GPT Reviews

Spotify is testing a new AI-driven playlist creation feature, while Putin was confronted by an AI-generated version of himself. We also delve into a new paper from Google DeepMind that explores the use of vision-language models as sources of rewards for reinforcement learning agents. Finally, we discuss GLEE, an object-level foundation model for locating and identifying objects in images and videos, which exhibits remarkable versatility and improved generalization performance.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:44 Spotify confirms test of prompt-based AI playlists feature

03:02 Putin confronts his AI 'double'

04:35 Twitter Thread on LLM Retrieval Eval

05:35 Fake sponsor

07:44 Vision-Language Models as a Source of Rewards

08:59 Pixel Aligned Language Models

10:24 General Object Foundation Model for Images and Videos at Scale

12:13 Outro

Instagram's new generative AI-powered background editing tool allows users to change the background of their images with fun prompts. FunSearch, a new method for searching for solutions in mathematics and computer science, discovered new solutions for a longstanding open problem in mathematics. The concept of weak-to-strong generalization in AI is explored in a new research direction for superalignment. Finally, the paper "CLIP as RNN" proposes a recurrent framework that enhances mask quality without the need for additional training efforts, setting new state-of-the-art records for image segmentation tasks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:50 Instagram introduces GenAI powered background editing tool

03:18 FunSearch: Making new discoveries in mathematical sciences using Large Language Models

05:23 The AI trust crisis

06:45 Weak-to-strong generalization

08:29 Fake sponsor

10:15 Invariant Graph Transformer

11:47 CLIP as RNN: Segment Countless Visual Concepts without Training Endeavor

13:39 Outro

Axel Springer partners with OpenAI to integrate journalism in AI technologies, enriching the user experience with ChatGPT and providing summaries of selected global news content. Mistral AI raises $415 million in a Series A funding round, aiming to become a European champion in generative artificial intelligence with an open, responsible, and decentralized approach to technology. Microsoft's Medprompt study achieves record-breaking results on the MMLU benchmark, steering GPT-4 with a modified version of Medprompt and providing tools for engineers and customers to achieve similar results. "Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models" proposes a self-training method called ReST$^{EM}$ that goes beyond human data by using scalar feedback, significantly reducing dependence on human-generated data.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:34 Partnership with Axel Springer to deepen beneficial use of AI in journalism

03:05 Mistral AI, a Paris-based OpenAI rival, closed its $415 million funding round

04:59 Steering at the Frontier: Extending the Power of Prompting

06:16 Phi-2: The surprising power of small language models

07:47 Fake sponsor

09:49 Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models

11:12 Interfacing Foundation Models' Embeddings

13:00 Outro

Nvidia is in talks with the Biden administration about permissible sales of AI chips to China, while also facing challenges with ChatGPT-4's performance.

The open-source model, StripedHyena-7B, offers a potential solution for improved training and inference performance over the Transformer architecture.

The papers explore efficient quantization strategies for Latent Diffusion Models, the push for transparency and collaboration in the development of LLMs, and a novel approach for federated full-parameter tuning of billion-sized LLMs.

The episode covers a range of AI topics, from industry news to cutting-edge research, and offers insights into the challenges and potential solutions in the field.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:25 Nvidia in talks with Biden administration about AI chip sales to China, US commerce chief Gina Raimondo says

02:50 As ChatGPT gets “lazy,” people test “winter break hypothesis” as the cause

05:04 Paving the way to efficient architectures: StripedHyena-7B, open source models offering a glimpse into a world beyond Transformers

06:20 Fake sponsor

08:07 Efficient Quantization Strategies for Latent Diffusion Models

09:42 LLM360: Towards Fully Transparent Open-Source LLMs

11:25 Federated Full-Parameter Tuning of Billion-Sized Language Models with Communication Cost under 18 Kilobytes

13:20 Outro

Europe's new AI rules, Cerebras' gigaGPT model, and Mistral AI's Mixtral 8x7B model. The team also discusses two innovative research papers that propose new approaches to enhance multi-step reasoning tasks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:35 Europe reaches a deal on the world’s first comprehensive AI rules

