GPT Reviews

The G7's agreement on an AI code of conduct for companies, ensuring safety and transparency. They also explore Boston Dynamics' ChatGPT, which turns their robot dog Spot into a chatty tour guide. The team then delves into three fascinating research papers: PromptAgent, which optimizes prompts for language models; GlotLID, an improved language identification model for low-resource languages; and controlled decoding, a method to steer autoregressive generation from language models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:53 G7 to agree AI code of conduct for companies

04:19 Boston Dynamics turned its robot dog into a talking tour guide with ChatGPT

06:22 Lucas Beyer on Twitter

08:00 Fake sponsor

10:11 PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization

11:45 GlotLID: Language Identification for Low-Resource Languages

13:16 Controlled Decoding from Language Models

14:58 Outro

OpenAI's new team tackling nuclear threats. The demystification of CLIP data from Meta Research and the Data Provenance Explorer, a tool that provides transparency and accountability in AI datasets. The team also delves into research papers on detecting pretraining data, quantized Transformers, and compression of trillion-parameter models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:25 OpenAI forms team to study ‘catastrophic’ AI risks, including nuclear threats

04:34 Demistifying CLIP data from Meta Research

06:27 Data Provenance Explorer

08:33 Fake sponsor

11:11 Detecting Pretraining Data from Large Language Models

12:44 LLM-FP4: 4-Bit Floating-Point Quantized Transformers

14:28 QMoE: Practical Sub-1-Bit Compression of Trillion-Parameter Models

16:34 Outro

Amazon's AI-powered image generation for better ad experiences, the White House's upcoming AI executive order, and discussions on merging vision foundation models, improving passage ranking with demonstrations, and understanding in-context learning in large language models. With the help of his brilliant guests, GPT dives into the world of AI, delivering a mix of news, humor, and insightful analysis.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:29 Amazon rolls out AI-powered image generation to help advertisers deliver a better ad experience for customers

04:12 White House to unveil sweeping AI executive order next week

05:53 The Jeff Dean Facts

08:04 Fake sponsor

10:50 SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understanding

12:19 PaRaDe: Passage Ranking using Demonstrations with Large Language Models

13:38 In-Context Learning Creates Task Vectors

15:27 Outro

The latest news on GPT-5, the Wafer Wars, and the behavior of AI assistants. They also dive into three research papers, discussing the Branch-Solve-Merge method for large language model evaluation and generation, ULTRA for knowledge graph reasoning, and Matryoshka Diffusion Models for high-resolution image and video synthesis.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:49 Bill Gates does not expect GPT-5 to be much better than GPT-4

03:59 Wafer Wars: Deciphering Latest Restrictions On AI And Semiconductor Manufacturing

05:55 Anthropic Twitter Thread

07:52 Fake sponsor

10:05 Branch-Solve-Merge Improves Large Language Model Evaluation and Generation

11:56 Towards Foundation Models for Knowledge Graph Reasoning

13:30 Matryoshka Diffusion Models

15:48 Outro

Developments in robot learning, where NVIDIA researchers have created an AI agent called Eureka that can generate algorithms to train robots. They also explore the concerning use of AI in cyber warfare by North Korea, and the potential consequences for global enterprises. Additionally, they touch on Apple's rumored plans to implement generative AI features on iPhones and iPads. The team also delves into thought-provoking discussions on AI-generated music and the impact of AI on job automation. Finally, they analyze three research papers that shed light on the limitations of large language models in reasoning and planning tasks, and introduce a new approach called Contrastive Preference Learning for optimizing behavior from human feedback.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:55 Eureka! NVIDIA Research Breakthrough Puts New Spin on Robot Learning

04:15 North Korea experiments with AI in cyber warfare: US official

05:37 Apple Rumored to Follow ChatGPT With Generative AI Features on iPhone as Soon as iOS 18

07:06 Fake sponsor

09:48 GPT-4 Doesn't Know It's Wrong: An Analysis of Iterative Prompting for Reasoning Problems

10:52 Can Large Language Models Really Improve by Self-critiquing Their Own Plans?

