TLDR AI 2023-11-15

Need more compute power (and choices) to fine-tune your generative AI?

✔️Learn with hands-on tutorials and real-world code samples.

✔️Evaluate performance and acceleration with different hardware configurations.

✔️Build heterogeneous applications, develop your own prototypes, and benchmark your AI workloads.

Start on the free tier to explore the latest Intel® AI tools

Intel® Developer Cloud accelerates your AI development with Intel®-optimized software on the latest Intel® Xeon® processors and GPU compute.

Choose from several configurations tuned to various workloads — AI and inference training, FPGA development, edge prototyping, preproduction deployment, and more.

Retool's State of AI report, which surveyed 1,500 tech workers, reveals generative AI's growing impact on business. MongoDB Atlas Vector Search leads in popularity among vector databases, vital for AI-powered apps enhancing productivity and customer experiences. However, challenges in AI adoption persist, notably in model accuracy and data security, with C-suite executives showing greater optimism than individual contributors.

You.com has launched APIs starting at $100 per month, enabling large language models like Meta’s Llama 2 to access up-to-date internet data and offering enhanced answer accuracy through web search and news functionalities.

At OpenAI’s DevDay, Microsoft announced plans to offer the new GPT-4 Turbo on Azure OpenAI Service by year-end, providing enhanced control and efficiencies at lower costs. OpenAI's Custom Models program will seamlessly integrate with Microsoft's ecosystem, boosting enterprises’ AI capabilities.

Nous research has released a very strong 34B model built on top of the bilingual Yi model. Capybara is trained primarily on multi-turn data with more than 1000 tokens per conversation, which makes it uniquely positioned for certain user-facing tasks with high information density.

When new models come out with astounding capabilities, it is natural to question whether they actually trained on the validation data or on benchmarks. Decontamination is the process of removing such testing data from the input. The LMSYS group found that if you rephrase test data in a way that passes decontamination but still contains information about the benchmark, you can fake your way to amazing performance.

This study introduces MonoDiffusion, a novel framework for self-supervised monocular depth estimation, which uniquely approaches the challenge as an iterative denoising process. It utilizes a pseudo ground-truth diffusion process guided by a pre-trained teacher model to improve depth map accuracy without using actual depth ground-truth in training.

Need to deliver real-time, low-latency data services to edge-based applications? Use NATS.io — the same technology used by thousands of companies across leading Financial Services, Automotive, Retail, and Government institutions.

➡️ Process massive amounts of data where it is generated in real-time

➡️ Train in cloud but execute models locally in any environment

➡️ Focus on the AI application, not the cloud infrastructure

Discover a better way to run AI at the Edge

This project introduces AMBER, a new benchmark designed to assess and reduce hallucinations in Multi-modal Large Language Models (MLLMs) without relying on other advanced language models.

This project introduces Knowledgeable Preference AlignmenT (KnowPAT), a novel approach to improving domain-specific question answering using large language models.

A compelling demo that uses GPT-V and a pleasant sketching interface to create HTML websites.

IBM has committed $500 million to invest in generative AI startups that cater to business clients, signaling a strategic shift and indicating its readiness to support emerging companies, with a preference for those that complement rather than compete with its own business offerings.

Humane's AI Pin is a $699 screenless wearable device with innovative interaction methods powered by a $24 monthly subscription that offers features like language translation, AI-assisted photography, and personal assistant tools on T-Mobile's network.

A study by Redwood Research reveals that large language models (LLMs) can master "encoded reasoning," a form of steganography allowing LLMs to embed intermediate reasoning steps in generated text, making them undecipherable to humans.

An AI applicant tracking and hiring tool.

Structured extraction in Python powered by OpenAI’s function calling API.

Samsung has unveiled Samsung Gauss, a generative AI model that integrates with its devices and enhances tasks like email composition, software development, and image processing.