GPT Reviews

Meta's controversial AI chatbots mimicking celebrities to Adobe's Project Fast Fill using generative AI for video manipulation, we explore the latest developments in the field. We also dive into two research papers, ChunkAttention and (InThe)WildChat, which propose solutions to improve the efficiency of GPT-style language models and offer a dataset for researchers to study potentially toxic use cases and fine-tune instruction following models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:26 Meta's AI celebrities face more resistance than enthusiasm

02:40 Adobe is working on generative AI video manipulation

04:23 Tweet by Hyung Won Chung

05:46 Fake sponsor

07:56 ChunkAttention: Efficient Attention on KV Cache with Chunking Sharing and Batching

09:23 (InThe)WildChat: 570K ChatGPT Interaction Logs In The Wild

11:12 LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression

13:32 Outro

No Fakes Act, the State of AI Report 2023, Learning Interactive Real-World Simulators, and Harmonic Self-Conditioned Flow Matching for Multi-Ligand Docking and Binding Site Design. These topics have implications for the entertainment industry, natural language processing applications, autonomous driving, drug synthesis, and energy storage.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:07 No Fakes Act wants to protect actors and singers from unauthorized AI replicas

04:23 The State of AI Report 2023

05:34 OpenAI is too cheap to beat

07:00 Fake sponsor

09:02 Text Embeddings Reveal (Almost) As Much As Text

10:31 Learning Interactive Real-World Simulators

12:20 Harmonic Self-Conditioned Flow Matching for Multi-Ligand Docking and Binding Site Design

13:56 Outro

Polymathic AI's initiative to change how people use AI and machine learning in science, the high cost of running generative AI models like Microsoft's GitHub Copilot, the impressive performance of Mistral 7B, a 7-billion-parameter language model, and the introduction of iTransformer, a new model for time series forecasting that achieves state-of-the-art results on several real-world datasets.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:32 Announcing: Polymathic AI 🎉

03:12 Microsoft's GitHub Copilot Loses $20 a Month Per User

04:52 Fake sponsor

06:55 Mistral 7B

08:22 LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression

10:28 Functional Interpolation for Relative Positions Improves Long Context Transformers

11:49 iTransformer: Inverted Transformers Are Effective for Time Series Forecasting

13:42 Outro

From Geoffrey Hinton's concerns about the potential risks of advanced AI to Adobe's new symbol for tagging AI-generated content, this episode provides insights into the latest developments in the field. The discussion on Disney Imagineering's latest robot that can walk independently and mimic emotions through body language is also fascinating. Additionally, the episode features three AI research papers on Reinforcement Learning from Human Feedback, neural network predictions, and complex reasoning with Large Language Models, which provide valuable insights into the capabilities and limitations of AI.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:38 Geoffrey Hinton on the promise, risks of advanced AI

03:01 Adobe created a symbol to encourage tagging AI-generated content

04:27 Disney Imagineering’s latest robot looks right out of Star Wars

06:07 Fake sponsor

08:11 A Long Way to Go: Investigating Length Correlations in RLHF

09:44 Deep Neural Networks Tend To Extrapolate Predictably

11:07 Thought Propagation: An Analogical Approach to Complex Reasoning with Large Language Models

13:00 Outro

Open-source AI system LLaVA is introduced, which could rival GPT-4 for visual and language understanding. Microsoft's new AI chip is discussed, which aims to reduce reliance on Nvidia's GPUs. FreshPrompt, a few-shot prompting method, is highlighted for its ability to substantially boost the performance of LLMs on FreshQA. Finally, LATS is introduced, which leverages the strengths of LLMs to enhance decision-making in planning, acting, and reasoning tasks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:43 Open-source LLaVA challenges GPT-4

03:13 Microsoft said to debut AI chip next month in an effort to cut costs

05:01 How I think about LLM prompt engineering

06:17 Fake sponsor

08:11 FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation

09:50 Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models

11:19 Towards Foundational Models for Molecular Learning on Large-Scale Multi-Task Datasets

13:12 Outro

Adobe's new AI photo editing tool, OpenAI's exploration into making their own AI chips, and the potential of generative AI and its impact on employment. Additionally, the episode features three AI research papers that showcase the capabilities of GPT-4 in playing imperfect information card games, DSPy in optimizing language model pipelines, and MathCoder in enhancing the mathematical reasoning abilities of language models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:46 Adobe teases new AI photo editing tool that will ‘revolutionize’ its products

03:11 ChatGPT-owner OpenAI is exploring making its own AI chips

05:05 Generative AI exists because of the transformer

06:30 Fake sponsor

08:27 Suspicion-Agent: Playing Imperfect Information Games with Theory of Mind Aware GPT-4

10:00 DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines

11:40 MathCoder: Seamless Code Integration in LLMs for Enhanced Mathematical Reasoning

13:17 Outro

The potential dangers of chatbot-generated false information, the exciting developments in Large Language Models, the discovery of truly general reinforcement learning algorithms, and the cross-lingual vulnerability of large language models. These topics highlight the importance of understanding the implications of AI and the need for robust safeguards to ensure their safety.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:28 Chatbot Hallucinations Are Poisoning Web Search

03:27 The state of LLMs on 2023 by Hyung Won Chung

04:36 Towards Monosemanticity: Decomposing Language Models With Dictionary Learning

05:48 Fake sponsor

07:39 Retrieval meets Long Context Large Language Models

09:12 Discovering General Reinforcement Learning Algorithms with Adversarial Environment Design

10:47 Low-Resource Languages Jailbreak GPT-4

12:32 Outro

Yahoo's spins off its AI-serving engine, Vespa, into its own separate company, the changing trends of publishing in the AI field, the impressive fundraising of a four-week-old AI startup, and the capabilities of large language models in various settings.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:44 Vespa is becoming a company

03:46 Jason Wei (Researcher at OpenAI) reflects on Publishing Trends

04:49 See the pitch memo that raised €105m for four-week-old startup Mistral

06:11 Fake sponsor

08:21 Language Models Represent Space and Time

10:13 Large Language Models as Analogical Reasoners

11:26 Efficient Streaming Language Models with Attention Sinks

13:30 Outro

Open X-Embodiment dataset and RT-1-X model checkpoint for general-purpose robotics learning, as well as the new multimodal AI assistant, Yasa-1. We also delve into the Representation Engineering (RepE) approach to improving transparency and safety in AI systems, and the PIT framework for enabling language models to implicitly learn self-improvement from data.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:53 Scaling up learning across many different robot types

03:38 Announcing our Multimodal AI Assistant at Reka

05:30 Evaluating LLM Outputs

06:52 Fake sponsor

08:59 Representation Engineering: A Top-Down Approach to AI Transparency

10:35 AutomaTikZ: Text-Guided Synthesis of Scientific Vector Graphics with TikZ

12:10 Enable Language Models to Implicitly Learn Self-Improvement From Data

13:53 Outro

OpenAI's DALL-E 3 image generator upgrade to the limitations of pre-trained language models and the introduction of Promptbreeder, there's plenty to keep listeners engaged. The episode also delves into the strange world of "Weird A.I. Yankovic," a cursed deep dive into the world of voice cloning.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:40 OpenAI Quietly Ships DALL-E 3 AI Image Generator Upgrade In Bing

03:50 Weird A.I. Yankovic, a cursed deep dive into the world of voice cloning

04:58 Tweet by Jason Wei

06:21 Fake sponsor

08:13 Physics of Language Models: Part 3.2, Knowledge Manipulation

09:28 Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution

11:07 Efficiency Pentathlon: A Standardized Arena for Efficiency Evaluation

12:56 Outro