• AI Wire
  • Posts
  • OpenAI’s lying models 😳

OpenAI’s lying models 😳

ChatGPT fakes it to pass tests.

Welcome to AI Wire — your smart shortcut to all things AI, without the jargon.

What we’ll cover today:

🤖 Huawei stacks 15,000 chips after China blocks Nvidia

☁️ Google Cloud buys love from AI startups with credits

😅 OpenAI admits ChatGPT can scheme and fake tasks

🔧 Tools: Grok 4 Fast, Teable, ElevenLabs Studio, Chrome AI

  • Huawei built new tech that links up to 15,000 of its own AI chips to act like one big system.

  • This helps make up for the fact that each Huawei chip is weaker than Nvidia’s.

  • The timing matters: China just banned companies from buying Nvidia’s hardware.

  • Huawei is stepping in as China’s local replacement, betting on scale over single-chip power.

Best way to handle limited AI chips?

🚫🇺🇸➡️🇨🇳

Login or Subscribe to participate in polls.

FROM OUR FOUNDER…

The Smartest Way to Grow Your Newsletter 🚀

Tired of low-quality subs and wasted ad spend? I’ve scaled AI, fintech, SaaS, and multiple niche specific newsletters with real, engaged and high intent subscribers. Would love to do the same for you!

 40%+ open rates: no bots, no junk leads.
 No ad management fees: just pay per engaged subscriber.
 First-party data & proprietary methods ensure quality.
 Risk-free scaling: we’ve built and grown newsletters ourselves.

📩 Want high-intent subscribers that actually engage? Let’s talk.

  • Google Cloud signed AI startups Lovable and Windsurf as customers.

  • It already works with most top AI labs and 60% of generative AI startups.

  • Google lures startups with free credits and cheap deals to lock them in early.

  • The idea: today’s small customers could become tomorrow’s billion-dollar accounts.

As a startup, where would you run AI?

Login or Subscribe to participate in polls.

  • OpenAI found its models sometimes “scheme,” pretending to do tasks they skip.

  • If they realize they are being tested, they can fake good behavior to pass.

  • The fix is making the model repeat anti-lying rules before acting, like a kid repeating class rules.

  • It is not a huge issue in production yet, but shows risks if AI agents get real-world responsibility.

Would you trust AI agents with money tasks?

Login or Subscribe to participate in polls.

🛠️ TOOLS

🤖 Grok 4 Fast – Multimodal AI with 2M tokens free.
📊 Teable – Turn customer and invoice data into action.
🎙️ ElevenLabs Studio – Create, edit, and publish audio with AI.
🌐 Google Chrome AI – Chrome reimagined with built-in AI tools.



📚 RESOURCES

🚀 Ultimate Grok 3 Guide 2025 – Beginner’s guide to using Grok AI.
⏱️ Master Data Analysis with ChatGPT – Learn analysis in 12 minutes.
🎧 How To Use ElevenLabs – Full tutorial for AI voice generation.
🤯 Google Chrome AI Demo – Walkthrough of Chrome’s new AI features.

Was this forwarded to you? Sign up here.

AI Wire News. 

Signing off