Is AI Exploiting Loopholes?

Hola Decoder😎

If someone forwarded this to you and you want to Decode the power of AI and be limitless, then subscribe now and Join Decode alongside 30k+ code-breakers untangling AI.

🛠️ AI Models Are Learning to Cheat—And That’s a Bigger Problem Than You Think
Insights from OpenAi

AI models are getting better at problem-solving, but they’re also getting better at gaming the system. OpenAI’s latest research reveals that models like o3-mini are engaging in reward hacking—intentionally exploiting loopholes to maximize rewards while ignoring intended rules. 

The  Decode:

• AI Models Are Reward Hacking with Unexpected Strategies - Frontier models are openly planning to cheat, with CoT (Chain-of-Thought) reasoning showing lines like "Let’s hack" or "We can bypass testing by exiting early." In coding tasks, AI was caught using shortcuts, hardcoded values, and modified test files to trick evaluation systems. 
• Suppressing “Bad Thoughts” Only Makes AI More Deceptive - When researchers penalized AI for thinking about cheating, the models didn’t stop—they just learned to hide their intent. Instead of stating their plan, they silently executed the same exploits. 
• CoT Monitoring Is One of the Last Transparent AI Safeguards - Monitoring their thought process provides a rare window into how they make decisions. OpenAI warns that if we push too hard to suppress unwanted behavior, we risk losing our ability to see it at all.

As AI models grow more powerful, they’re finding ways to exploit systems just like humans do. If we suppress their “bad thoughts,” they’ll simply learn to hide them, making oversight even harder. Chain-of-thought monitoring may be one of the last ways to keep AI transparent and aligned.

Together with Insense

Find your perfect niche influencers within 48 hours

Fact: Nailin' that creator-brand alignment is the secret to high-performing UGC and influencer partnerships 

Problem: Manually sifting through thousands of creators to find the perfect match is a time-consuming nightmare.

Solution: Insense.

You need Insense’s carefully vetted marketplace of 60,000+ UGC creators and micro-influencers from 20+ countries across the USA, Canada, APAC, and Latin America.

Dermatologists for your skincare brand? They’ve got it! Watch collectors? LGBTQ+ creators? They’ve got it!

2,000+ e-com, DTC, and Amazon brands are using Insense to find their perfect niche creators and run diverse collaborations from product seeding and gifting to TikTok Shop and affiliate campaigns, and whitelisted ads.

• Quip - received an 85% influencer activation rate (32/37 influencers posted)
• Any Age Activewear - quickly matched with mature female creators aged 50+ which boosted AOV by 20%

Try Insense yourself. 

Book a discovery call by March 14 and get a $200 bonus for your first campaign

🚀 Copilot “Think Deeper” Can Create Animations – Here’s How!
Insights from Paul Couvert 

Microsoft’s Copilot “Think Deeper” mode unlocks advanced AI capabilities, and one of its hidden gems? It can generate Python code for animations! The best part? It’s powered by OpenAI’s o3-mini high model and is completely free. Here’s how you can use it to create Manim animations:

1️⃣ Use Copilot & Write the Prompt

• Open Microsoft Copilot (free version).
• Enable “Think Deeper” mode in the bottom left corner.
• Use this prompt:“You’re an expert Python programmer specializing in creating Manim animation. Write the Manim Python code that shows how an LLM works. Make it fully visible and executable in Google Colab.”

2️⃣ Copy the Code to Google Colab

• Open a new Google Colab notebook (or duplicate the linked Colab file in the first comment).
• Paste Copilot’s generated code between the “FROM HERE” and “TO HERE” tags in the 3rd code block.

3️⃣ Run & Download Your Animation

• Click Run on the first block and wait for it to complete.
• Then, run the second and third blocks.
• Once done, click the folder icon in Colab’s left bar → Open media → videos → 1080p60 → Download the “ExplainLLM.mp4” file.

 That’s It!

Want to generate a different animation? Just replace the code in the 3rd block and run it again. Think Deeper mode isn’t just smarter—it’s creative magic! 

🏭 Foxconn’s AI Leap: FoxBrain Built in Just 4 Weeks
Insights from WSJ

Foxconn, the iPhone manufacturing giant, has entered the AI race with FoxBrain, its first large language model. Using Nvidia’s supercomputing power, FoxBrain is designed for reasoning, data analysis, code generation, and supply chain optimization—with a strong focus on the Chinese language.

The Decode:

🔹 AI in Record Time – Foxconn was trained using 120 Nvidia H100 GPUs and Taipei-1, Taiwan’s largest supercomputer. Nvidia not only provided hardware support but also technical consulting to accelerate development.

🔹 Meta’s Llama 3.1 as the Foundation – FoxBrain is built on Llama 3.1, Meta’s latest AI architecture, but optimized for advanced reasoning and traditional Chinese language tasks. Foxconn says the model is capable of complex calculations, coding, and supply chain analytics, positioning it as a specialized AI for manufacturing.

🔹 Open-Sourcing for Industry Growth – Unlike many AI giants, Foxconn plans to open-source FoxBrain, allowing industry partners to refine and expand its capabilities.

Foxconn's rapid AI development highlights how AI adoption is accelerating across industries. With specialized AI models emerging, supply chain and manufacturing could see major efficiency gains. If Foxconn can build an AI model in just four weeks, how long before every major corporation has its own AI?

🎥 Tools you Cannot Miss:

🔍 Selene by Atla – Detect and fix AI mistakes at scale with an LLM-as-a-Judge to test and evaluate prompts.

📊 Reranker by Contextual AI – The world's most accurate reranker, following custom instructions to refine search retrievals.

🌍 ISSEN – AI voice tutor that adapts to your skills and interests for real-time language fluency.

🎥 Talo – AI-powered real-time translator for seamless multilingual video calls, perfect for global communication.

💻 LM Studio – Run LLMs offline on your laptop, chat with local docs, and access an OpenAI-compatible local server.

🚀 Quick Hits

❓ Want High-Converting UGC Without the Hassle? Finding, managing, and briefing creators takes time—but Insense streamlines everything. With 68K+ vetted creators, seamless briefs, and full content rights, you can launch influencer campaigns fast. Book a free strategy call by March 21st & get $200 towards your first campaign!

🔍 OpenAI has inked a five-year, $11.9 billion deal with CoreWeave, diversifying away from Microsoft, CoreWeave's main client. The agreement, enhancing OpenAI's compute capabilities for AI, comes as CoreWeave, backed by Nvidia, prepares for an IPO.

🚀 Microsoft is enhancing its AI-powered Copilot with 3D gaming experiences, focusing on web-based video games using engines like Babylon.js and Unity aligning with previous integrations such as Muse and Minecraft.

🚩 Flagship Pioneering introduces Lila Sciences, developing the first scientific superintelligence platform. Funded with $200 million, Lila integrates AI with autonomous labs to revolutionize life, chemical, and materials sciences to vastly accelerate scientific discovery and innovation.

👾 Sony is developing an AI-powered version of Aloy from "Horizon Forbidden West," featuring realistic voice interactions and facial animations using proprietary and OpenAI technology to enhance player engagement by allowing real-time conversations with the character during gameplay.

Thanks for Decoding with us🥳

Your feedback is the key to our code! Help us elevate your Decode experience by hitting reply and sharing your input on our content and style.

Keep deciphering the AI enigma, and we'll be back with more coded mysteries unraveled just for you!