TLDR AI 2024-09-13

There's valuable data in emails, but it's often buried under layers of text, images, and HTML. Nylas ExtractAI is a new way to get that data into the hands of developers, faster and with zero effort.

👉 Try ExtractAI for free

✅ Retrieve first-party data from your users' inboxes with advanced ML, NLP, and LLMs .

✅ Extract features and classify emails: online orders, travel reservations, invoices, and more.

✅ Sync directly with your user's inbox in real time. No forwarding required!

✅ Create high-quality labeled datasets to train your own extractors.

👷‍♀️ Get your free API key and start building

OpenAI has released its next model, which was trained to think before it answers. The new model was trained with reasoning traces and spends time considering before it answers. In some domains, this has led to super human performance. The model will be rate limited to 30 or so queries per user per week, but OpenAI hopes to lift that restriction soon.

Google is rolling out Gemini Live, its conversational AI feature, to free Android users after a month of advanced user access. Users can interrupt responses with new information and receive text transcripts of their interactions. Although Gemini Live doesn't support extensions like Gmail yet, it offers ten new voice options, and more features are promised soon.

Policy Filtration for Proximal Policy Optimization (PF-PPO) is a method designed to improve the accuracy of reinforcement learning from human feedback (RLHF) in code generation tasks.

Researchers have developed a new data augmentation method for improving saliency prediction models, which traditionally suffer from limited labeled data.

Jina has released two new state-of-the-art models that take noisy HTML and parse it into clean and usable Markdown for training and reasoning.

Hi3D is a new model that improves the generation of multi-view consistent, high-resolution 3D images from a single input. It uses a video diffusion approach to tackle the lack of 3D awareness in traditional 2D methods, leveraging temporal consistency from video models to enhance geometry across views.

Executives want to see numbers—are you delivering? Harnessing the power of people analytics helps HR get a seat at the table. Our guide breaks down:

- The definition and uses of data and reporting

- Finding the right metrics to track for your business

- How to put data into action

Download Now

Agent Workflow Memory (AWM) is a method that helps language model-based agents learn reusable task workflows from past experiences to handle complex, long-horizon tasks.

Axolotal AI has partnered with Lambda Labs to show how to use its one click cluster to fine-tune the Llama 3.1 405B model. This requires 64 GPUs, but can be done with minimal infra setup due to new tools.

SUPER is a new benchmark designed to assess how well LLMs can reproduce tasks from research repositories.

Researchers have developed an AI tool that automates scientific processes, performing tasks from hypothesis generation to experiment execution and paper writing. Its accuracy and coherence still need improvement. Critics highlight that, while AI can handle simulations in fields like quantum computing and materials science, it risks narrowing research questions and producing less meaningful knowledge. Proponents believe this AI could optimize early research stages, potentially aiding scientists in conceptualizing and scoping research projects.

An AI-assisted web scraper that uses OpenAI's GPT-4o to extract structured data from HTML tables, with varying results on complex and merged row tables.

Amazon's shopping-focused chatbot, Rufus, will soon display sponsored ads based on search and conversational context.

Google co-founder Sergey Brin has returned to Google to focus on AI because of its rapid advancements.