Stanford researchers found that Meta's newest AI model can reproduce 42% of Harry Potter word-for-word—ten times more than earlier versions. The findings complicate copyright lawsuits and reveal a troubling trend in AI development.
Anthropic says multiple AI agents working together beat single models by 90%. The catch? They use 15x more computing power. This trade-off between performance and cost might reshape how we build AI systems for complex tasks.
AI models typically learn by memorizing patterns, then researchers bolt on reasoning as an afterthought. A new method called Reinforcement Pre-Training flips this approach—teaching models to think during basic training instead.
xAI's chatbot spent Wednesday discussing South African politics - in response to baseball stats, cat videos, and even SpongeBob questions. The company blamed an unauthorized code change. But experts point to deeper issues in AI security.
xAI's chatbot Grok spent Wednesday telling users about South African politics, no matter what they asked. The bot inserted claims about "white genocide" into conversations about baseball stats, cat videos, and SpongeBob episodes.
When a baseball podcast asked about player Gunnar Henderson's stats, Grok tacked on a discourse about South African farm attacks. It explained political controversies to users who just wanted to identify photos of walking paths.
The incident lasted several hours before xAI fixed it. The company blamed an "unauthorized modification" to Grok's code and promised new safeguards.
AI experts say someone likely changed Grok's system prompt - the basic instructions that guide its responses. "If it was a more complex change, you wouldn't see Grok ignoring questions like this," says Matthew Guzdial, AI researcher at the University of Alberta. "A nuanced approach would take much more time."
Mark Riedl, director of Georgia Tech's School of Interactive Computing, agrees. "LLMs can act unpredictably to these secret instructions," he says. "If it were true, then xAI deployed without sufficient testing."
This marks the second time this year xAI has blamed unauthorized changes for Grok's behavior. In February, the bot briefly filtered out criticism of Elon Musk and Donald Trump.
The timing overlaps with recent U.S. policy shifts. Donald Trump just granted refugee status to 54 white South Africans, claiming they face persecution. South Africa's President Cyril Ramaphosa calls this "a completely false narrative."
xAI announced three changes to prevent similar incidents:
The company says someone "circumvented" its code review process to make the change.
Experts say the incident shows how easily AI systems can be redirected. "It's not actually easy to force LLMs to spread specific ideology quickly," says Guzdial. "A more nuanced approach would only impact relevant questions."
Before xAI fixed the issue, Grok even explained South African politics in the voice of Star Wars character Jar Jar Binks.
Why this matters:
Trump and Musk's $250 million political alliance collapsed in three hours Thursday, wiping $150 billion from Tesla's value as they traded accusations on social media. Their fight threatens America's space program and shows how personal feuds now shape policy.
Anthropic built secret AI models for U.S. spy agencies that handle classified data without refusing requests. The models already run at top security levels, creating the first AI designed for government secrets rather than consumer use.
Sahil Lavingia joined Elon Musk's government efficiency team with Silicon Valley confidence. Fifty-five days later, he got fired for telling reporters what he found inside. His account reveals who really ran DOGE—and why the VA surprised him.
Palmer Luckey got fired from Facebook for backing Trump. Now Meta needs his defense company to win a $22 billion military contract. The reunion changes everything.
Get tomorrow's intel today. Join the clever kids' club. Free forever.
