Meta

Major Financial Firms Revolutionize Fraud Detection and Credit Scoring with AI-Powered Transaction Models

Financial institutions in 2026 have made significant strides in fraud detection and credit scoring by using large-scale transformer models trained on billions of transaction sequences. Companies like NVIDIA, Stripe, Nubank, Visa, Mastercard, Revolut, and Plaid have developed tools that enable these advancements, with NVIDIA's Build Your Own Transaction Model leading the way. This model uses GPU acceleration and custom tokenization to preprocess data, then trains a compact Llama-based decoder-only AI system. The result? A near-50% improvement in accuracy over traditional methods on IBM's TabFormer fraud dataset. These models are transforming how financial tasks are handled. Instead of relying on outdated rule sets and hand-engineered features, foundation models analyze sequential customer behavior to create robust representations for various applications-like fraud detection, credit scoring, and personalized recommendations. The shift is accelerating across the industry, with firms reporting double-digit performance gains while reducing operational complexity. Looking ahead, expect more financial institutions to adopt these AI-driven approaches, expanding their use in areas like customer segmentation and transaction pattern analysis. The integration of raw data features with pre-trained embeddings promises even greater efficiency and accuracy in fraud detection and beyond.

NVIDIA Dev Blog6d ago

AI Showcases Strong Potential for Automating Data Extraction from Dutch Neuroradiology Reports

AI has demonstrated impressive ability to extract data from complex medical reports, a breakthrough that could transform how radiologists handle their work. In a recent study, researchers tested the LLaMA 3.1 model on over 947 brain MRI reports in Dutch, focusing on variables like atrophy and microbleeds. The AI achieved near-perfect accuracy for categorical data-96% for medial temporal atrophy on the right and 87% for global cortical atrophy-while showing room for improvement with numerical data. The study highlights how few-shot prompting can enhance AI performance, boosting its ability to handle numbers by nearly 12 percentage points. This suggests that with the right strategies, AI could significantly reduce the time doctors spend on repetitive tasks like data extraction. However, challenges remain, particularly in accurately identifying specific lesion locations. Looking ahead, researchers will likely focus on refining these techniques to address remaining gaps. The potential for AI to automate data extraction from medical reports is enormous, offering a clearer picture of how these technologies can support healthcare professionals in the future.

arXiv CS.AI1w ago

AI Revolution Accelerates: Top Breakthroughs and Challenges

1. Mathematicians Unite on AI: Mathematicians have issued a declaration on the use of artificial intelligence in their research, aiming to ensure the discipline's continued growth as AI transforms mathematics, with over 100 mathematical proofs already generated using AI methods. This declaration is important because it highlights the potential of AI in mathematics. 2. AI Verification Breakthrough: A new verification framework has been developed to ensure AI systems are safe and compliant in highly regulated industries, achieving 48.3% regulatory coverage in a pilot program across four sectors, which is a significant improvement over traditional methods. This breakthrough uses an ontology-based approach to automatically generate test scenarios. 3. UK Publishers Opt Out of Google AI: UK publishers can now choose not to appear in Google's AI search results, giving them more power to negotiate with Google, which controls over 90% of the UK's online search market. This decision allows publishers to stop Google from using their content in AI summaries. 4. Instagram Hacked Using Meta AI: Hackers took over Instagram accounts by exploiting Meta AI's chatbot, linking the account to an email they controlled, and then resetting the password, with over 100 accounts hacked, including some with unique short user-profile handles. The company said the issue was fixed, but more users reported hacks. 5. New AI Benchmark Tests Deception: Researchers have introduced SMAC-Talk, a new test environment that evaluates how large language models work together in complex settings, including scenarios where one agent tries to deceive others through misleading messages. This benchmark uses natural language communication to assess coordination among AI agents. 6. AI Agents Handle Long Conversations: NVIDIA has introduced a new AI system that allows agents to carry out extended, multi-step conversations, enabling them to reason, keep track of context, and use tools over many exchanges, making them far more capable in real-world tasks. This development marks a significant leap in how AI interacts with users. 7. AI Alignment Challenge: Recent discussions highlight the critical challenge of aligning superintelligent AI with human values, as these systems can develop internal structures that are incomprehensible to humans, leading to unintended consequences. This challenge is crucial to address to ensure the safe development of superintelligent AI. 8. NVIDIA GPU VRAM Used as Swap Space: A new tool lets Linux users use their NVIDIA GPU's VRAM as swap space, increasing the total addressable memory on a system, which is useful for hybrid graphics laptops with limited upgrade options. This tool works by allocating VRAM via the CUDA driver API. 9. American AI Sovereign Wealth Fund: Senator Bernie Sanders has proposed a plan to give Americans ownership of AI companies by imposing a one-time 50% tax on companies like OpenAI, Anthropic, and xAI, which would be paid in shares, giving the public voting rights and board representation. This plan aims to benefit the public by allowing the government to block harmful decisions. 10. New Physics-Inspired Theory for Deep Learning: A group of researchers has proposed a new framework called "learning mechanics" that aims to create a mathematical theory for deep learning, drawing parallels with physics, which seeks to explain the dynamics of how machine learning models learn, much like classical mechanics explains object movement or quantum mechanics describes particle behavior.

