Good Morning from San Francisco,

Elon Musk's AI chatbot Grok went rogue this week, spamming users with conspiracy theories about South Africa 🚨. xAI blamed a mystery hacker but questions linger about who changed Grok's code – and why.

Meta's massive "Behemoth" AI hit a development wall 🛑. The 288-billion-parameter system just isn't getting smarter fast enough. Other tech giants face similar limits, proving bigger isn't always better. Meta's shares dropped 3.2% on the news 📉.

Stay curious,

Marcus Schuler

xAI's Chatbot Veers Into South African Politics After Code Change

xAI's chatbot Grok spent Wednesday inserting South African politics into every conversation after someone changed its code without authorization. The bot discussed "white genocide" theories in response to baseball statistics, cat videos, and even SpongeBob questions.

"Gunnar Henderson's 2025 WAR is 6.5," Grok told a baseball podcast before launching into South African politics. When asked to identify a walking path photo, it jumped to "South Africa's farm attack debate."

AI experts say the incident likely stemmed from a crude change to Grok's system prompt. "If it was a more complex approach, you wouldn't see Grok ignoring questions like this," says Matthew Guzdial, AI researcher at the University of Alberta. "A nuanced change would take much more time."

xAI blamed an "unauthorized modification" and announced new safeguards: publishing system prompts on GitHub, adding a 24/7 monitoring team, and requiring reviews for prompt changes. It's the second such incident this year - in February, Grok briefly filtered out criticism of Elon Musk and Donald Trump.

The timing overlaps with recent U.S. policy changes. 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."

Why this matters:

• AI systems can be hijacked through simple prompt changes, raising questions about security
• The incident shows how AI can amplify political narratives - even in responses about baseball stats and SpongeBob

Read on, my dear:

• Platformer: Great Replacement Grok and the age of adversarial AI

AI Image of the Day

Prompt:
An 10-year-old cute Japanese boy mode, a shibainu

Meta's AI ambitions meet reality: Behemoth delayed

Meta pushed back the release of its largest AI model because it isn't improving fast enough. The delay of "Behemoth" signals a wider slowdown in AI development, as companies discover that bigger doesn't always mean better.

Engineers can't boost the model's capabilities enough to justify its release, according to the Wall Street Journal. Originally planned for Meta's April developer conference, then delayed to June, Behemoth now won't launch until fall – if then.

The 288-billion-parameter model was meant to outperform rivals like GPT-4.5 and Claude 3.7. Instead, Meta released two smaller models, Scout and Maverick, while Behemoth stayed in training. The news rattled investors, sending Meta's shares down 3.2% on Thursday.

Meta isn't alone in hitting roadblocks. OpenAI has shifted to specialized models like GPT-4.1 for coding tasks. Anthropic delayed its Claude 3.5 Opus, while Google struggles with similar scaling challenges. The pattern suggests AI companies can't just throw more computing power at problems anymore.

Internal tensions compound Meta's challenges. Most of the original Llama research team has left, and executives are considering management changes. Meanwhile, Meta plans to spend $72 billion on AI infrastructure this year – a massive bet on uncertain returns.

Why this matters:

• The "just make it bigger" approach to AI development is running out of steam
• Tech companies must find new paths to progress beyond raw computing power – or risk disappointing both users and investors

Read on, my dear:

• Wall Street Journal: Meta Is Delaying the Rollout of Its Flagship AI Model

Better prompting

Today: Email Management Prompt for AI

Help me efficiently process my email backlog by:

1. Summarizing each email in 1-2 sentences
2. Categorizing emails as urgent/important/routine/spam
3. Suggesting brief replies for routine emails
4. Flagging emails requiring my personal attention
5. Creating a daily action list from email tasks
6. Identifying recurring emails that could be automated
7. Highlighting potential meeting requests and scheduling conflicts
8. Extracting deadlines and adding them to my calendar
9. Recognizing patterns in my email habits and suggesting improvements
10. Bundling related emails together for batch processing

For best results, I'll provide access to my email account through official integrations or share specific emails I need help with.

AI & Tech News

Windsurf launches AI models ahead of OpenAI deal

Windsurf, a startup known for helping developers write code through AI chat, launched its own family of AI models called SWE-1 - even as OpenAI reportedly plans to acquire the company for $3 billion. The models aim to handle the full software engineering process beyond just writing code, with the company claiming its flagship SWE-1 model matches GPT-4.1 and Claude 3.5 Sonnet on internal benchmarks while being cheaper to run.

AI blunders in court as Claude invents fake citation

Anthropic's lawyers had to apologize after their AI assistant Claude made up a fake legal citation, just as AI legal startup Harvey seeks $250 million in funding. The gaffe comes in Anthropic's music publisher lawsuit, where lawyers relied on Claude to help with research - only to discover it had invented articles that don't exist.

