AIR#21 - March 25, 2024

Good morning, AI aficionados! As you're pouring that first cup of coffee, let's dive into today's edition of AIR: The AI Recon, where we're unpacking the latest feats and challenges in the realm of artificial intelligence. Topping our stories, GPT-4's remarkable achievement of outshining its predecessor without any specialized training has the AI community buzzing. This breakthrough not only marks a significant moment in AI experimentation but also raises intriguing questions about the future capabilities and applications of generative models. It's a clear sign that the march towards more sophisticated and intuitive AI is not just continuing; it's accelerating.

In other news, the specter of data poisoning looms large, casting shadows on machine learning models with its potential to introduce errors and biases. As we delve into the complexities of defending against such attacks, it's evident that the road to secure and reliable AI is fraught with obstacles that require innovative solutions. Meanwhile, Google's Gemini AI is poised to transform iPhones into AI powerhouses, potentially marking a mainstream moment for smartphone technology. This collaboration hints at a future where AI is not just a niche interest but an essential feature of our everyday devices, seamlessly integrated into our digital lives.

And let's not overlook the groundbreaking work by UCLA engineers, who've developed an AI wearable that offers a voice to those without vocal cords. This invention not only showcases the compassionate potential of AI but also reminds us of the technology's power to profoundly impact human lives. As we explore these stories and more, today's edition is a testament to AI's dynamic evolution, from enhancing computational power and securing data integrity to revolutionizing personal technology and offering new hope for inclusivity. So, grab your coffee and let's embark on this journey through the latest and greatest in AI, where innovation meets impact.

Business

Google's Gemini AI May Transform iPhones into AI Powerhouses
Google's Gemini AI could turn iPhones into mainstream AI hubs, marking a pivotal shift from niche to essential smartphone technology.

Google Leads the 2024 AI Race Despite OpenAI Challenge
Google likely to win the 2024 AI race, leveraging YouTube's data and video generation, despite OpenAI's challenge and Microsoft's cautious stance.

Is Apple Really Falling Behind in AI Innovation?
Apple's AI focus on privacy & on-device ML surprises with advanced tasks & cutting-edge research, debunking innovation lag myths.

Andreessen Horowitz Insights: The Evolution of Generative AI in Enterprises
Andreessen Horowitz reveals enterprises are massively upscaling generative AI investment, prioritizing multi-model, open-source approaches for diverse, production-level applications. #AIRevolution in enterprise skyrockets.

Stability AI CEO Emad Mostaque Steps Down Amid Company Turmoil
Emad Mostaque resigns as Stability AI CEO to focus on decentralizing AI, amid company challenges and a mission to keep AI open.

Engineering

Understanding ML Model Data Poisoning: Attacks and Defenses
Data poisoning attacks manipulate ML model training, causing errors and biases. Examples include Microsoft's Tay and art theft. Defenses include input checks and anomaly detection.

Simon Willison's Adventure with Claude and ChatGPT: GeoJSON Sidequests Unlocked
Simon Willison turns GeoJSON quest into a 6-minute AI-powered adventure with Claude and ChatGPT, transforming GIS data into actionable insights.

[Paper] Rose: Revolutionizing Autodiff on the Web with Speed and Extensibility
Rose revolutionizes web autodiff, offering a 173x speed boost over TensorFlow.js, with custom derivative support and dynamic AD functions.

Mistral AI Labs Unveils Mistral 7B v0.2 Base: 32k Context Window, Rope Theta, No Sliding Window
Mistral AI Labs drops Mistral 7B v0.2 Base: a game-changer with a 32k context window, Rope Theta tech, and no sliding window. Learn to fine-tune!

CS Fundamentals: Our Best Bet Against the AI Apocalypse
Diving deep into CS fundamentals, not AI, is our future safeguard. Learning to code remains vital amidst AI's rise. #LearnToCode

[GitHub] MRML Switches to MIT License: A Major Update by jdrouet
MRML project by jdrouet switches to MIT license, opening doors for wider use in software development.

[Paper] PERL: Boosting Reinforcement Learning with Human Feedback Using Low-Rank Adaptation
PERL makes AI training with human feedback faster and less memory-intensive, promising wider adoption for aligning models with human preferences.

[Paper] Terrain Diffusion Network: Generating Realistic Terrains with Climate and Geological Sketches
New AI creates realistic terrains from sketches, factoring in climate and geology for better virtual worlds.

Apple's AI Could End "Hey Siri" Phrase: Exploring Voice-Activated Assistance Without Triggers
Apple explores AI to ditch "Hey Siri" trigger, aiming for seamless voice recognition, sparking privacy concerns.

Academic

🔥 GPT-4 Outperforms $10M GPT-3.5 Model with No Specialized Training
GPT-4 outshines $10M GPT-3.5 with zero specialized training, marking a pivotal AI experiment moment.

[Paper] Enhancing Transformers' Reasoning with Chain of Thought: A Computational Power Analysis
Transformers with "chain of thought" reasoning show potential to solve complex problems previously deemed impossible, advancing AI's computational power.

🔥 UCLA Engineers Develop AI Wearable for Speech Without Vocal Cords
UCLA engineers create AI wearable that lets people without vocal cords speak by translating throat muscle movements into speech.

[Paper] EXPLORER: Neurosymbolic Agent for Text-Based Games on arXiv
EXPLORER, a new AI, excels in text-based games by blending neural exploration with symbolic reasoning, outperforming others in unseen challenges.

🔥 Study Reveals "Emergent" Abilities in AI Develop Predictably, Not Suddenly
Study finds AI "emergent" abilities grow gradually, predictably, challenging sudden leap theory. Measuring methods key to understanding growth.

The Ingenious AI Hoax of Mark V. Shaney: Usenet's Markov Chain Masterpiece
Mark V. Shaney, a fake Usenet user created by algorithms, fooled many with posts generated by Markov chains, showcasing early digital deception.

Stanford Study Debunks Emergence in Large Language Models as a Measurement Mirage
Stanford study claims large language models' "emergent abilities" are predictable, challenging the notion of AI unpredictability.

The Deepfake Porn Crisis: Kids and Celebrities Exploited for Millions of Views
Deepfake porn crisis exploits kids and celebrities, garnering millions of views while victims, including a 15-year-old, fight for legal recourse and tech solutions.

[Paper] Enhancing Radiology with Small Language Models: The RAD-PHI2 Study
RAD-PHI2 study shows small language models can revolutionize radiology, offering precise answers and enhancing workflows.

Risks of Replacing Human Research Participants with AI Explored
Scientists warn against replacing human research participants with AI, citing risks of inaccurate and ethically compromised data.

UC Berkeley Symposium on AI's Role in Climate Innovation
UC Berkeley symposium unites AI and climate science experts to innovate against global warming, aiming for rapid, impactful solutions.