ThursdAI - The top AI news from the past week

In this podcast, they discuss the importance of datasets in training AI models and mention specific individuals and techniques that have contributed to effective model training. They also talk about text wrangling and dataset visualization, improving dataset classification through clustering, and an open-source cloud service for dataset processing and computation. Additionally, they discuss the process of building datasets, introduce the latest release from RWKV called Eagle, and explore potential improvements in transformer architectures.

Hosts discuss new developments in open AI models including the release of CodeLama 70B, function calling in Mistral and Mixtral, and the non-transformer based Eagle 7B model. They also mention hosting open AI models locally and breaking news from the Allen Institute. Challenges of evaluating language models and open source considerations are explored, along with the release of BGM3 and Nomic AI's embedding models. Copybar, DPO training, and the Capybar dataset are discussed, as well as the Technium Hermes dataset. The release of Whisper Kit and the performance of LLaVa 1.6 are mentioned.

In this episode, Maxime Labonne, Senior Machine Learning Scientist at JPMorgan and author of Hands on GNNs book, talks about the increasingly popular technique of model merging and its impact on the AI community. He discusses the creation of LazyMergeKit, a wrapper on top of MergeKit by Charles Goddard, which has become a widely used library for model merging. The episode also covers the challenges of running benchmarks and creating leaderboards, the use of LNMA FITU for evaluations, and Maxim's involvement in training small language models for playing chess.

This podcast covers topics such as multi-modal transformer models, open source language models, video generation, open source medical language models, and fraud prevention. They also discuss lucid dreaming, creating open source datasets, and AI partnerships. The hosts touch on various AI-related topics including high resolution image synthesis, deepfake audio, government partnerships, and the National Artificial Intelligence Research Resource initiative.

The podcast discusses topics such as OpenAI's new election guidelines, DeepMind's AlphaGeometry neurosymbolic model for solving geometry, Samsung's AI capabilities in the S24 flagship phone, and the merging of models in AI. It also covers reinforcement learning technique DPO, optimization techniques for Asian frameworks, technology detection, Microsoft's copilot announcement and dispute, hackathons, GPU training, and the progress of Llama 3 as a multi-model system.

Creator of CrewAI, João Moura, discusses the inspiration, technical challenges, and success of CrewAI. Jon Durbin talks about Bagel merges, including its origins and fine-tuning stage. They also explore the use of fake facts and context in training AI models.

The podcast covers various interesting topics such as advancements in language models, Mistral's rise as one of the top LOMs, and OpenAI's launch of GPT store. They also discuss conditional training techniques using CRLFT and upcoming hackathons with sponsors. The Luma team updates and enhancements to the Genie are also highlighted.

New WizardCoder achieves 79% HumanEval score. Teknium's Hermes2 on SOLAR 10.7B. Microsoft's E5 SOTA text embeddings with Mistral. Alibaba updates QWEN-VL PLUS to 14B. Nvidia + Suno release NeMo Parakeet beating Whisper on english ASR. Stanford's Mobile ALOHA bot - Open source cooking robot.

Topics covered include new multimodal OSS releases, OpenAI being sued by the NYT, predictions for 2024, open source LLMs, Apple's ML-Ferret model, controversy surrounding the Upstage model, lawsuit implications on licensing and bias, copyright concerns for B6's image generation, and predictions for open-source language model systems in 2024.

This week's podcast covers controversial topics in AI, including the takedown of the LAION 5B dataset due to CSAM allegations, challenges in evaluating AI models, and the advancements in transformer architectures. They also discuss Apple's MLX platform, Midjourney v6's image generation capabilities, and Microsoft Copilot's new plugins for AI music generation. The hosts emphasize the importance of evaluation frameworks and highlight the implications of data set controversies.