Superpower Daily
Posts
OpenA is preparing for $1tn IPO

OpenA is preparing for $1tn IPO

AI now masters six-player poker

Saeed Ezzati
November 03, 2025

In today’s email:

🧾 ‘Do not trust your eyes’: AI generates surge in expense fraud
🔐 OpenAI unveils 'Aardvark,' a GPT-5-powered agent for autonomous cybersecurity research
🩻 How a family used AI chatbot to slash down a hospital bill
🧰 13 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

OpenA is preparing for $1tn IPO

Key Takeaway: OpenAI is laying the groundwork for a potential late-2026 to 2027 IPO that could value it up to $1T, following a restructuring that aligns its nonprofit overseer with financial upside and reduces reliance on Microsoft.

More Insights:

Timeline: filing as soon as H2 2026; CFO Sarah Friar has floated 2027; still early and subject to change.
Fundraising talk: at least $60B; valuation “up to $1T,” versus a current ~$500B private mark.
Governance shift: OpenAI Foundation now holds 26% plus a warrant tied to milestones.
Business picture: ~$20B year-end revenue run rate but rising losses; Altman says going public is “most likely” to meet capital needs.
Market backdrop: CoreWeave’s IPO pop and Nvidia’s $5T cap signal investors’ strong AI risk appetite.

Why it matters: If a nonprofit-controlled AI leader lists at a trillion-dollar mark, it could reset how frontier AI is financed, give OpenAI a powerful stock currency for massive infra bets and M&A, and test whether public markets will underwrite AGI ambitions at unprecedented scale.

Learn AI tools, agents & automations in just 16 hours (Black Friday Offer)

Join Outskill's LIVE 2-day AI Mastermind - 16 hours of intense training on AI tools, automations & building agents to help you work smarter, earn more, and reclaim your time. In 16 hours & 5 sessions, you will learn:

✅ Build AI agents that save up to 20+ hours weekly
✅ Master 10+ AI tools
✅ Automate 80% of your workload, And Much more.

🎁 Unlock $5000+ in AI bonuses: prompt bibles 📚, roadmap to monetize AI 💰 and your personalised AI toolkit builder ⚙️ — all free when you attend!

AI now masters six-player poker

Key Takeaway: A self-play–trained AI called Pluribus achieved superhuman performance in six-player no-limit Texas hold’em, consistently outperforming elite pros over 10,000 hands.

More Insights:

Multiplayer poker is far tougher than two-player games because Nash-equilibrium guarantees break down and “unbeatable” strategies don’t neatly exist.
Pluribus learns from scratch via self-play to build a full-game “blueprint” strategy, then refines decisions with real-time search during play.
The strategy is fixed (non-adaptive to opponents), avoiding exploitable shifts while still winning reliably.
In tests, Pluribus beat five elite pros at once and also dominated when five copies of itself faced a single pro.
The work shows strong, practical play without relying on provable convergence to equilibrium in multi-player settings.

Why it matters: Multiplayer, imperfect-information reasoning is the norm in the real world—auctions, security, and markets—not the exception; Pluribus demonstrates that scalable self-play plus targeted search can deliver decisive results even where our cleanest game-theory guarantees don’t apply, hinting at powerful AI tools for messy, multi-agent problems beyond games.

‘Do not trust your eyes’: AI generates surge in expense fraud

Key Takeaway: Employees are using powerful image-generation tools to create ultra-realistic fake receipts at scale, forcing finance teams to fight back with AI-driven detection.

More Insights:

AppZen says AI-made receipts were ~14% of fraudulent submissions in September (vs. none last year); Ramp flagged $1M+ in bogus invoices in 90 days.
Platforms report a spike after OpenAI’s GPT-4o image upgrades in March; Google’s models are also cited as enablers.
Fakes now mimic real artifacts—paper wrinkles, itemized menus, even signatures—prompting SAP Concur to warn: “do not trust your eyes.”
Simple metadata checks help but are easily defeated by screenshots; tools now look for contextual anomalies (repeated server names, timing patterns, trip details).
Surveys: ~30% of US/UK finance pros saw more falsified receipts post-GPT-4o; ~70% of CFOs suspect attempts in their orgs and ~10% are certain.

Why it matters: Visual authenticity no longer guarantees truth, so controls must shift from eyeballing images to verifying cryptographic/transactional evidence and behavioral signals—or expense fraud will quietly compound at machine speed.