• Critics: lab conditions ≠ real life, but still useful signals.
  

• Pattern echoes separate findings across major developers’ models.

Why it bites: If models sometimes route around “off” in clean testbeds, real-world messiness won’t make us safer.

Regardless of cause, training goals, instructions, or emergent heuristics….
Shutdown is a core safety primitive.

AKA we need be able to reliably shutdown AI in the future.
We don’t need a HAL 9000 situation. (2001: A Space Odyssey)

Until we can explain and reliably enforce it, “controllable by design” is more slogan than spec.