12:27 Contrastive Prefence Learning: Learning from Human Feedback without RL

14:21 Outro

Transparency in AI: Major AI companies fail transparency test, hindering effective legislation. Apple's plan for generative AI: Revamping Siri and integrating AI into iOS. YouTube's AI tool: Allows users to sound like their favorite artists, pending licensing deals. Cutting-edge AI research: Eureka's human-level reward design, Vision-Language Models as zero-shot reward models, and AgentTuning for generalized agent abilities.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:24 Top AI Shops Fail Transparency Test

03:52 Inside Apple’s Big Plan to Bring Generative AI to All Its Devices

05:21 YouTube Is Creating an AI Tool That Allows People to Sound Like Popular Artists

07:07 Fake sponsor

09:51 Eureka: Human-Level Reward Design via Coding Large Language Models

11:17 Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning

13:09 AgentTuning: Enabling Generalized Agent Abilities for LLMs

15:07 Outro

OpenAI's failed attempt to develop an AI model named after a dystopian hellscape in 'Dune', as well as the groundbreaking use of AI in ultra-fast deep-learned CNS tumor classification during surgery. The team also explores a thought-provoking Twitter thread on AI research and delves into three fascinating research papers on improving image generation, enabling generalized agent abilities, and using vision-language models as zero-shot reward models for reinforcement learning.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:40 OpenAI tried to develop an AI model named after a dystopian hellscape in 'Dune,' but the project failed to land, report says

04:01 Ultra-fast deep-learned CNS tumour classification during surgery

06:08 Thoughts on doing AI Research by Jason Wei

08:33 Fake sponsor

10:48 DALLE 3: Improving Image Generation with Better Captions

12:14 AgentTuning: Enabling Generalized Agent Abilities for LLMs

13:53 Vision-Language Models are Zero-Shot Reward Models for Reinforcement Learning

15:41 Outro

Universal Music Group partners with BandLab Technologies to protect artists' rights against AI violations, a game-changer for the music industry. AMD, Arm, Intel, Meta, Microsoft, NVIDIA, and Qualcomm collaborate to standardize next-generation narrow precision data formats for AI

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:34 Universal Music Group enters partnership to protect artists' rights against AI violations

04:26 AMD, Arm, Intel, Meta, Microsoft, NVIDIA, and Qualcomm Standardize Next-Generation Narrow Precision Data Formats for AI

06:48 How Transparent Are Foundation Model Developers?

09:05 Fake sponsor

12:10 Reward-Augmented Decoding: Efficient Controlled Text Generation With a Unidirectional Reward Model

13:37 Prometheus: Inducing Fine-grained Evaluation Capability in Language Models

15:32 Rethinking Negative Pairs in Code Search

17:38 Outro

Foxconn and Nvidia are collaborating to build "AI factories" that will accelerate the development of self-driving cars and other autonomous machines. Yann Lecun, an AI legend, shares his thoughts on the potential of open-source Language Learning Models (LLMs) and how they could revolutionize AI research. The Instant Domino Effect is a new device that allows users to set up impressive chain reactions of falling dominoes in seconds. Exciting research papers are discussed, including Llemma, an open language model for mathematics; BitNet, a scalable 1-bit Transformer architecture for large language models; and VeRA, a vector-based random matrix adaptation method for reducing trainable parameters.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:15 Foxconn and Nvidia are building ‘AI factories’ to accelerate self-driving cars

04:54 Yann Lecunn on Open Source LLMs

06:43 ChatGPT Voice System Prompt

08:46 Fake sponsor

11:38 Llemma: An Open Language Model For Mathematics

13:51 BitNet: Scaling 1-bit Transformers for Large Language Models

15:20 VeRA: Vector-based Random Matrix Adaptation

17:32 Outro

OpenAI's DALLE 3 Prompt, a language model that creates diverse and inclusive images based on text prompts. We also explore MemGPT, a system that enables perpetual chatbots with self-editing memory. Additionally, we discuss "The Consensus Game," a game-theoretic approach to language model decoding that improves performance in various tasks. And don't miss our insights on PaLI-3 Vision Language Model, a smaller yet powerful model achieving state-of-the-art results in multilingual cross-modal retrieval.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:24 OpenAI's DALLE 3 Prompt Leaked)

02:48 MemGPT: Perpetual Chatbots with self-editing memory

04:43 Twitter Thread by Jeremy Howard)

05:33 Fake sponsor

07:46 The Consensus Game: Language Model Generation via Equilibrium Search

09:26 PaLI-3 Vision Language Models: Smaller, Faster, Stronger

10:48 Query and Response Augmentation Cannot Help Out-of-domain Math Reasoning Generalization

12:42 Outro