TLDR AI 2024-09-10

The world (still) runs on email. Invoices, product orders, travel plans - they're all in the inbox. But unstructured data and wildly disparate email formats make extracting the parts you need difficult.

ExtractAI by Nylas is the new way to sync, filter, and extract user data. With just a few lines of code, you can now deploy advanced ML, NLP, and LLMs to retrieve first-party data from your users' inboxes:

• 92% cost savings compared to other AI / LLM tools
• 99.9% accuracy SLA for order data
• Real-time sync with no email forwarding required

See what you can build with email data →

Apple introduced Apple Intelligence at WWDC 2024, focusing on integrating AI features into existing apps like Mail, Messages, and Siri. The platform uses large language models for tasks such as text summarization and image generation, aiming for a seamless user experience. A beta version will roll out in the U.S. this October, with broader international support planned for 2025.

Google's AI Overviews, powered by the Gemini language model, faced heavy criticism for inaccuracies and dangerous suggestions after its U.S. launch. Despite the backlash, Google expanded the feature to six more countries, raising concerns among publishers about reduced traffic and misrepresented content. AI strategists and SEO experts emphasize the need for transparency and better citation practices to maintain trust and traffic.

HyperWrite's Reflection 70B, a variant of Meta's Llama 3.1 LLM, is facing scrutiny after third-party evaluators couldn't replicate its claimed performance. Issues were attributed to model weights being corrupted during the upload to Hugging Face, which led to discrepancies. The AI community awaits further responses and updates to clarify the model's actual capabilities.

Proxy Token Diffusion Transformer reduces redundant computations in diffusion transformers by using fewer representative tokens for attention.

UniDet3D is a powerful 3D object detection model trained on multiple indoor datasets.

Much of modern language modeling is a specific case of imitation learning. Fortunately, there is a depth of study in the field broadly. This paper explores how to use inverse reinforcement learning to imitate entire sequences instead of just tokens. The results are promising and point in the direction of continued RL in language model training pipelines.

Image processing and cropping that can be run on the GPU.

A great tool that uses Mast3r and some clever optimizations to reconstruct 3D scenes from a few 2D images.

Get reliable, up-to-date monolingual and bilingual data in an accessible, scalable environment. Develop for a range of communities with large quantities of quality data from widely used and digitally under-represented languages — Hindi, Latvian, Portuguese, Swahili, and many more. Watch the Oxford Dictionaries API tour

Researchers conducted a study comparing seven different visual encoding models for 3D scene understanding and found that the best model depends on the task. DINOv2 performed the best overall. Video models excelled in object-level tasks and diffusion models were best for geometric tasks. Language-pretrained models had surprising limitations.

Diffusion models and autoregressive models are fundamentally similar, both using iterative refinement processes. Using Fourier transform techniques, this author shows how diffusion models operate like approximate autoregression in the frequency domain, particularly for visual data. This insight highlights potential directions for unifying generative modeling approaches across different data types.

The rise of AI and diverse intelligences challenges outdated distinctions between "real beings" and machines. Instead of focusing solely on AI, people should understand and ethically engage with diverse cognitive systems, including bioengineered, robotic, and hybrid entities. Expanding our definitions of intelligence and compassion will help navigate the ethical landscape of emerging technologies.

A One-DM model that generates handwritten text mimicking any style using just a single sample.

OpenAI can significantly reduce operational costs with the adoption of more efficient chips, especially as its user base grows.

Apple is heavily promoting the Mac as the "best AI PC" and directly targeting Microsoft's Copilot+ PCs.

Shortwave has significantly upgraded its AI Assistant, enabling it to perform complex, multi-step tasks such as advanced searches, calendar lookups, and email analysis.

It's likely that AI will not just write job descriptions, but actually interview candidates in the near future. Professor of Law at Emory University, Ifeoma Ajunwa, discusses potential risks and possibilities for companies using AI in recruitment. Read the article by Specialist Staffing Group