GPT Reviews

From the potential misuse of AI-generated imagery for disinformation campaigns to the use of AI chatbots to frustrate telemarketing scammers, the episode highlights the versatility of AI in addressing persistent issues. The development of Hint-ReLIC and Statler state-maintaining language models also show promising advancements in neural algorithmic reasoning and large language model-based robot reasoning.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:36 Political Satirist Slammed for Creating Deepfakes of Trump, Biden Cheating on their Wives

02:51 California man's business is frustrating telemarketing scammers with chatbots

04:26 The industry behind the industry behind AI

05:39 Fake sponsor

07:30 Neural Algorithmic Reasoning with Causal Regularisation

09:15 Statler: State-Maintaining Language Models for Embodied Reasoning

10:40 A Survey on Generative Diffusion Model

12:46 Outro

Elon Musk blames AI companies for new paywalls on reading tweets. European VCs and tech firms warn against over-regulation of AI in EU la. GPT-4's big secret is revealed: it's a mixture of 8 smaller models using a 2-year-old technique. AI research papers discussed include Preference Ranking Optimization for Human Alignment, ChatGPT for Robotics, and Large Language Models are Effective Text Rankers with Pairwise Ranking Prompting.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:36 Elon Musk blames data scraping by AI startups for his new paywalls on reading tweets

02:51 European VCs and tech firms sign open letter warning against over-regulation of AI in draft EU laws

04:43 GPT-4's Secret Has Been Revealed

06:06 Fake sponsor

08:18 Preference Ranking Optimization for Human Alignment

09:49 ChatGPT for Robotics: Design Principles and Model Abilities

11:25 Large Language Models are Effective Text Rankers with Pairwise Ranking Prompting

13:34 Outro

From the first fully AI-generated drug entering clinical trials to a new method for improving out-of-distribution performance of visual detectors, there's plenty to pique your interest. The OBELISC dataset is also discussed, which could have important implications for training large multimodal models. Finally, the potential catastrophic risks associated with AI are explored, with practical suggestions proposed for mitigating these dangers.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:33 How AI Influences What You See on Facebook and Instagram

03:13 The first fully A.I.-generated drug enters clinical trials in human patients

05:27 Your guide to AI: July 2023

07:00 Fake sponsor

08:52 Test-time Adaptation with Slot-Centric Models

10:23 OBELISC: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents

12:05 An Overview of Catastrophic AI Risks

14:07 Outro

Amazon's AI marketplace, Stability AI's recent changes, potential biases in large language models, and the use of LLMs by crowd workers. These discussions shed light on important issues in the AI industry and highlight the need for transparency and oversight in the development and use of AI models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:41 Amazon reveals AI strategy

03:37 Stability AI’s head of research quits

05:21 Optimize & Deploy BERT on AWS inferentia2

06:29 Fake sponsor

08:41 Towards Measuring the Representation of Subjective Global Opinions in Language Models

10:44 Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks

12:02 What Makes ImageNet Look Unlike LAION

13:52 Outro

ERNIE 3.5 has made significant improvements in efficacy, functionality, and performance, while OpenAI's work assistant may put the company at odds with Microsoft's copilots. "Learning to Rank in Generative Retrieval" proposes a novel framework that achieves state-of-the-art performance among generative retrieval methods. Finally, "System-Level Natural Language Feedback" proposes a framework for using natural language feedback to improve machine learning models at a system-level, leading to further gains in model performance.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:24 Introducing ERNIE 3.5: Baidu’s Knowledge-Enhanced Foundation Model Takes a Giant Leap Forward

02:56 Would an OpenAI work assistant put the company in conflict with Microsoft?

04:24 Fast machines, slow machines

05:46 Fake sponsor

07:56 Learning to Rank in Generative Retrieval

09:24 Extending Context Window of Large Language Models via Positional Interpolation

11:24 System-Level Natural Language Feedback

13:09 Outro

DeepMind is developing a chatbot called Gemini that will rival ChatGPT, while Reka has received $58 million in funding to advance AI research. We also explore the emerging discipline of prompt engineering and the use of large language models as weak learners in machine learning pipelines.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:34 DeepMind claims its next chatbot will rival ChatGPT

02:53 Announcing our $58M funding to Build Generative Models and Advance AI Research

04:11 Prompt Engineering Guide

05:57 Fake sponsor

07:45 Language models are weak learners

09:02 Supervised Pretraining Can Learn In-Context Reinforcement Learning

10:59 SequenceMatch: Imitation Learning for Autoregressive Sequence Modelling with Backtracking

12:51 Outro

Databricks' acquisition of MosaicML for $1.3 billion, Ramp's acquisition of Cohere.io to improve its customer service, and two research papers on self-supervised evaluation for large language models and scaling MLPs. These topics provide valuable insights into the growing demand for generative AI tools, the importance of realistic data evaluation, and the limits of MLPs' performance on vision tasks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:31 Databricks Strikes $1.3 Billion Deal for Generative AI Startup MosaicML

03:33 As the generative AI craze rages on, Ramp acquires customer support startup Cohere.io

05:36 Inside China's underground market for high-end Nvidia AI chips

07:36 Fake sponsor

09:31 Bring Your Own Data! Self-Supervised Evaluation for Large Language Models

11:28 System-Level Natural Language Feedback

12:55 Scaling MLPs: A Tale of Inductive Bias

14:50 Outro

Stability AI's SDXL 0.9, which has potential applications in film, television, music, design, and industrial fields, to archaeologists using AI to translate 5,000-year-old cuneiform tablets, this episode is packed with fascinating developments. We also delve into MIT and Stanford's "From Word Models to World Models," which proposes a computational framework for language-informed thinking, and Google's AudioPaLM, a large language model that fuses text-based and speech-based language models into a unified multimodal architecture that can process and generate both text and speech.

Contact: [email protected]

Timestamps:

00:34 Introduction

01:44 Stability AI launches SDXL 0.9: A Leap Forward in AI Image Generation

03:14 Archaeologists use artificial intelligence (AI) to translate 5,000-year-old cuneiform tablets

04:50 Extending Context is Hard…but not Impossible

05:59 Fake sponsor

08:17 From Word Models to World Models: Translating from Natural Language to the Probabilistic Language of Thought

10:06 Learning from Visual Observation via Offline Pretrained State-to-Go Transformer

11:42 AudioPaLM: A Large Language Model That Can Speak and Listen

13:22 Outro

DeepMind's self-improving robotic agent, RoboCat, and Microsoft's new AI model, Orca. They also discuss two research papers, "RED FM" and "SqueezeLLM," that introduce new resources for multilingual relation extraction and a post-training quantization framework, respectively.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:37 DeepMind Announces RoboCat: A self-improving robotic agent

03:35 Meet Orca, Microsoft’s new 13 billion parameter AI model that can imitate GPT-4

05:40 Levanter — Legible, Scalable, Reproducible Foundation Models with JAX

06:57 Fake sponsor

09:15 RED$^{\rm FM}$: a Filtered and Multilingual Relation Extraction Dataset

11:06 Fast Diffusion Model

12:38 SqueezeLLM: Dense-and-Sparse Quantization

14:25 Outro