GPT Reviews

Jony Ive and OpenAI are teaming up to create the "iPhone of artificial intelligence" with over $1 billion in funding. Mistral 7B is the most powerful language model for its size to date, outperforming Llama 2 13B on all benchmarks. Tom Hanks warns fans about a fake dental plan ad that uses his image created with AI. Belinda, the AI research expert, discusses advancements in long-context scaling models, multi-agent motion forecasting, and vision transformers.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:36 Details emerge on Jony Ive and OpenAI’s plan to build the ‘iPhone of artificial intelligence’

02:59 Mistral 7B is Released

04:23 Tom Hanks says AI version of him used in dental plan ad without his consent

05:49 Fake sponsor

07:35 Effective Long-Context Scaling of Foundation Models

09:14 MotionLM: Multi-Agent Motion Forecasting as Language Modeling

10:44 Vision Transformers Need Registers

12:39 Outro

ChatGPT can now browse the internet to provide users with current information, but concerns about accuracy and reliability remain. Meta has introduced social profiles for its AIs, allowing users to interact with them directly on Instagram, Messenger, and WhatsApp. The paper "Finite Scalar Quantization: VQ-VAE Made Simple" proposes a simpler alternative to vector quantization in VAEs, which could be a game-changer for the field. "Show-1: Marrying Pixel and Latent Diffusion Models for Text-to-Video Generation" proposes a hybrid model that combines pixel-based and latent-based VDMs for more efficient and accurate text-to-video generation.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:46 ChatGPT can now access up to date information

03:44 Introducing Social Profiles for Meta’s AIs

05:27 WebGPU Technical Report

06:29 Fake sponsor

09:02 Finite Scalar Quantization: VQ-VAE Made Simple

10:28 Emu: Enhancing Image Generation Models Using Photogenic Needles in a Haystack

12:09 Show-1: Marrying Pixel and Latent Diffusion Models for Text-to-Video Generation

14:04 Outro

OpenAI CEO Sam Altman caused a stir by posting a comment that AGI has been achieved internally, only to edit the post later to dispel the news.  Meta is bringing its AI assistant to WhatsApp, Messenger, and Instagram, and releasing dozens of AI characters based on celebrities like MrBeast and Charli D’Amelio.  "Aligning Large Multimodal Models with Factually Augmented RLHF" proposes a new approach to address the issue of misalignment between modalities in large multimodal models, achieving a remarkable 94% performance level on the LLaVA-Bench dataset. "InternLM-XComposer" is a vision-language large model that enables advanced image-text comprehension and composition, seamlessly integrating images into generated articles and consistently achieving state-of-the-art results on various benchmarks for vision-language foundational models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:32 Sam Altman says (trolls) saying AGI has been achieved

03:04 Meta is putting AI chatbots everywhere

05:08 26% of the top 100 websites are now blocking GPTBot

06:04 Fake sponsor

08:19 Aligning Large Multimodal Models with Factually Augmented RLHF

10:07 VideoDirectorGPT: Consistent Multi-scene Video Generation via LLM-Guided Planning

11:59 InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition

13:58 Outro

The rise of AI girlfriends and their impact on America's future population is a thought-provoking topic discussed in this episode. The use of AI for competitive intelligence and its advantages for companies is explored in-depth, with a focus on a new company called Prelaunch.com. The papers discussed in this episode shed light on important topics such as understanding large language models, training instabilities, and causality for machine learning. The insights provided by the experts on the show offer valuable perspectives on the implications of AI for various industries and society as a whole.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:45 AI girlfriends are ruining an entire generation of men

03:06 AI Driven Competitive Intelligence Accelerates Research Productivity

05:03 Causality for Machine Learning

06:11 Fake sponsor

08:02 Studying Large Language Model Generalization with Influence Functions

09:26 Embers of Autoregression: Understanding Large Language Models Through the Problem They are Trained to Solve

11:26 Small-scale proxies for large-scale Transformer training instabilities

13:20 Outro

OpenAI's ChatGPT-4 can now see, hear, and speak, making it more intuitive for daily use. Amazon is investing up to $4 billion in OpenAI rival Anthropic, which is developing a new AI chatbot called Claude 2. Spotify is partnering with OpenAI to clone podcasters' voices and translate their shows into other languages. Lastly, Text2Reward is a new framework that automates the generation of dense reward functions for reinforcement learning.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:32 ChatGPT can now see, hear, and speak: Announcing GPT-4 multimodal

03:16 Amazon to invest up to $4bn in OpenAI rival Anthropic

04:51 Spotify is going to clone podcasters’ voices — and translate them to other languages

06:14 Fake sponsor

08:11 ChatGPT v Bard v Bing v Claude 2 v Aria v human-expert. How good are AI chatbots at scientific writing? (ver. 23Q3)

09:49 DePT: Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning

11:35 Text2Reward: Automated Dense Reward Function Generation for Reinforcement Learning

13:21 Outro

Meta's plan to release a "sassy robot" for younger users and concerns about potential implications of these chatbots. Google's new Bard Extensions feature and its performance in retrieving emails and drafting responses. Research on language models with Claude's long context window and the need for models that can effectively use information from the middle of long contexts. The development of MetaMath, a fine-tuned language model that specializes in mathematical reasoning, and its superior performance on mathematical reasoning benchmarks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:33 Meta’s AI chatbot plan includes a ‘sassy robot’ for younger users

03:17 I tried Google's new Bard Extensions feature which integrates with apps like Gmail. The AI assistant isn't perfect, but it has one clear strength.

