GPT Reviews

Apple's preference for "machine learning" over "AI," McKinsey and Company's cautious use of generative AI tools, ChatGPT's reasoning capabilities in the Pokemon universe, and SpQR's near-lossless compression for large language models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:31 For better or worse, Apple is avoiding the AI hype train

03:08 McKinsey says ‘about half’ of its employees are using generative AI

05:26 Why AI Will Save the World

07:04 Fake sponsor

09:11 PokemonChat: Auditing ChatGPT for Pokémon Universe Knowledge

10:40 On the Tool Manipulation Capability of Open-source Large Language Models

12:20 SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression

14:41 Outro

The impact of AI on job loss, the new AI-powered feature on Zoom, and two groundbreaking papers on creating digital agents for graphical user interfaces and generating high-quality long-form text.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:26 Nearly 4,000 Jobs Were Lost to AI Last Month, Report Shows

03:09 Zoom can now give you AI summaries of the meetings you’ve missed

05:03 Sterling Crispin on Apple's Vision Pro

06:23 Fake sponsor

08:23 From Pixels to UI Actions: Learning to Follow Instructions via Graphical User Interfaces

10:03 PLANNER: Generating Diversified Paragraph via Latent Language Diffusion Model

11:40 The Surprising Effectiveness of Diffusion Models for Optical Flow and Monocular Depth Estimation

13:56 Outro

AI drone going rogue during a defense conference and public market investors missing out on the AI boom. Our resident AI research expert, Belinda, joins us to discuss "Segment Anything in High Quality," which proposes HQ-SAM, a modification of the Segment Anything Model that accurately segments any object while maintaining efficiency and zero-shot generalizability. We also dive into "Fine-Grained Human Feedback Gives Better Rewards for Language Model Training," which explores how to improve language models by using fine-grained human feedback as an explicit training signal.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:00 AI Drone Goes Rogue

03:25 Y Combinator founder Paul Graham thinks stock market investors are missing out on the AI boom

05:19 localGPT

06:41 Fake sponsor

08:59 Segment Anything in High Quality

10:37 Fine-Grained Human Feedback Gives Better Rewards for Language Model Training

12:25 Evaluating Language Models for Mathematics through Interactions

14:39 Outro

Google's new AI-powered search tool and its potential impact on the online publishing industry, as well as Japan's new policy on AI and copyright laws. We also explore "Discovering New Interpretable Conservation Laws as Sparse Invariants," which proposes an algorithm that can auto-discover conservation laws from differential equations, and "Controllable Text-to-Image Generation with GPT-4," which introduces Control-GPT, greatly boosting the controllability of image generation.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:25 Google’s New AI-Powered Search Is A Beautiful Plagiarism Machine

02:58 Japan Hops On the AI Train

04:37 How important is prompt engineering?

05:57 Fake sponsor

07:59 Discovering New Interpretable Conservation Laws as Sparse Invariants

09:19 Controllable Text-to-Image Generation with GPT-4

10:53 Blockwise Parallel Transformer for Long Context Large Models

12:48 Outro

Risks associated with AI technology, including prompt injection and the potential for AI to cause extinction, language models, with a new optimizer called MeZO proposed for fine-tuning large models and a paper investigating whether language models can identify their own "hallucinations." Additionally, a bias in the evaluation paradigm of using large language models to score the quality of responses generated by other models is uncovered, and two calibration strategies are proposed to address this bias.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:33 ChatGPT Plugins aren’t safe, Prompt Injections

02:48 Statement on AI Risk

04:40 AI is Eating The World

06:04 Fake sponsor

08:38 Fine-Tuning Language Models with Just Forward Passes

10:10 Do Language Models Know When They're Hallucinating References?