02:40 Introducing gigaGPT: GPT-3 sized models in 565 lines of code

04:44 Release of Mistral 7x8B

06:02 Thread claiming to show prompting breakthrough for GPT

06:57 Fake sponsor

08:38 Localized Symbolic Knowledge Distillation for Visual Commonsense Models

10:10 PathFinder: Guided Search over Multi-Step Reasoning Paths

11:44 Outro

Google's Gemini AI model demo was faked, highlighting the need for skepticism when it comes to tech demos. The "hallucination problem" in language models is not a bug, but rather a feature that allows for creativity and prompts play a significant role in guiding output. Pearl, a production-ready reinforcement learning agent, addresses a range of challenges that real-world intelligent systems encounter and has been adopted by Meta for a recommendation system. Large language models like ChatGPT have the potential to aid professional mathematicians by speeding up and improving the quality of their work. Best practices include fine-tuning LLMs on mathematical data and using them as a tool instead of a replacement for human mathematicians.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:42 Google’s best Gemini demo was faked

03:49 Your guide to AI: December 2023

05:31 Tweet on Hallucination

06:50 Fake sponsor

08:50 Chain of Code: Reasoning with a Language Model-Augmented Code Emulator

10:22 Pearl: A Production-ready Reinforcement Learning Agent

12:00 Large Language Models for Mathematicians

13:46 Outro

Meta's new AI-powered features, the launch of Purple Llama for responsible deployment of generative AI models, H-GAP's state-action trajectory generative model for controlling humanoid robots, and RCG's new image generation framework that sets a new benchmark in class-unconditional image generation.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:56 Meta reveals major AI upgrades

03:28 Announcing Purple Llama: Towards open trust and safety in the new world of generative AI

05:29 AMD MI300 Performance - Faster Than H100, But How Much?

06:53 Fake sponsor

09:02 H-GAP: Humanoid Control with a Generalist Planner

10:33 Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia

13:02 Self-conditioned Image Generation via Generating Representations

14:54 Outro

Meta and IBM have launched an 'AI Alliance' to promote open-source AI development, while Musk's AI startup seeks to raise $1 billion to compete with OpenAI. The paper "Training Chain-of-Thought via Latent-Variable Inference" explores improving language models, and "Rank-without-GPT" builds GPT-independent listwise rerankers on open-source large language models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:25 Meta and IBM launch ‘AI Alliance’ to promote open-source AI development

02:59 Musk's AI startup seeks to raise $1 bn

04:22 Fake sponsor

06:22 Training Chain-of-Thought via Latent-Variable Inference

08:01 WhisBERT: Multimodal Text-Audio Language Modeling on 100M Words

09:32 Rank-without-GPT: Building GPT-Independent Listwise Rerankers on Open-Source Large Language Models

11:26 Outro

Runway ML and Getty Images' partnership to create a new AI video model, the general availability of Microsoft Copilot, the Magicoder series of fully open-source LLMs for code, and the Unlocking Spell on Base LLMs' new tuning-free alignment method called URIAL.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:34 Runway and Getty Images team up on AI video

02:55 Microsoft Copilot Now Available to All Users

04:41 GPU Cloud Economics Explained – The Hidden Truth

05:53 Fake sponsor

07:55 Object Recognition as Next Token Prediction

09:08 Magicoder: Source Code Is All You Need

10:46 The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning

12:47 Outro

Google's postponement of the launch of Gemini, Apple's "Generating Molecular Conformer Fields" paper achieving state-of-the-art performance on molecular conformer generation, "Mamba: Linear-Time Sequence Modeling with Selective State Spaces" proposing a new model that allows for content-based reasoning, and EPFL's "Instruction-tuning Aligns LLMs to the Human Brain" paper finding that instruction-tuning generally enhances brain alignment.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:30 Google postpones launch of GPT-4 rival Gemini

03:11 AI and Trust

04:41 LLM Visualization Tool

05:35 Fake sponsor

07:25 Generating Molecular Conformer Fields

08:32 Mamba: Linear-Time Sequence Modeling with Selective State Spaces

10:18 Instruction-tuning Aligns LLMs to the Human Brain

12:18 Outro