GPT Reviews

the intersection of AI and music with the Grammys' new rules for AI use. We also dive into the OpenLLaMA project, an open source reproduction of Google's LLaMA language model. Our AI research expert, Belinda, joins us to discuss two papers: the Recurrent Memory Decision Transformer, which proposes a model for reinforcement learning tasks, and the Block-State Transformer, which combines State Space Models and Block Transformers for language modeling.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:40 And the award goes to AI ft. humans: the Grammys outline new rules for AI use

03:51 OpenLLaMA: An Open Reproduction of LLaMA

05:38 GPT Engineer

06:22 Fake sponsor

08:47 Recurrent Memory Decision Transformer

10:25 Block-State Transformer

12:08 Demystifying GPT Self-Repair for Code Generation

14:01 Outro

OpenAI's ChatGPT takes the AI world by storm with 1 million users in five days, while Google revolutionizes online shopping with a virtual try-on feature. The essay "Imaginary Problems Are the Root of Bad Software" sheds light on the importance of clear communication in software development. Meanwhile, research papers explore TryOnDiffusion, image captioning, and the BBF agent's super-human performance in Atari games. It's a whirlwind of AI advancements and insights that will leave you craving more.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:07 OpenAI’s Mira Murati: the woman charged with pushing generative AI into the real world

04:03 How AI makes virtual try-on more realistic

06:21 Imaginary Problems Are the Root of Bad Software

07:54 Fake sponsor

11:51 TryOnDiffusion: A Tale of Two UNets

13:35 Image Captioners Are Scalable Vision Learners Too

15:16 Bigger, Better, Faster: Human-level Atari with human-level efficiency

17:19 Outro

Amazon's leaked document on ChatGPT, Japan's decision on AI training copyrights, and Israel's support for copyrighted works in machine learning. We also explore two research papers on end-to-end reinforcement learning for robotic mobile manipulation and benchmarking general conditional image similarity.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:51 A leaked document of Amazon's ideas for using ChatGPT and AI at work lists 67 ways to take advantage of the ChatGPT boom

03:36 Japan Goes All In: Copyright Doesn’t Apply To AI Training

04:49 Israel Ministry of Justice Issues Opinion Supporting the Use of Copyrighted Works for Machine Learning

06:38 Jim Fan on the future of Open Source Language Models

08:00 Fake sponsor

09:53 Galactic: Scaling End-to-End Reinforcement Learning for Rearrangement at 100k Steps-Per-Second

11:51 GeneCIS: A Benchmark for General Conditional Image Similarity

13:47 Outro

Mistral AI secures $113 million seed funding to compete against OpenAI with a unique approach, while Hugging Face partners with AMD to optimize transformer performance. The Beatles announce a final record using John Lennon's voice via AI assist, raising concerns about the ethics of AI in music. We also dive into research papers exploring the efficiency of language models, the use of retrieval-enhanced models, and advanced techniques for next-gen language models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:39 France’s Mistral AI blows in with a $113M seed round at a $260M valuation to take on OpenAI

03:34 Hugging Face and AMD partner on accelerating state-of-the-art models for CPU and GPU platforms

05:12 The Beatles will release a final record, using John Lennon's voice via an AI assist

06:27 Fake sponsor

07:58 Orca: Progressive Learning from Complex Explanation Traces of GPT-4

10:03 A Quantitative Review on Language Model Efficiency Research

11:47 Retrieval-Enhanced Contrastive Vision-Text Models

13:36 Outro

New function calling capability in the Chat Completions API, and the release of I-JEPA, a new AI model based on Yann LeCun's vision for more human-like AI. They also discuss the collaboration between OpenAI, Google DeepMind, and Anthropic with the UK government, and the potential risks of misuse. Finally, the team explores two papers, "Augmenting Language Models with Long-Term Memory" and "Benchmarking Neural Network Training Algorithms," which propose a framework for language models to memorize long history and a new benchmark to reliably identify training algorithm improvements.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:45 Function calling and other API updates

03:31 I-JEPA: The first AI model based on Yann LeCun’s vision for more human-like AI

05:09 OpenAI, DeepMind will open up models to UK government

07:16 Hot takes on open-source Language Models

08:24 Fake sponsor

10:23 Augmenting Language Models with Long-Term Memory

11:54 Benchmarking Neural Network Training Algorithms

13:44 Outro

OpenAI's ChatGPT leaked potential new features for a business variant of the chatbot, while Aya aims to accelerate multilingual AI progress through open collaboration. LEACE and Leveraging Large Language Models for Scalable Vector Graphics-Driven Image Understanding propose new methods for improving fairness, interpretability, and computer vision.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:33 Leaked ChatGPT Docs Reveal Potential New Features

03:08 Introducing Aya: An Open Science Initiative to Accelerate Multilingual AI Progress

06:01 Apple execs on Facebook

07:01 Fake sponsor

08:40 LEACE: Perfect linear concept erasure in closed form

10:23 Can Large Language Models Infer Causation from Correlation?

11:42 Leveraging Large Language Models for Scalable Vector Graphics-Driven Image Understanding

13:44 Outro

They discuss Meta's MusicGen AI model that can generate music with very little data, Google's Bard AI language model that is improving at mathematical tasks, BlenderBot 3x that is trained using organic conversation and feedback data, and MIMIC-IT, a dataset comprising 2.8 million multimodal instruction-response pairs for vision-language tasks that has been used to train the Otter model that outperformed existing models on several tasks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:34 Meta just released MusicGen, a simple and controllable model for music generation

03:07 Bard is getting better at logic and reasoning

04:47 Prompts for Work & Play: Launching the Wolfram Prompt Repository

05:37 Fake sponsor

07:36 Improving Open Language Models by Learning from Organic Interactions

09:09 MIMIC-IT: Multi-Modal In-Context Instruction Tuning

10:41 Rewarded soups: towards Pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards

12:17 Outro

Microsoft grants government access to GPT-4, while Google Bard improves by 30% in reasoning abilities. AlphaDev discovers faster sorting algorithms, and LLMZip proposes a lossless text compression algorithm using large language models. These advancements have the potential to transform how we automate responses, program computers, and compress and store text.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:08 Microsoft Gives The Govt. GPT-4

03:30 Google Bard Just got 30% better

05:12 AlphaDev discovers faster sorting algorithms

06:17 Fake sponsor

07:56 Benchmarking Foundation Models with Language-Model-as-an-Examiner

10:17 LLMZip: Lossless Text Compression using Large Language Models

11:35 Large Language Models are In-Context Semantic Reasoners rather than Symbolic Reasoners

13:18 Outro

Instagram's AI chatbot, Tim Cook's take on AI and ChatGPT, and Google Cloud's new no-cost generative AI training courses. They also delve into research papers that explore improving the trustworthiness of Large Language Models, training interpretable Transformers, and augmenting LLMs with databases as their symbolic memory.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:46 Instagram is apparently testing an AI chatbot that lets you choose from 30 personalities

03:11 Tim Cook Talks ChatGPT and AI

05:00 Seven new no-cost generative AI training courses to advance your cloud career

06:28 Fake sponsor

08:52 Deductive Verification of Chain-of-Thought Reasoning

10:24 Learning Transformer Programs

12:08 ChatDB: Augmenting LLMs with Databases as Their Symbolic Memory

14:04 Outro