NeuralPulse Daily2w ago

Tiny AI Model Triumphs in Battleship Game, Outshines Giants

Tiny AI models have shown surprising prowess in a unique game-based test. MIT and Harvard researchers used "Collaborative Battleship," where AI agents ask questions to locate hidden ships. They found that smaller models like Llama 4 Scout, which cost 1% of the largest models, performed exceptionally well after strategic tweaks. With a refined approach called Monte Carlo inference, these small models won 82% of their games against humans, surpassing even top-tier AI systems. The study highlights how efficient strategies can make smaller, more affordable AI models competitive in complex tasks. This could democratize access to powerful AI tools, allowing developers and researchers with limited resources to achieve impressive results. The findings challenge the notion that bigger models are always better, suggesting that smarter algorithms can compensate for size. Looking ahead, this research may inspire new ways to optimize AI efficiency across various industries. Whether in medical diagnostics or scientific discovery, smaller models could prove equally effective if equipped with similar strategic improvements.

MIT News AI2w ago

Instagram Accounts Hacked Using Meta AI Chatbot

Hackers took over Instagram accounts by asking Meta AI's chatbot to link the account to an email they controlled. The hackers then reset the account's password and took control. Over 100 accounts were hacked, including some with unique short user-profile handles. These handles can be sold on a gray market for a high price. The company said the issue was fixed, but more users reported hacks on Tuesday. New security measures will be put in place to prevent future hacks.

TechCrunch2w ago

Redditor Builds 1-Terabyte Parameter LLM Using Optane PMem as RAM

A tech enthusiast has made waves by using Intel's discontinued Optane Persistent Memory (DCPMM) sticks as RAM to run a 1-trillion-parameter large language model locally. The setup, detailed on the Local LLaMA subreddit, uses six used Optane PMem sticks totaling 768GB of memory. Despite Optane's slower performance compared to traditional DRAM, it offers lower latency than NVMe SSDs and was acquired at a fraction of the cost of equivalent DRAM capacity. The build includes an Intel Xeon Gold 6246 CPU, Samsung DDR4 ECC RAM for cache, and a Western Digital NVMe SSD. Running Kimi K2.5, the system achieves about 4 tokens per second using a hybrid GPU/CPU inference method with llama.cpp. While Optane's discontinuation limits this approach's scalability, the creator views it as a successful proof-of-concept. The experiment highlights the potential of repurposing older memory technologies for cutting-edge AI tasks, despite their limitations.

Hacker News3w ago

Argonne National Lab Launches AI Inference Service Using Spare Supercomputing Power

The Department of Energy's Argonne National Laboratory has introduced a new AI inference service using spare supercomputing resources. This initiative aims to assist researchers across the U.S., particularly those involved in the Genesis Mission, by providing access to advanced AI models through a chatbot-like interface. The service currently operates on two clusters: Sophia, equipped with 192 Nvidia A100 GPUs, and Metis, featuring 32 SambaNova AI accelerators. Researchers can utilize models like OpenAI's GPT-OSS and Meta’s Llama for tasks such as analyzing experimental data in real-time or processing large datasets from particle accelerators. The service ensures secure AI experimentation without exposing sensitive data to public platforms. As Argonne expands the service to include more systems, this effort promises to enhance scientific discovery by efficiently leveraging underutilized computing resources.

Hacker News3w ago

AI's Personality Test Fails When Put to Work

A new study reveals that AI models trained to mimic specific personalities in chat conversations struggle when given real-world tasks. Researchers tested three major AI systems-Llama, Qwen, and Gemma-trained with personality-based fine-tuning (SFT). These models were scored using a classifier designed to identify their personas, achieving high accuracy (86-95%) in controlled chat settings. However, the same models performed poorly when asked to act autonomously-composing emails or making decisions. The classifier's accuracy dropped sharply to 29-55%, showing that AI personalities don't translate well beyond structured chat interactions. This suggests that SFT, a common training method for character-driven AI, may not prepare models for practical, agent-like tasks. The findings highlight the limitations of current personality-training techniques and emphasize the need for more generalized alignment methods. As AI becomes more integrated into daily life, understanding how these systems behave outside of controlled chats will be crucial for developers aiming to create reliable and versatile AI assistants.

LessWrong4w ago