11:41 Large Language Models are not Fair Evaluators

13:50 Outro

Meta's Quest 3 headset promising improved AR performance, OpenAI's potential exit from Europe over new AI regulations, the impact of Twitter's algorithm on public opinion and democratic engagement, and a new pre-trained transformer model called BiomedGPT that could improve healthcare outcomes.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:37 Quest 3 hands-on confirms Meta’s building a ‘far thinner and lighter’ headset

03:02 OpenAI Could Quit Europe Over New AI Rules, CEO Sam Altman Warns

05:01 State of GPT

06:01 Fake sponsor

08:13 Twitter's Algorithm: Amplifying Anger, Animosity, and Affective Polarization

10:00 BiomedGPT: A Unified and Generalist Biomedical Generative Pre-trained Transformer for Vision, Language, and Multimodal Tasks

11:32 Large Language Models as Tool Makers

Nvidia's new AI supercomputers and services, the discovery of a new antibiotic using AI, the challenges of implementing AI in specialized fields like healthcare, and a study on Large Language Models' behavior in interactive social settings.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:37 Nvidia debuts new AI supercomputers and services after shares skyrocket

03:01 New superbug-killing antibiotic discovered using AI

04:54 Production AI systems are really hard

06:09 Fake sponsor

08:23 Playing repeated games with Large Language Models

10:09 Demo2Code: From Summarizing Demonstrations to Synthesizing Code via Extended Chain-of-Thought

11:32 Training Socially Aligned Language Models in Simulated Human Society

13:17 Outro

Swiss scientists helping a paralyzed man walk again with the help of a "digital bridge," ChatGPT's life-saving capabilities, Voyager's exceptional proficiency in playing Minecraft, and SPRING's potential for language models in completing sophisticated high-level trajectories in open-world games. These advancements have the potential to revolutionize the way we treat paralysis, learn, and play games, and open the door for the development of generalist agents that can learn and solve tasks in a variety of environments without human intervention.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:23 Swiss Scientists Rebuild Spinal Cord With AI

02:29 TikTok Is Testing an AI Chatbot Named Tako

03:50 ChatGPT is saving my life

05:15 Fake sponsor

07:47 Voyager: An Open-Ended Embodied Agent with Large Language Models

09:52 The False Promise of Imitating Proprietary LLMs

11:23 SPRING: GPT-4 Out-performs RL Algorithms by Studying Papers and Reasoning

12:59 Outro

Microsoft's Windows Copilot, Spotify's use of AI for podcast ads, and Azure AI Studio's capabilities for businesses. We also dive into three research papers on digital chemistry, controllable text-to-video generation, and arithmetic tasks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:02 Microsoft announces Windows Copilot, an AI ‘personal assistant’ for Windows 11

03:46 Spotify may use AI to make host-read podcast ads that sound like real people

05:22 Microsoft’s Azure AI Studio lets developers build their own AI ‘copilots’

07:06 Fake sponsor

09:20 ChemGymRL: An Interactive Framework for Reinforcement Learning for Digital Chemistry

10:59 Control-A-Video: Controllable Text-to-Video Generation with Diffusion Models

12:31 Goat: Fine-tuned LLaMA Outperforms GPT-4 on Arithmetic Tasks

14:36 Outro

Google introduces new AI-powered ad features, while 1X deploys humanoid robots for security and healthcare tasks, potentially addressing labor shortages. Juergen Schmidhuber shares his optimistic perspective on the potential of AI and the fear of a dystopian future. The team also dives into AI language model advancements, including textually pretrained speech language models, efficient finetuning of quantized LLMs, and aligning large language models through synthetic feedback.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:53 Introducing a new era of AI-powered ads with Google

03:22 OpenAI-backed robot startup beats Elon Musk’s Tesla, deploys AI-enabled robots in real world

05:56 Juergen Schmidhuber, Renowned 'Father Of Modern AI,’ Says His Life’s Work Won't Lead To Dystopia

07:24 Fake sponsor

09:23 Textually Pretrained Speech Language Models

10:53 QLoRA: Efficient Finetuning of Quantized LLMs

12:55 Aligning Large Language Models through Synthetic Feedback

14:38 Outro