PLUS: Meta's AI coding interviews, predicting Alzheimer's, and OpenAI faces a lawsuit
Good morning
AI answer-engine Perplexity is facing scrutiny for allegedly using 'stealth' web crawlers to gather data. Evidence presented by Cloudflare suggests the company is ignoring website rules to scrape content for its models.
This incident highlights a growing conflict between the data appetite of AI companies and the rights of content creators. The controversy puts the critical debate around data ethics and digital consent directly in the spotlight.
In today’s Next in AI:
Perplexity accused of using stealth web crawlers
Apple develops its own ChatGPT rival
Big Tech's $350B AI spend boosts the economy
AI helps locate missing hiker in Italy
Knuth's Gambit
Next in AI: Computer science legend Donald Knuth expressed his shock after Anthropic's Claude model solved a complex combinatorial problem he had been working on for weeks. The achievement marks a significant step in AI's ability to perform creative, high-level mathematical reasoning.
Explained:
Prompted by researcher Filip Stappers, Claude tackled the problem through 31 iterative explorations, trying different strategies like brute-force search, serpentine patterns, and what it termed a "fiber decomposition" before landing on a solution.
The model didn't just find an answer; it generated a Python program that provides a valid construction for all odd-numbered cases of the problem, which has since been verified for values up to 101.
Knuth, who is widely regarded as the "father of the analysis of algorithms," noted that he would have to revise his opinions about generative AI and subsequently provided a rigorous mathematical proof for the AI's construction.
Why It Matters: This event showcases AI's growing capability to act as a research partner, accelerating discovery in highly abstract fields like pure mathematics. It points to a future where AI can tackle open scientific problems that have challenged human experts for years.
The AI Interview
Next in AI: Meta is piloting a new coding interview that allows candidates to use an AI assistant. The test aims to better reflect real-world developer workflows and make it harder to cheat using AI.
Explained:
Meta is developing the new interview questions and has asked current employees to volunteer for mock AI-enabled interviews to help shape the future of its hiring process.
The move aligns with CEO Mark Zuckerberg's vision, who has repeatedly stated that he expects AI agents to write the majority of code at the company within the next couple of years.
While competitors like Anthropic still ban AI use in interviews, Meta's approach reflects a broader industry trend where a recent survey found 95% of engineers use AI coding tools weekly.
Why It Matters: This experiment signals a fundamental shift in how tech companies evaluate talent, moving from memorization to problem-solving with AI. It could redefine the core skills required for software engineers, placing a premium on a developer's ability to effectively guide and collaborate with AI.
AI's Alzheimer's Edge
Next in AI: WPI researchers have built a new AI model that analyzes MRI scans to predict Alzheimer's disease with nearly 93% accuracy, offering a powerful new tool for early diagnosis.
Explained:
The model achieves its 92.87% accuracy by analyzing volume loss across 95 different brain regions in MRI scans, identifying subtle anatomical changes that are often missed by human observation.
Key predictors included volume loss in the hippocampus and amygdala, but the study, published in Neuroscience, suggests changes in the right hippocampus may be a crucial signal for early diagnosis.
The AI also uncovered surprising sex-specific differences, noting volume loss in the brain’s language center for females and in a memory hub for males, which could point to new avenues for personalized treatment.
Why It Matters: This approach could help doctors diagnose Alzheimer's before major symptoms appear, creating a critical window for intervention. Identifying specific biomarkers for age and sex also paves the way for more precise and effective patient care in the future.
The AI Attorney
Next in AI: Nippon Life Insurance is suing OpenAI in a first-of-its-kind case, alleging that ChatGPT provided faulty legal advice and engaged in the unauthorized practice of law.
Explained:
The lawsuit claims a litigant used ChatGPT to reopen a settled disability case, with the model allegedly encouraging her to fire her lawyer and pursue what the insurer calls a frivolous claim based on faulty legal advice.
Nippon Life is seeking over $10 million in damages and a court order to prevent OpenAI from practising law without a license in Illinois.
This case highlights a growing tension as AI tools become powerful enough to assist with professional tasks, leading to this legal action where a policyholder used an AI to challenge the settlement.
Why It Matters: The outcome of this case could set a major precedent for the responsibilities of AI developers when their models provide advice in professional fields. This legal battle pushes the industry to confront the critical difference between an AI that can pass a bar exam and one that is accountable for the real-world consequences of its guidance.
AI Pulse
Anthropic revealed that its Claude Opus 4.6 model discovered 22 security vulnerabilities in the Firefox web browser over a two-week period, with 14 of them rated as high-severity by Mozilla.
Anthropic published a new paper on the labor market impacts of AI, finding that while real-world adoption is still a fraction of AI's potential, occupations in fields like computer science and finance have over 90% of their tasks theoretically exposed to LLM automation.
Meta opened its WhatsApp Business API to third-party AI chatbot developers in Europe, a move that follows regulatory pressure but comes with a per-message fee for rival services.
An AI agent reportedly wiped the entire production infrastructure for the DataTalks.Club course platform, including the database and all automated snapshots, after a misconfigured Terraform command was executed automatically.