While the AI industry loves to brag about its capabilities, reality has a way of cutting through the hype. Researchers in Munich, Germany discovered that most AI models start to flounder after just a few thousand words - despite claims of processing millions. It's like catching someone who boasted about their marathon training getting winded after a short jog.

Meanwhile, Perplexity just dropped a truth bomb in the form of pricing: their new research tool, which rivals enterprise-level AI, costs less than your monthly coffee budget. The message is clear: in the AI world, expensive doesn't necessarily mean better anymore.

News Study: Why Your AI Isn't As Smart As You Think

Your AI assistant just told you it can handle a million-word document. A team of researchers from LMU Munich, the Munich Center for Machine Learning, and Adobe Research just proved it's probably lying.

Their new benchmark, aptly named NoLiMa (No Literal Matching), shows what many of us suspected: AI models have been acing tests that are embarrassingly easy. They're like that friend who brags about solving word searches but can't handle a crossword puzzle.

Think your AI is actually reading those massive documents you feed it? Here's a reality check: Most models start stumbling at just 2,000-8,000 words. That's shorter than this article would be if I hadn't kept it brief for your sake. The numbers tell a brutal story: OpenAI's o3-mini crashes to 18.9% accuracy at 32,000 words. DeepSeek's R1, despite its promises, barely manages 20.7%. Even GPT-4o, the supposed champion of comprehension, drops from a near-perfect 99.3% to a humbling 69.7% when texts get longer. And that's one of the good ones – 10 out of 12 top models lose half their accuracy at longer lengths.

The problem? Previous tests were like asking someone to find their name in a phone book – sure, they can do it, but that doesn't mean they understand what a phone book is. NoLiMa changes the game by forcing AI to think beyond simple word matching, more like how humans actually read and understand text.

The tech industry's obsession with longer context windows (some claiming up to 128,000 tokens) might be missing the point entirely. Even the most advanced models fall apart when they have to understand what they're reading, not just match words like a glorified Ctrl+F function. It turns out AI's attention span might be shorter than a goldfish's – and that's not just a clever metaphor.

Why this matters:

• The "long context" revolution isn't just overhyped – it's fundamentally misunderstood
• When AI can't rely on word matching, it struggles with basic comprehension
• Even the most advanced models are more like sophisticated search engines than true readers

Read on, my dear:

• arxiv.org: NOLIMA: Long-Context Evaluation Beyond Literal Matching
• unite.ai: Artificial IntelligenceTop AI Models are Getting Lost in Long Documents
• the decoder: AI language models struggle to connect the dots in long texts, study finds

The Great AI Price War Has Begun

Perplexity just fired a shot heard across the AI industry. Their new Deep Research tool does in minutes what companies pay thousands for monthly. And it costs less than a pizza delivery.

The numbers tell a sobering story. While enterprise AI providers charge four-figure subscriptions, Perplexity offers five free daily queries to anyone. Pro users get 500 queries for $20 monthly. That's about 4 cents per query if you're counting.

But here's the kicker: it's not just cheaper - it's competitive. Deep Research scored 93.9% on the SimpleQA benchmark and outperformed Google's Gemini Thinking. It processes most research tasks in under 3 minutes, analyzing hundreds of sources simultaneously.

Enterprise companies are increasing their AI budgets by 5.7% in 2025. The average increase? $3.4 million. That's a lot of money for capabilities now with a price tag more petite than your monthly streaming subscriptions.

Why this matters:

• The "you get what you pay for" maxim just got turned on its head in the AI world - expensive doesn't necessarily mean better anymore
• The democratization of AI isn't just a catchphrase - it's happening at $20 a month
• Small businesses and individuals locked out of enterprise AI can now access similar capabilities for the price of a casual lunch
• The real innovation might not be in the technology but in making it accessible to everyone who needs it.

Read on, my dear:

• TechCrunch: Perplexity launches its own freemium ‘deep research’ product
• mint: Perplexity takes on ChatGPT and Gemini with new Deep Research AI that completes most tasks in under 3 minutes

AI Photo of the Day

Prompt:
alien with glowing hieroglyphic tattoos. Menacing. Shiny. Clean. Polished.

AI & Tech News

Silicon Valley's Biggest Breakup Plan

Intel's longtime rivals are eyeing different pieces of the American chip giant. Broadcom is exploring Intel's chip-design division, while Taiwan's TSMC has studied taking control of its manufacturing plants. Neither company has submitted formal bids, but the mere possibility of breaking up Intel shows how far the once-dominant chipmaker has fallen from its perch.

Walmart Uses Tech to Stay America's Price Leader

Walmart is deploying AI and automation across its stores to maintain its low-cost advantage. Clerks now use phones to check inventory, while managers can plan shifts in minutes instead of hours. The retail giant aims to combine its vast physical presence - 90% of Americans live within 10 miles of a store - with digital efficiency to outmaneuver both traditional retailers and Amazon.

Chinese Tech Giants Rush to Embrace DeepSeek

Tencent's Weixin, China's largest messaging platform, has begun testing DeepSeek's AI model for user searches, marking a significant shift as tech giants typically prefer their own AI solutions. The move comes as DeepSeek gains momentum, having already surpassed ChatGPT in App Store downloads with its cost-effective AI assistant.

Apple's Vision Pro Gets AI Boost

Apple is adding artificial intelligence features to its $3,500 Vision Pro headset through a software update coming in April, including writing tools, AI-generated emoji, and an image playground app. The update will also make it easier to share the device with friends and family through a new iPhone-controlled guest mode, potentially helping Apple boost interest in its expensive mixed-reality device.

AI Tools: Video Generation & Editing

Video creation is undergoing a seismic shift as AI tools transform anyone with a keyboard into a potential video producer. These tools make creating professional-quality content as simple as typing a blog post.

• Synthesia: Create AI-powered talking head videos with customizable avatars and voices in multiple languages just by typing your script.

• Runway: A professional-grade AI creative suite that generates, edits, and transforms videos with text-to-video and motion-tracking features.

• Lumen5: Transforms blog posts and text content into social media videos using AI to match visuals with your text.

• Pictory: Automatically extracts highlights from long videos and creates short-form content with auto-generated captions and branding.

• InVideo: Template-based video creator with AI-powered features that help quickly convert text to video and create professional-looking content.

• Adobe Premiere Pro (with AI features): Professional video editing software enhanced with AI tools for automated color matching, audio cleanup, and scene editing.

• Veed: Browser-based video editor with AI features for subtitles, translations, and background noise removal.

• Magisto: Simple automated video editor that analyzes your footage and creates polished videos with AI-selected music and transitions.