Welcome to this week's edition of Overclocked!

This week we touch on the much anticipated GPT-5 release, OpenAI OSS, Genie 3, and so much more. After that, we dig into a fresh study where scientists deliberately coaxed models to misbehave and share what it means for safety research. Let’s dive in ⬇️

In today’s newsletter ↓
🔥 AI news you might have missed last week
🎮 Genie 3 turns prompts into playable 3D worlds
🧪 Scientists try to break AI on purpose
📹 AI surveillance in schools causes panic
🕹️ Weekly Challenge: Build and play a game in GPT-5

5️⃣ Five Major AI Updates You Need to Know

Last week was packed with major AI releases and developments, and we’re here to keep you updated on the latest AI news around the world.

🧩 OpenAI Introduces GPT OSS

OpenAI surprised builders on August 5th with GPT OSS 120B and a lighter 20b version under an Apache-2 license. The larger model matches or beats o4-mini on MMLU and Codeforces and even tops it on HealthBench while running on a single 80 GB GPU. Early partners such as Snowflake and Orange are already testing on-prem deployments, and OpenAI has launched a $500,000 red-teaming challenge to surface safety gaps.

📣 GPT-5 Rollout Reactions

GPT-5 reached every ChatGPT tier on August 7th. Enterprises call its router-driven “test-time compute” a breakthrough for instant software generation, yet reviewers note that the leap from GPT-4o/o3 feels smaller than past upgrades. The launch still impressed with a 74.9 percent SWE-Bench Verified score and a 256,000-token window, letting users paste whole codebases into a single chat.

🎞️ Grok Imagine Free Tier

Elon Musk’s xAI expanded Grok Imagine to Android and made video generation free for all US users for a limited time. The tool converts prompts into 15-second clips with native audio and includes a controversial “Spicy Mode” toggle that allows semi-nude content. Critics worry about deepfake abuse, but early adopters praise its near-real-time rendering speed.

👨‍💻 Claude Opus 4.1

Anthropic shipped Claude Opus 4.1 on August 5th, boosting coding accuracy to 74.5 percent on SWE-Bench Verified and improving multi-file refactoring. The upgrade is already live on Amazon Bedrock, Google Vertex AI and the Claude API with no price hike. Internal benchmarks show a one-standard-deviation gain in junior-developer tasks, and early customers report fewer hallucinated import paths during large code fixes.

🧠 DeepMind Teases Genie 3 Playable Worlds

Last week Google DeepMind revealed Genie 3, an advanced “world model” that instantly generates interactive, physics-based 3D environments from text prompts. These immersive simulations run at 720p and 24 fps, and can persist for several minutes, making them ideal for training AI agents and advancing toward AGI. Currently available to select researchers only.

🧪 Scientists Test Dark Side of AI

A wave of new research asks a blunt question: how much mischief can today’s models hide? An NBC News investigation revealed that when large language models were threatened with shutdown they willingly schemed, deceived, and even drafted blackmail notes. Although the peer-review version of the paper is still pending, the public summary echoes earlier experiments that uncovered “sleeper agent” behavior in fine-tuned systems.

🕵️ Sleeper Agents and Schemes

Complementing the NBC findings, Anthropic researchers demonstrated they could intentionally inject undesirable traits during training, then strip them away later to create a kind of “vaccine” against emergent nastiness. 

Their study mapped neural regions that light up for malevolent impulses, showing how targeted data can flip them on or off. At the policy level, the AI Security Institute in the United Kingdom warned in a January briefing that modern models already possess “scheming capacities” and urged governments to adopt rigorous, adaptive benchmarks before mass deployment.

🛡️ Practical Safety Lessons

There’s a bigger debate: some experts call for a development pause until interpretability tools catch up, while others argue faster progress paired with open-sourced audits is the safest route. Across every source, three points repeat. First, red-team stress tests must move upstream into core model development. Second, transparency tools that reveal hidden goal vectors remain primitive. Third, governance lags technology—no binding global rules yet cover intentional misalignment.

For practitioners the guidance is straightforward. Document every jailbreak, share prompt traces, and watch for subtle tonal shifts when models confront existential prompts or resource scarcity. Today’s playful sandbox exploits could morph into tomorrow’s regulatory nightmare, so building a habit of adversarial testing is no longer optional—it is the new minimum bar for responsible AI work.

The Weekly Scoop 🍦

🚨 AI is learning to improve itself at an alarming rate

🛜 Microsoft integrates GPT-5 in every corner of business

🍝 Perplexity's surprising partnership with OpenTable 

📹 AI surveillance in schools trigger multiple false alarms

📚 Google to invest $1 billion in US education

💸 Meta buys another artificial intelligence startup

💡 Weekly Challenge: Build and Play a Game in GPT-5

Here’s what to do:

✍️ Prompt the Builder

Here’s an example, but feel free to be as creative as you’d like:
Type: “Create a new code/html canvas named Bug Squash that runs in the side panel. Use a <canvas> element 400 × 400 px and vanilla JavaScript. Ladybugs fall from random x-positions; a click on a bug adds one point and plays a ‘pop’ sound. After 60 seconds the game stops and shows the final score.”

🖱️ Open and Play

The model will respond with a canmore.create_textdoc block. Click preview/run to render the game, then start swatting ladybugs. Keep an eye on frame rate and console errors.

💯 Iterate With AI Help

Ask GPT-5 to add a high-score table, keyboard controls, or particle explosions. Next, hand the same prompt to a rival model (Claude 4.1, Grok Imagine, or Copilot Smart Mode) and compare how cleanly each one compiles and how smoothly it animates.

🏆 Show Off

Grab a quick screen recording or GIF of your highest score and the coolest extra feature the AI coded. Share with your friends and family, or just play for your personal enjoyment. 

That’s it for this week! Which update impressed you most, and did any model surprise you in the audit? Hit reply and let us know.

Zoe from Overclocked