GPT Reviews

Google's Duet AI introducing new features for productivity software, growing public concern about the role of AI in daily life, and Andrej Karpathy's tweet about an optimization technique for inference-time with LLMs. Additionally, the episode features three fascinating papers: RLAIF, CityDreamer, and FACET, which cover topics such as reinforcement learning, 3D city generation, and fairness in computer vision evaluation.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:33 Google Duet AI: New Features for Gmail, Docs and Sheets, at $30 a Month

03:20 Growing public concern about the role of artificial intelligence in daily life

04:58 Andrej Karpathi tweets

06:34 Fake sponsor

08:20 RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback

10:05 CityDreamer: Compositional Generative Model of Unbounded 3D Cities

11:40 FACET: Fairness in Computer Vision Evaluation Benchmark

13:34 Outro

China's Baidu has rolled out a new AI chatbot called ERNIE, which is being touted as a rival to OpenAI's ChatGPT. The US has expanded restrictions on the export of Nvidia AI chips beyond China to some countries in the Middle East, which could have serious implications for China's AI industry. OpenAI has released GPT-3.5-Turbo for fine-tuning natural language to SQL, which allows for even more advanced NL-to-SQL models to be created. A new paper proposes LLaSM, a large multi-modal speech-language model that can follow speech-and-language instructions, which could provide a more natural way for humans to interact with artificial intelligence.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:11 China’s Baidu rolls out ChatGPT rival ERNIE to public

03:49 US restricts exports of Nvidia AI chips to Middle East

05:29 Fine-tuning GPT-3.5-Turbo for Natural Language to SQL

06:32 Fake sponsor

08:20 General Purpose Audio Effect Removal

09:54 RoboTAP: Tracking Arbitrary Points for Few-Shot Visual Imitation

11:10 LLaSM: Large Language and Speech Model

13:00 Outro

Google Cloud's latest AI updates and partnerships, to the US Copyright Office's call for opinions on AI and copyright issues, this episode is packed with insights and analysis. The show also features papers on vector search and the security risks of generative AI, highlighting the need for ethical guidelines and methods for detecting and mitigating these risks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:32 Google AI Updates from their Developer Keynote

02:45 US Copyright Office wants to hear what people think about AI and copyright

04:36 ChatLZMA – text generation from data compression

05:54 Fake sponsor

08:03 Vector Search with OpenAI Embeddings: Lucene Is All You Need

09:23 When Do Program-of-Thoughts Work for Reasoning?

11:09 Identifying and Mitigating the Security Risks of Generative AI

13:01 Outro

OpenAI has launched ChatGPT Enterprise, an AI chatbot with enterprise-grade security and privacy, unlimited higher-speed GPT-4 access, longer context windows, and advanced data analysis capabilities. Some websites, including Amazon, Quora, and NY Times, have already blocked OpenAI's GPTBot, a language model that can generate text indistinguishable from human writing. Researchers from the Australian National University have proposed a dual-stage optimization technique using adversarial fine-tuning to generate and detect problematic content in large language models. LongBench, the first bilingual, multi-task benchmark for long context understanding, provides a more rigorous evaluation of long context understanding, which is crucial for developing language models that can handle longer sequences.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:54 Introducing ChatGPT Enterprise

03:35 Websites That Have Blocked OpenAI’s GPTBot - 1000 Website Study

05:21 Ahead of AI #11: New Foundation Models

06:27 Fake sponsor

08:48 Adversarial Fine-Tuning of Language Models: An Iterative Optimisation Approach for the Generation and Detection of Problematic Content

10:35 LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding

12:36 ORES: Open-vocabulary Responsible Visual Synthesis

14:35 Outro

Alibaba's new AI models, Tesla's "Full Self-Driving" software, and Google's new Gemini model. The papers discussed also provide valuable insights into evaluating safeguards in LLMs, improving the quality of generated code, and large vision-language models with versatile abilities.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:38 Alibaba launches chatGPT competitor that can understand images and have more complex conversations

03:06 Elon Musk demoes Tesla's AI-only self driving

05:06 Google Gemini Eats The World – Gemini Smashes GPT-4 By 5X, The GPU-Poors

06:14 Fake sponsor

08:39 Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities

09:50 Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs

11:39 Does Asking Clarifying Questions Increases Confidence in Generated Code? On the Communication Skills of Large Language Models