04:59 Prompt engineering for Claude's long context window

06:42 Fake sponsor

08:30 Lost in the Middle: How Language Models Use Long Contexts

09:52 Evaluating Large Language Models for Document-grounded Response Generation in Information-Seeking Dialogues

11:21 MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models

13:22 Outro

OpenAI's privacy lawsuit has been dismissed, but the company is still facing legal controversies. YouTube is introducing new AI-powered tools for creators, including AI-generated backgrounds and personalized music recommendations. We also discuss the potential impact of open source AI on language and image models, and a new study that shows AI systems have a much lower carbon footprint than humans when it comes to tasks like writing and illustrating.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:25 A lawsuit alleging privacy violations by OpenAI was dismissed

03:01 YouTube is going all in on AI

04:51 Why Open Source AI Will Win

06:30 Fake sponsor

08:42 Chain-of-Verification Reduces Hallucination in Large Language Models

10:02 Kosmos-2.5: A Multimodal Literate Model

11:31 The Carbon Emissions of Writing and Illustrating Are Lower for AI than for Humans

13:37 Outro

OpenAI's latest release of DALL-E 3 integrates with ChatGPT, making it more accessible for people who struggle with prompts. Google's BART Extensions for the Workplace uses a reinforced learning model to pull relevant information from all their Google tools and services, potentially setting it apart from its competitors. Neuralink has opened recruitment for their first-in-human clinical trial for their brain-computer interface, which aims to help those with paralysis control external devices with their thoughts. The research papers discussed in this episode explore the compression capabilities of large language models, a new text generation method called Contrastive Decoding, and Compositional Foundation Models for Hierarchical Planning, which leverages multiple expert foundation models to solve long-horizon tasks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:38 OpenAI releases third version of DALL-E

03:04 Google Announces BART Extensions for the Workplace

04:31 Neuralink’s First-in-Human Clinical Trial is Open for Recruitment

06:02 Fake sponsor

08:06 Language Modeling Is Compression

09:46 Contrastive Decoding Improves Reasoning in Large Language Models

11:24 Compositional Foundation Models for Hierarchical Planning

13:12 Outro

Microsoft's AI research team accidentally exposed 38 terabytes of private data while publishing open-source training data on GitHub, posing a significant security risk. The UK's new AI principles focus on accountability and transparency, seeking views from leading AI developers and governments to ensure the development and use of foundation models evolves in a way that promotes competition and protects consumers. Two new papers explore ways to improve the efficiency and quality of large language models, including a new inference scheme called self-speculative decoding and the ability to prune pretraining data while still retaining performance. A third paper introduces a new type of prompt called the "Chain of Density" or CoD, which generates increasingly dense summaries without increasing their length, resulting in more abstractive and human-preferred summaries.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:45 38TB of data accidentally exposed by Microsoft AI researchers

03:05 UK focuses on transparency and access with new AI principles

04:40 Jason Wei Tweet on the role of task-specific LLMs

06:11 Fake sponsor

07:43 Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding

09:20 When Less is More: Investigating Data Pruning for Pretraining LLMs at Scale

10:55 From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting

12:37 Outro

Amazon is limiting new Kindle books due to the rapid evolution of generative AI, which has flooded the market with low-quality content. DeepMind's co-founder believes that interactive AI is the future, which can carry out tasks by calling on other software and people to get things done. "Compositional Foundation Models for Hierarchical Planning" proposes a solution for effective decision-making in novel environments with long-horizon goals. "Scaling Laws for Sparsely-Connected Foundation Models" explores the impact of parameter sparsity on the scaling behavior of transformers trained on massive datasets, which can lead to more efficient and scalable models in the future.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:42 Citing “rapid evolution of generative AI,” Amazon limits new Kindle books

02:56 DeepMind’s cofounder: Generative AI is just a phase. What’s next is interactive AI.

05:10 Mitigating LLM Hallucinations: a multifaceted approach

06:27 Fake sponsor

08:52 Compositional Foundation Models for Hierarchical Planning

10:39 Replacing softmax with ReLU in Vision Transformers

11:56 Scaling Laws for Sparsely-Connected Foundation Models

13:37 Outro