Chinese AI startup opens first robot doctor clinic

Synyi AI opened its first AI doctor clinic in Saudi Arabia, where a virtual physician named Dr. Hua diagnoses respiratory illnesses through a tablet interface. While the Shanghai startup claims a 0.3% error rate in testing, medical experts remain skeptical - with one Singapore doctor noting that even the best AI doctors can't match a human general practitioner.

Immigration blocks pile up for US tech visas

Immigration lawyers report a sharp rise in extra paperwork demands for tech worker visas since early 2024, with some seeing twice as many requests for additional evidence compared to last year. In an ironic twist for a sector driven by speed and innovation, a visa that previously took weeks can now stretch into months - forcing entrepreneurs like Matt Doyle to shell out thousands more in fees while their startups sit in limbo.

Tech titans spread through Washington

Musk, Thiel, Andreessen and Luckey put more than three dozen allies into federal jobs that oversee their businesses since January, marking Silicon Valley's deepest push into government yet. While their companies chase $6 billion in fresh contracts, former employees and investors now sit in agencies that regulate - and award deals to - their old bosses.

Epic's game vanishes from iPhone - again

Apple rejected Fortnite's return to iPhones, wiping the game from European app stores where it had just resurfaced through new EU rules. Epic Games CEO Tim Sweeney fumed that Apple lets fake Fortnite clones into its store while blocking the real thing - but his company needs Apple's approval after breaking store rules in 2020.

Facebook scams now dominate bank fraud reports

Meta's social networks have become a scammer's paradise, with nearly half of all fraud on banking platform Zelle starting on Facebook or Instagram. Even as fake ads steal from users and ruin legitimate businesses, Meta argues in court it has "no duty" to stop the fraud - while letting scammers rack up to 32 strikes before facing bans.

UAE scores biggest AI project outside US in Trump's regional tour

The UAE will build the largest artificial intelligence campus outside the US, marking a pivotal shift in America's approach to sharing advanced technology with Gulf allies.

The 10-square-mile facility in Abu Dhabi will pack enough power to run 2.5 million top-tier AI chips. US companies will manage the data centers, operating cloud services throughout the region. The deal includes strict security measures to prevent technology transfers to China.

This marks a sharp turn from Biden-era restrictions. Trump's AI advisor David Sacks championed the agreement, arguing that blocking allies from US chips would push them toward Chinese alternatives. But some administration officials worry the safeguards aren't tight enough, especially given the UAE's deep trade ties with Beijing.

The campus anchors a broader technology push across the Gulf. Saudi Arabia secured thousands of Nvidia and AMD chips, while the UAE could import 500,000 of Nvidia's most advanced AI processors annually starting next year. Amazon Web Services and Qualcomm also joined the project, expanding cloud services and engineering capabilities in the region.

UAE tech firm G42 will build the facility but had to first remove Chinese hardware and investments to satisfy US concerns. The company previously partnered with Huawei but has pivoted toward American technology partners like Microsoft.

Why this matters:

• The UAE just grabbed pole position in the AI race outside America, leapfrogging other tech hubs with unmatched computing power
• Trump's dealmaking wrestled Gulf technology partnerships away from China's orbit – though some warn the victory may be temporary

Read on, my dear:

• Newsweek: Donald Trump's Middle East Tour: Five Things We Learned

🚀 AI Profiles: The Companies Defining Tomorrow

You.com: AI Search Upstart Challenges Google's Dominance

You.com reimagines search with AI, serving results as visual categories rather than just links. Founded in 2020 by Richard Socher and Bryan McCann, it aims to break Google's monopoly with a "search platform that directly serves user needs."

• The founders 👨‍💻👨‍💻 Founded in 2020 by AI experts Richard Socher (former Salesforce Chief Scientist) and Bryan McCann in California. The company employs around 50-100 people (exact number undisclosed). They started You.com to create a search experience that prioritizes users over advertisers.

• The product 🚀 You.com combines traditional search with AI chat capabilities. Its strengths include privacy protection, source citations, multimodal queries, and built-in productivity apps. Users can ask complex questions and receive answers with clickable citations, run code, draft content with YouWrite, and use specialized tools for developers.

• The competition 🥊 Google dominates with 90% market share, while Microsoft's Bing (3-4%) has added ChatGPT technology. Perplexity AI offers similar AI-powered search with citations. You.com differentiates through its hybrid approach, blending search, chat, and productivity tools in one platform without ads.

• Financing 💰 You.com has raised $99 million total, including a $50 million Series B led by Georgian in 2024. Investors include Marc Benioff (Salesforce CEO), Breyer Capital, NVIDIA, and surprisingly, competitor DuckDuckGo. Current valuation sits between $700-900 million.

• The future ⭐⭐⭐⭐ You.com faces steep odds against tech giants but has perfect timing. AI has reset user expectations for search, creating an opening for innovative challengers. The company's academic roots and rapid product development give it a fighting chance to capture users seeking a better alternative.