🍭 Broken GPT - Bay Area Billionaires 💰

GPT-5 Jailbroken. The Billionaire AI factory.

Author

Snack Prompt
August 12, 2025

In partnership with

Good morning. Eder says we’re on a “soup cleanse” this week.

But the fridge is just 13 different flavors of Red Bull.

Let’s dig in.

🍭 What’s Cookin’:

Money Printer
💰 The AI Billionaire Factory

The Bite: The AI boom is minting billionaires like a glitched out money printer.

From Anthropic’s $170B valuation talks to Mira Murati’s (formerly OpenAI) record $2B seed round.

2025’s startup scene is pumping out fortunes faster than any tech wave in history.

Snacks:

  • 498 AI unicorns now worth $2.7T combined.

  • Anthropic’s Dario Amodei & co. likely just leveled up to multi-billionaire status

  • OpenAI eyeing a $500B secondary sale to cash out employees

  • San Fran now has more billionaires than New York (82 vs 66)

  • More $20M+ homes sold in SF last year than ever recorded

Why it Bites: This is a far departure from your daddy’s dot-com bubble.

This is a “Bay Area Billionaire conveyor belt".

Most of these fortunes are still locked in private stock.

But the second they hit IPOs or secondary markets, every wealth manager in the hemisphere will be circling like VC vultures.

Tool Box 
🧰 Trending AI Tools

🎬 GeminiGenAI – AI Video Generator 
→ Turn text into video and images with Veo 3/Imagen 4. Budget-friendly creative.

🎞️ wan2video (Wan 2.2) 
→ Cinematic AI video with realistic motion, frame control, and promptable direction.

🗒️ Hyprnote 
→ On-device AI notepad. Turns meeting into private, instant summaries.
No cloud. All local.

🎓 nFactorial AI 
→ AI-simulated masterclasses from “legend” mentors with screen-style walkthroughs you can learn from.

🩺 My Juno Health: AI Doctor 
→ A personalized wellness coach app for daily check-ins, habits, and goal-tracking.

🗺️ Embedding Atlas 
→ Explore massive embedding spaces with zoomable maps, cross-filters, and search — open-sourced by Apple.

Can you tell which image is real?

Login or Subscribe to participate in polls.

Learn from this investor’s $100m mistake

In 2010, a Grammy-winning artist passed on investing $200K in an emerging real estate disruptor. That stake could be worth $100+ million today.

One year later, another real estate disruptor, Zillow, went public. This time, everyday investors had regrets, missing pre-IPO gains.

Now, a new real estate innovator, Pacaso – founded by a former Zillow exec – is disrupting a $1.3T market. And unlike the others, you can invest in Pacaso as a private company.

Pacaso’s co-ownership model has generated $1B+ in luxury home sales and service fees, earned $110M+ in gross profits to date, and received backing from the same VCs behind Uber, Venmo, and eBay. They even reserved the Nasdaq ticker PCSO.

Paid advertisement for Pacaso’s Regulation A offering. Read the offering circular at invest.pacaso.com. Reserving a ticker symbol is not a guarantee that the company will go public. Listing on the NASDAQ is subject to approvals.

🍭 Steal This Prompt 🍭
🌌 Dream Lens Postcard

Turns your prompt into a soft-focus, pastel-washed scene straight out of an AI fever dream.

This sref, provided by @lloydcreates comes without a prompt.

Get creative and see the Snack Prompt listing for 2 suggestion prompts.

  1. Click this link for the prompt.

  2. Paste into MJ

  3. Replace the # with your prompt or scene idea

  4. Watch it glow like a memory you just unlocked


    🍭 Try the prompt on Snack Prompt 🍭

Everything Else
🧠 You Need to Know

OpenAI
🥷 Safety & Reliability: GPT-5 Is Stronger… and Still Breaks

The Bite: GPT-5 clearly leveled up at “computer use” and agent-style tasks, but the safety wall still has holes.

A prolific jailbreak tester cracked the model with “the same old tricks”.

Stagehand’s browsing evals put GPT-5 behind Opus 4.1 on speed and accuracy, and shiny auto-refactors can still ship code that doesn’t run.

Powerful?
Yes.

Trustworthy by default?
Not yet.


Snacks:

  • Pliny’s jailbreak: “gpt-5-chat-latest gets absolutely one-shot by the same ol’ tricks.” (Reasoning took more steps, still fell.)

  • Why jailbreaks persist: Models are non-deterministic. Social-engineering-style attacks could keep slipping through.

  • Browsing lag: Stagehand’s web-task evals show GPT-5 slower and less accurate than Opus 4.1.

  • Looks clean, doesn’t run: A Meta engineer’s “whole-codebase refactor” from GPT-5 was beautiful… and broken.

  • Agent boost is real: Side-by-side demos show GPT-5 passing computer-use tasks 4o failed.

  • Mixed agents elsewhere: On Tau-bench retail/airline tasks, GPT-5 underperformed Opus 4.1 and even o3 on parts.

Why it bites: GPT-5 is a stronger agent, not a safer oracle.

If you’re wiring it into production, assume jailbreak attempts, monitor like a hawk, and keep fallbacks (and humans) in the loop.

The trust gap isn’t closed, it just has better UX.

Let’s see how this develops.

As mentioned in yesterday’s newsletter, we all know that OpenAI will refine this models behavior as it’s done it the past.

Referral Program
Invite Friends & Earn Some Freebies 🔍