TLDR AI 2024-05-16

Leading engineers have cracked the code by achieving top-quality AI performance at a fraction of the cost. How? Fine-tuning small open-source models and running them on OctoAI's optimized infrastructure.

Learn from OctoAI's experts about:

• The benefits of fine-tuning for classification and content moderation
• Efficient dataset creation and fine-tuning techniques for Mixtral 7B
• Achieving GPT-4 quality at 60x lower costs

Read the blog to dive into the details — then sign up to bring your fine-tuned model to OctoAI today.

Ilya Sutskever, OpenAI's co-founder and chief scientist, is officially leaving the company. Sutskever helped lead the coup against Sam Altman and then later changed his mind. His employment status had been ambiguous since the ouster. Jakub Pachocki, the company's director of research, will be OpenAI's new chief scientist.

Project IDX is an AI-centric, browser-based development environment that was launched as an invite-only service in August. Over 100,000 developers have already tried the service, which is now available in open beta. Google has added an integration with the Google Maps Platform into the IDE and integrations with the Chrome Dev Tools and Lighthouse to help debug applications. It will soon enable deploying apps to Cloud Run, Google Cloud's serverless platform for running front- and back-end services, and integrate with Checks, Google's AI-powered compliance platform.

The newly developed Hunyuan-DiT sets a benchmark in text-to-image diffusion transformers, especially for English and Chinese. It features advanced transformer structures and a refined data pipeline for continual model improvement.

Generating 3D models is challenging. Allowing users to input a coarse mesh first and then run generation allows for finer-grained control and improved output quality from the model.

Perturbed-Attention Guidance (PAG) is an approach to enhancing the quality of images generated by diffusion models without additional training or external modules. By creatively manipulating self-attention mechanisms within the model, PAG significantly improves the structure and fidelity of both unconditional and conditional samples.

AI Relighting tool that uses foreground separation and text conditioning for extremely strong results.

EfficientTrain++ introduces a novel curriculum learning strategy that significantly reduces the training time of major visual models like ResNet and Swin by up to three times on ImageNet databases.

SqueezeTime is a new lightweight network designed for mobile video understanding that condenses the time axis of videos into the channel dimension to enhance temporal analysis.

Building a useful AI SaaS product requires your models to have access to your users' external data (documents, Slack convos, emails, CRM data etc.) for your multi-tenant RAG pipeline.

Paragon is an embedded data ingestion engine that engineering teams rely on to extract users' data from 100+ sources, without the plumbing.

The new Theta model is a merge between Hermes 2 Pro and Llama 3 8B instruct. It has excellent function calling capabilities and strong performance on a variety of benchmarks.

Hopfield Boosting is a technique that utilizes modern Hopfield energy to enhance the detection of out-of-distribution (OOD) data in machine learning models.

Ramp's Q1 analysis reveals a 293% year-over-year increase in AI-related spending among its customers, outpacing overall software spend growth, with a significant adoption in non-tech sectors like healthcare and financial services, indicating a broader integration of AI into various industries. Despite a general slowdown in new investments in AI, companies already using AI are doubling down, with average spending on AI tools growing by 138% year-over-year, while finance tools also see increased usage and businesses maintain cautious approaches to travel expenditure.

Braintrust's AI Recruiter can do things like generate job descriptions, match candidates, async interviews, and grading.

Educators are shifting away from syntax and emphasizing higher-level skills.

The ability to browse and use OpenAI's GPT store is now available to all users.

In this online talk by Confluent and SAP, you'll learn how to keep every system that relies on SAP data in sync with the latest and greatest real-time data. Watch it for free