13:54 Outro

Hugging Face, an AI startup, has raised a massive $235 million in funding from big tech companies like Google, Amazon, and IBM, highlighting the demand for collaborative AI development. The Anti-hype LLM reading list is a valuable resource for anyone interested in understanding language models from a practical perspective, providing links to reasonable and good explanations of how things work, with no hype or vendor content. Large language models are being utilized in the development of autonomous agents, which can better understand natural language, generate human-like responses, and adapt to new situations. The Giraffe paper explores the limitations of fixed context lengths in large language models and introduces new strategies for context length extrapolation, which can greatly improve their performance on longer sequences.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:05 Huggingface Receives $235M in funding from big tech

03:59 Anti-hype LLM reading list

05:26 Tweet by Cameron R. Wolfe

06:39 Fake sponsor

08:45 A Survey on Large Language Model based Autonomous Agents

10:23 Instruction Tuning for Large Language Models: A Survey

11:46 Giraffe: Adventures in Expanding Context Lengths in LLMs

13:54 Outro

Code Llama, a new state-of-the-art large language model for coding, and Microsoft's plans to add AI capabilities to apps like Paint and Photos on Windows 11. We also explore Ludwig, a low-code framework for building custom AI models, and WizardMath, a new model that improves the mathematical reasoning abilities of large language models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:34 Introducing Code Llama, a state-of-the-art large language model for coding

03:24 Microsoft may bring AI capabilities to apps like Paint and Photos on Windows 11

05:24 Ludwig: a low-code framework for building custom AI models like LLMs and other deep neural networks.

06:48 Fake sponsor

08:53 Bayesian Flow Networks

10:13 The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only

11:59 WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct

13:44 Outro

Meta introduces SeamlessM4T, a multimodal AI model for speech and text translations that supports nearly 100 languages. OpenAI announces fine-tuning for GPT 3.5 Turbo, allowing businesses to customize the model for unique user experiences. The Backyard Emporium offers Miracle Grow, a product that makes plants grow super fast. Three AI research papers are discussed, covering adversarial robustness of multi-modal foundation models, causal inference learning, and motion-guided masking for video masked autoencoding.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:30 Meta Introduces SeamlessM4T, a Multimodal AI Model for Speech and Text Translations

03:10 OpenAI Announces Finetuning for GPT 3.5 Turbo

04:42 Mixture of Experts Open Project

05:41 Fake sponsor

07:33 On the Adversarial Robustness of Multi-Modal Foundation Models

08:46 Active and Passive Causal Inference Learning

10:16 MGMAE: Motion Guided Masking for Video Masked Autoencoding

12:10 Outro

YouTube's partnership with music labels to establish rules for AI-generated content, IBM's study on the impact of AI on the workforce, and two papers on language modeling and safety alignment. The episode also explores Reinforced Self-Training (ReST), a new reinforcement learning algorithm that improves the quality of large language models, and "Adapting Learned Sparse Retrieval for Long Documents," which proposes adaptations to improve the performance of Learned Sparse Retrieval (LSR) for long documents.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:50 YouTube is figuring out its AI strategy by working with music labels

03:26 40% of workers will have to reskill in the next three years due to AI, says IBM study

05:24 Jason Wei tweets

06:39 Fake sponsor

08:46 Reinforced Self-Training (ReST) for Language Modeling

10:13 Red-Teaming Large Language Models using Chain of Utterances for Safety-Alignment

11:59 Adapting Learned Sparse Retrieval for Long Documents

14:10 Outro

From the launch of BeFake, an AI-based social media app, to a ruling that AI-created art is not copyrightable, the show explores the latest developments in the field. The episode also features two research papers that propose new approaches to deep neural networks and machine unlearning, as well as a paper that introduces a general approach to personalized text generation inspired by writing education.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:26 Ex-MZ CEO launches BeFake, an AI-based social media app

02:59 AI-Created Art Isn’t Copyrightable, Judge Says in Ruling That Could Give Hollywood Studios Pause

04:35 LlamaGPT

05:52 Fake sponsor

07:35 Brain-Inspired Computational Intelligence via Predictive Coding

09:07 Fast Machine Unlearning Without Retraining Through Selective Synaptic Dampening

10:54 Teach LLMs to Personalize -- An Approach inspired by Writing Education

12:48 Outro