From Hallucinations to Headsets

👋 Hey Reader!

AI isn’t slowing down — it’s hallucinating less, whispering your thoughts, and making billion-dollar cloud deals. From Microsoft hedging its bets to Apple teaching AirPods new tricks, today’s brew is packed. Let’s sip. ☕🤖

👁️ OpenAI Cracks the Code on Hallucinations

TL;DR: OpenAI dropped a paper arguing that AI “hallucinations” happen not because models love lying, but because current training rewards guessing over honesty.

• The problem: Models are trained to predict the next word, so when they don’t know, they still guess. Current evaluation metrics often give them points for lucky guesses, but penalize them for saying “I don’t know.”

• The fix: Rethink how models are graded — rewarding accuracy and humility.

• Why it matters: If AI can admit what it doesn’t know, trust in systems like ChatGPT skyrockets. Fewer “hallucinated birthdays,” more reliable answers.

👉 Bottom line: Honesty might be the best policy… even for AI.

🤫 Alterego’s “Near-Telepathic” Wearable

Arthur C. Clarke once said, “Any sufficiently advanced technology is indistinguishable from magic.” Alterego is leaning into that.

• Boston-based, MIT-born startup Alterego unveiled a headset that detects tiny jaw/throat muscle movements to interpret “silent speech.”

• Founder Arnav Kapur demoed tasks like coding, texting, and even chatting with other wearers — all without speaking a word.

• Works in noisy areas, supports multiple languages, and even detects intent to speak.

🚨 Why it’s wild: This is BCI-level stuff without drilling into your brain. If it works, it could reshape how we interact with AI — and each other.

💰 Microsoft’s $17.4B GPU Grab

Europe’s GPU drought is about to get some relief.

• Microsoft inked a five-year, $17.4B deal with Nebius to roll out 400 MW of AI-ready capacity across Europe.

• Phase 1: Finland, with liquid-cooled Nvidia H100 clusters online by mid-2026.

• Both companies pledge 100% renewable energy by 2028.

• Bonus: Prices for compute could drop 10–15%.

⚡ Why it matters: Cheaper, greener GPUs = faster AI progress. And if you’re an Azure customer, you might want to lock in early-access pricing.

🔄 Microsoft Bets on Anthropic Too

OpenAI may be Microsoft’s BFF, but it’s not the only friend in the sandbox anymore.

• Microsoft signed a multi-year deal to bring Anthropic’s Claude into Azure and Microsoft 365 Copilot.

• Claude Sonnet 4 reportedly matches GPT-4o in accuracy with lower latency.

• Anthropic’s “constitutional AI” makes compliance easier for regulated industries.

📌 Takeaway: By dual-sourcing, Microsoft gains pricing power and flexibility. OpenAI just got some competition… in its own house.

📰 Other News

• 🎧 Apple’s AirPods Pro 3: Soon to feature live translation and heart-rate tracking. Price tag: $249.

• 📊 Claude upgrades: Can now create/edit files like spreadsheets, docs, and decks inside chat.

• 🏜️ UAE’s AI play: Unveiled “K2 Think,” a 32B-parameter open-source reasoning model that rivals larger ones.

• 📉 AI adoption dips: Usage among large U.S. companies fell from 13.5% (June) to 12% (Aug).

• 🖼️ Meta x Black Forest Labs: A $140M partnership to expand AI-generated images.

• 💊 Eli Lilly: Rolled out TuneLab, an AI drug discovery platform trained on $1B worth of proprietary data.

• 🔒 AI-fueled scams: Deepfake fraud is spiking in India, with scams costing just pennies to run.

• 🎬 Veo 3 upgrade: Costs nearly halved, now supports HD + vertical video formats for creators.

🙌 That’s a Wrap

From hallucination fixes to silent speech wearables, AI’s future feels a little more magical today. Thanks for reading Not So Artificial — where AI news is always grounded (and never hallucinated 😉).

Catch you tomorrow,
— The Not So Artificial Team