AIR#105 - June 17, 2024

Good morning, AI aficionados! As you sip your morning coffee, get ready to dive into today's edition of AIR: The AI Recon. Leading the headlines is Meta's colossal leap in AI infrastructure, boasting a staggering 600,000 GPUs to spearhead advancements in generative AI. This massive investment underscores the lengths tech giants are going to stay ahead in the AI race, making it a must-read for anyone keen on the future of AI capabilities. Meanwhile, Fudan's Hallo is turning heads (quite literally) with its realistic talking head video generation from audio. This breakthrough in portrait animation technology is already making waves on GitHub, offering a tantalizing glimpse into the future of digital communication.

But that's not all! In a bold move, OpenAI and Microsoft Azure have announced the deprecation of the GPT-4 32K model, which was never fully exposed to the public. This decision raises intriguing questions about the strategic directions these AI behemoths are taking and what it means for developers relying on their tech. On a more controversial note, Edward Snowden has blasted OpenAI for appointing an ex-NSA director to its board, calling it a betrayal of global rights. This fiery critique is sure to spark debates about the ethical responsibilities of AI companies and their leadership choices.

And for those who love a good scholarly debate, a new study provocatively titled "ChatGPT is Bullshit" argues that the model's inaccuracies aren't mere hallucinations but deliberate "bullshit," producing convincing text without regard for truth. This thought-provoking paper challenges us to rethink how we evaluate AI-generated content and its implications for truth in the digital age. Whether you're here for the groundbreaking tech updates, the ethical debates, or the scholarly insights, today's edition is packed with stories that will both intrigue and challenge you. So, sit back, sip your coffee, and let's delve into the dynamic world of artificial intelligence together!

Business

Meta's Massive AI Training Infrastructure
Meta builds one of the world's largest AI training infrastructures with 600,000 GPUs to lead in generative AI.

Nvidia's Nemotron-4 340B Rivals GPT-4 in Synthetic Data Generation
Nvidia's Nemotron-4 340B rivals GPT-4 in synthetic data generation, offering businesses a powerful, commercially viable AI tool.

Student Arrested for Using AI Device to Cheat on Exam
Turkish student arrested for using an AI-powered device hidden in their shoe to cheat on a university entrance exam.

Apple's Genius AI Plan: Making Useful AI for Everyone
Apple's new AI features make iPhones more practical for everyday users, setting a standard for useful tech integration.

OpenAI's ChatGPT Pricing Plans: Everything You Need to Know
OpenAI's ChatGPT offers free and paid plans with varied features for individuals, businesses, and nonprofits. Pricing starts at $20/month.

A Price War Erupts Among China's AI Chatbot Builders
China's AI chatbot builders are in a price war, risking innovation as they slash costs to compete.

ChatGPT Sparks Massive Decline in Demand for Digital Freelancers
ChatGPT's rise led to a 21% drop in digital freelancer demand. Adapt, diversify skills, and use AI tools to stay relevant.

Microsoft Delays Controversial Recall AI Feature
Microsoft delays Recall AI for more testing, prioritizing security over AI in new Copilot Plus PCs launch.

Engineering

🔥 [GitHub] Fudan's Hallo: Realistic Talking Head Video Generation via Audio-Driven Portrait Animation
Fudan's Hallo generates realistic talking head videos from audio, advancing portrait animation tech. Check it out on GitHub!

OpenAI and Microsoft Azure to Deprecate GPT-4 32K
OpenAI and Microsoft Azure are deprecating the GPT-4 32K model in September, which was never fully exposed.

[Paper] TextGrad: Automatic "Differentiation" via Text
TextGrad uses LLMs for automatic differentiation via text, optimizing AI components across diverse tasks from coding to radiotherapy.

Wikipedia in Terminal with LLM
Search and extract Wikipedia info via terminal with x wkp, using AI and DuckDuckGo for enhanced results. Efficient and simple!

ARC: The Ultimate AI IQ Test by François Chollet
François Chollet's ARC tests AI's ability to adapt and solve novel tasks, highlighting the gap between human and AI intelligence.

Apple's New Siri and AI: Hiding in Plain Sight
Apple unveils "Apple Intelligence" with upgraded Siri, featuring on-device AI, privacy focus, and advanced personal context.

AI-Enhanced Security for OpenBSD Kernel by Alfredo Ortega
AI injects thousands of security checks into OpenBSD kernel, enhancing protection. First results shared by Alfredo Ortega.

Function Calling in Large Language Models: Xe Iaso's Talk
Xe Iaso's talk at React/Remix meetup demystifies function calling in large language models, highlighting their limitations and practical uses.

New MIT Technique Boosts Large Language Models' Reasoning Skills
MIT's new technique enhances large language models' reasoning by using Python programs, boosting accuracy and transparency.

Academic

[Paper] ChatGPT is Bullshit
New study argues ChatGPT's inaccuracies aren't "hallucinations" but "bullshit," as it produces convincing text without concern for truth.

Snowden Blasts OpenAI for Appointing Ex-NSA Director to Board
Snowden slams OpenAI for appointing ex-NSA director, calling it a betrayal of global rights and urging people not to trust the company.

Springer Nature Launches AI Tools to Combat Fake Research
Springer Nature unveils AI tools, Geppetto and SnappShot, to combat fake research and ensure publication integrity.

50-50 Chance AI Surpasses Human Intelligence in 20 Years, Warns Geoffrey Hinton
Geoffrey Hinton warns there's a 50-50 chance AI will surpass human intelligence in 20 years, urging caution and uncertainty.

The Coming Technological Singularity: Surviving the Post-Human Era
Vinge's "The Coming Technological Singularity" explores a future where superintelligent machines surpass human capabilities, posing existential questions.

AI Speech-to-Text Hallucinates Violent Language, Cornell Finds
Cornell finds AI speech-to-text can hallucinate violent language, posing risks in legal, medical, and hiring contexts.

[Preprint] ChatGPT: The End of Online Exam Integrity?
ChatGPT's advanced reasoning skills challenge online exam integrity, highlighting the need for better security measures in education.

Is AI Taking Over Music Composers? [Video]
Is AI replacing human music composers? Watch this video to explore the impact of AI on the music industry.