AIR#56 - April 29, 2024
Good morning, AI aficionados! Dive into today's edition of AIR: The AI Recon, where the digital realm buzzes with groundbreaking advancements and eyebrow-raising controversies that are reshaping the landscape of artificial intelligence. Salesforce and collaborators have just lifted the veil on OSWorld, a new benchmark for testing AI's prowess in navigating the complex world of real computer tasks. With 369 diverse challenges laid out, this initiative shines a light on the existing gaps in AI's operational knowledge, pushing us to rethink how we train our digital counterparts to interact with the digital environment more effectively.
In the realm of fine-tuning large models, LoRA+ emerges as a game-changer, boasting a dual learning rate strategy that not only doubles the speed but also enhances performance by 1-2%. This leap forward is not just about efficiency; it's about unlocking new potentials in AI's learning capabilities, setting a new bar for what's possible in model optimization. Meanwhile, the GitHub universe is abuzz with the release of Ollama v0.1.33, introducing not one but three heavyweight modelsβLlama 3, Phi 3 Mini, and Qwen 110B. This update isn't just a routine rollout; it's a testament to the relentless pace of innovation in AI, offering new tools and features that promise to revolutionize how we approach parallel processing and model deployment.
Yet, as we marvel at these technological feats, the AI world is not without its dramas. The self-citation scandal involving AI expert Juan Manuel Corchado serves as a cautionary tale about the ethical pitfalls in the pursuit of scientific impact. This controversy, alongside Meta's AI-induced challenges turning apps like Facebook and Instagram into bewildering platforms, underscores the complex interplay between innovation, ethics, and usability in the digital age. So, as you sip your morning coffee, let these stories fuel your curiosity and spark lively debates on the future of AI, reminding us that the journey of discovery is as much about navigating challenges as it is about celebrating breakthroughs.
Business
Meta's AI Overload: A Path to Unusable Apps
Meta's AI integration turns apps like Facebook & Instagram into spammy, unusable platforms, confusing users & diluting personal content.
π₯ Urgent: Stop California's SB 1047 Bill Threatening Open-Source AI and Tech Industry
Urgent call to action: Stop SB 1047 in California, a bill that threatens open-source AI and tech with draconian measures. Act now!
BBC Presenter's Image Misused in Ad Due to AI Voice Scam
BBC presenter Liz Bonnin's image used in scam ad after company fooled by AI-generated voice impersonation. A cautionary tale of AI's dark side.
Meta's AI Ad Tools: A Costly Misfire for Marketers
Meta's AI ad platform overcharges and underdelivers, draining budgets and frustrating marketers with lack of transparency and support.
Bank CIO Advocates for Critical Thinkers Over AI Experts in the Workforce
Esquire Bank's CIO says forget AI experts; critical thinkers who challenge AI are the real need in the workforce.
Pentagon's Quest for Soldier-Robot Trust: Engineering the Future of Military Teams
Pentagon aims for trust between soldiers & robots, exploring emotional bonds & engineering trust for future military teams.
Elon Musk in Beijing to Negotiate Full Self-Driving Activation for Tesla in China
Elon Musk in Beijing to negotiate enabling Tesla's Full Self-Driving in China, aiming for data transfer abroad amid safety and competition concerns.
Athletic Director Uses AI to Create Fake Racist Audio of Principal, Faces Charges
Athletic director charged for using AI to fake racist audio of principal, sparking legal and ethical AI use debate.
Gurman: iPad Pro's Next Leap with M4 Chip's AI Capabilities
Next-gen iPad Pro may leap to M4 chip, boasting AI enhancements, skipping M3. Expect OLED, new accessories, and AI focus in upcoming iOS 18.
Engineering
π₯ [Paper] LoRA+: Boosting Finetuning of Large Models with Efficient Low Rank Adaptation
LoRA+ revolutionizes large model finetuning with dual learning rate strategy, achieving up to 2X speed and 1-2% performance boost.
π₯ [GitHub] Ollama v0.1.33 Release: Introducing Llama 3, Phi 3 Mini, and Qwen 110B Models
Ollama v0.1.33 drops: Llama 3, Phi 3 Mini & Qwen 110B debut, alongside fixes & new parallel processing features.
[Paper] FP6-LLM: Boosting Large Language Model Inference with FP6 and GPU Kernel Design
FP6-LLM revolutionizes AI by shrinking large model sizes without losing quality, enabling faster, cheaper AI on a single GPU.
[Npm] Ollama Web: Elegant Browser Interface for Ollama AI by Anas Dew
Ollama Web launches: An elegant browser interface for easy access to Ollama AI, offering a selection of models and a sleek UI.
[Github] Cognita: Truefoundry's Open Source Framework for Document Retrieval with RAG
Truefoundry launches Cognita, an open-source framework making document retrieval smart and simple with RAG technology. Dive into a sea of docs effortlessly!
[GitHub] rasbt's LLMs-from-scratch: Fine-tuning Strategies for Classification Tasks
Rasbt's GitHub experiments reveal fine-tuning strategies for LLMs, showing training all layers & using larger models can enhance classification tasks.
[GitHub] RepoQA: Benchmarking Long-Context Code Understanding Across Multiple Languages
RepoQA launches as the first benchmark for long-context code understanding across 5 languages, aiming to refine AI's real-world coding skills.
Stanford CS25: Deep Dive into Transformers [Video]
Dive into the world of AI with Stanford CS25's deep dive into Transformers. Watch now for a transformative learning experience!
[Paper] EyeFormer: Personalized Scanpath Prediction with Transformer and RL
EyeFormer revolutionizes GUI design by predicting individual gaze paths using Transformer & RL, enabling personalized interface optimization.
[Paper] Enhancing LLM Context Utilization with IN2 Training: Introducing FILM-7B
FILM-7B revolutionizes LLMs, mastering long-context info retrieval and boosting real-world task performance. Say goodbye to lost-in-the-middle!
FastSD CPU 1.0.0 Beta 30: Hyper-SD SDXL 1-Step Inference Release
FastSD CPU 1.0.0 Beta 30 launches with Hyper-SD SDXL for quicker StableDiffusion inferences.
Academic
[Paper] OSWorld: A Benchmark for Multimodal Agents in Real Computer Tasks by Salesforce Research & Collaborators
Salesforce & collaborators unveil OSWorld, a groundbreaking benchmark for multimodal AI in real computer tasks, challenging agents with 369 diverse tasks. Results show a gap in AI's operational knowledge.
The Seven Lies of an AI Expert: Self-Citation Scandal Unveiled
AI expert Juan Manuel Corchado caught in self-citation scandal, artificially inflating his scientific impact. Claims of class exercises debunked.
Amazon Scholar Cracks Century-Old Geometry Puzzle with AI
Amazon Scholar and team solve a 100-year-old geometry puzzle, proving 30 points ensure a convex hexagon with AI, streamlining future proof-checking.
[Paper] Comprehensive Survey on Knowledge Graph Embedding Models and Applications by Manita Pote
Manita Pote's paper offers a deep dive into Knowledge Graph embedding models, addressing drawbacks and applications in AI and social networks.
The 1956 Dartmouth Conference: The Dawn of Artificial Intelligence
1956 Dartmouth Conference marks the birth of AI, uniting brilliant minds to explore machine learning, language use, and self-improvement.