TLDR AI 2024-06-19

Systems powered by Nvidia's Hopper architecture dominated the results of two new tests from MLPerf, an AI benchmarking suite that compares the fine-tuning of large language models and training of graph neural networks.

Geoff Hinton and Max Welling, preeminent researchers in the field of AI, have assembled a strong team and are working to build AI systems that help with material science for carbon capture.

Perplexity is leveraging its strategic partnership with SoftBank to expand its reach in Japan. It is offering a free year of its premium AI-powered search engine, Perplexity Pro. SoftBank aims to attract users with AI services without developing in-house solutions. Perplexity, valued at $1 billion, is advancing its funding and investor base, which includes notable tech leaders and venture firms.

ERASE is a new approach for updating language models. Unlike traditional methods that focus on improving retrieval during prediction, ERASE incrementally deletes or rewrites entries in the knowledge base when new documents are added.

Duoduo CLIP is a 3D representation learning model that uses multi-view images instead of point-clouds.

CAMixerSR performs image super-resolution by smartly applying convolution for simpler parts and deformable window-attention for detailed textures.

OGNI-DC is a new framework for depth completion using "Optimization-Guided Neural Iterations" (OGNI). It refines a depth gradient field and integrates the depth gradients into a depth map.

Chameleon is an early fusion model that operates on images and text tokens simultaneously. Its team released the paper a few weeks ago. They have now released model checkpoints along with inference code.

Subobject tokenization is a new way for vision models to understand images. Instead of breaking images into fixed square patches, models using subobject tokenization look at images in terms of meaningful segments, like parts of objects.

Open models with state-of-the-art code performance that do depth upscaling to initialize larger models. Unlike many code models, these models also excel at fixing code, explaining code, and evaluating areas for improvement.

Open source local agent Open Interpreter has just had a big upgrade. It can control the computer seamlessly and works completely offline and locally.

LlamaIndex introduces the Property Graph Index, enhancing knowledge graph capabilities with better modeling, storage, and querying features. The new index allows flexible construction of graphs, supports schema-guided, implicit, and free-form entity extraction, and integrates with vector databases for hybrid searches. It provides querying through keyword expansion, vector similarity, Cypher queries, and custom traversal.

Open model that matches GPT4-o on function calling benchmarks trained on top of Llama 3 70B.

Experience increased productivity and accelerated growth with AI-powered features integrated into the HubSpot customer platform.

Decagon is building human-like AI agents for customer support. It raised a $30M Series A from Accel and $5M Seed from a16z. Decagon's product handles global support for companies like Eventbrite, Rippling, Webflow, BILT, and Substack.