GPT Reviews

Apple's investment in generative AI and CoreWeave's $2.3 billion loan for cloud infrastructure. We also dive into two research papers, "Retroformer" and "MM-Vet," which explore optimizing language agents and evaluating large multimodal models, respectively.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:42 Apple has been quiet about ChatGPT. Now Tim Cook says its hefty $22.6 billion research spend is down to generative AI.

02:58 CoreWeave, which provides cloud infrastructure for AI training, secures $2.3B loan

05:17 Analysis: While AI takes the spotlight, infrastructure stocks shine

06:56 Fake sponsor

08:53 Retroformer: Retrospective Large Language Agents with Policy Gradient Optimization

10:11 MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities

11:49 Training Data Protection with Compositional Diffusion Models

13:42 Outro

OpenAI's trademark application for GPT-5 could signify a continued advancement in natural language processing and machine learning. Google's "mind-reading" AI raises ethical concerns about potential implications and future uses of the technology. Soft MoE, a new type of mixture of expert architecture proposed by Google DeepMind, outperforms standard Transformers and popular MoE variants in visual recognition. PerceptionCLIP, a method proposed by the University of Maryland and the Bosch Center for Artificial Intelligence, improves zero-shot image classification by inferring and conditioning on contextual attributes, achieving better generalization, group robustness, and interpretability.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:23 OpenAI Files Trademark Application For GPT-5

02:44 Google's 'mind-reading' AI can tell what music you listened to based on your brain signals

04:26 Nvidia H100 GPUs: Supply and Demand

05:17 Fake sponsor

07:24 From Sparse to Soft Mixtures of Experts

09:11 More Context, Less Distraction: Visual Classification by Inferring and Conditioning on Contextual Attributes

10:47 Music De-limiter Networks via Sample-wise Gain Inversion

12:45 Outro

YouTube is experimenting with AI-generated video summaries, potentially changing the way creators structure their videos. AudioCraft, a generative AI for audio, is now available to all and could revolutionize the way we create and experience audio. PointOdyssey, a synthetic dataset and data generation framework, aims to advance the state-of-the-art in long-term fine-grained tracking algorithms. A new attack method on aligned language models demonstrates the vulnerability of even aligned models to adversarial attacks, raising important questions about preventing objectionable content.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:33 YouTube uses AI to summarize videos in latest test

02:58 Open sourcing AudioCraft: Generative AI for audio made simple and available to all

04:52 Show HN: Learn a language quickly by practising speaking with AI (prettypolly.app)

05:50 Fake sponsor

07:54 PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking

09:35 Three Bricks to Consolidate Watermarks for Large Language Models

11:14 Universal and Transferable Adversarial Attacks on Aligned Language Models

13:39 Outro

Billionaire CEO of military technology supplier Palantir advocates for AI weapons, sparking controversy and raising questions about the risks and benefits of such technology.

Meta is developing AI-powered chatbots with different personalities, which could collect vast amounts of data on users' interests, but also raises concerns around privacy and potential manipulation.

Google is planning to update Assistant with features powered by generative AI, which could allow it to answer questions based on information gleaned from across the web, but also raises potential privacy implications.

Cutting-edge AI research papers have been published, including one on developing adaptable control policies for autonomous robots and another on improving the tool-use capabilities of open-source large language models. These papers have important implications for the future of AI technology.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:46 Billionaire CEO of military technology supplier Palantir advocates for AI weapons: 'We must not grow complacent'

03:24 Meta prepares chatbots with personas to try to retain users

04:32 Google will ‘supercharge’ Assistant with AI that’s more like ChatGPT and Bard

06:13 Fake sponsor

08:14 Discovering Adaptable Symbolic Algorithms from Scratch

09:46 ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs

11:43 MovieChat: From Dense Token to Sparse Memory for Long Video Understanding

13:16 Outro

The RT-2 model translates vision and language into action, showing improved generalization capabilities and semantic and visual understanding beyond the robotic data it was exposed to.

OverflowAI is a new space for Stack Overflow's community and customers to explore the future of knowledge sharing together, featuring semantic search, enterprise knowledge ingestion, Slack integration, a Visual Studio Code extension, and AI community discussions.

"Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback" surveys the fundamental limitations and open problems of RLHF, as well as techniques to improve and complement it in practice.

"Robust Distortion-free Watermarks for Language Models" proposes a way to plant watermarks in text generated by language models that are robust to perturbations without changing the distribution over text up to a certain maximum generation budget.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:43 RT-2: New model translates vision and language into action

03:17 Announcing OverflowAI, the future of community & AI

04:57 Computer Scientists Discover Limits of Stochastic Gradient Descent

06:00 Fake sponsor

07:55 Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback

09:39 Uncertainty in Natural Language Generation: From Theory to Applications

11:59 Robust Distortion-free Watermarks for Language Models

13:37 Outro

Netflix's controversial AI job ad, Google's impressive Q2 earnings, and advancements in autonomous agents and large language models for code. The episode also features discussions on a realistic web environment for building autonomous agents and a new framework for boosting pre-trained Code LLMs.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:41 Netflix touts $900k AI jobs amid Hollywood strikes

02:55 Google stock jumped 10% this week, fueled by cloud, ads and hope in A.I.

04:44 Worldcoin: a solution in search of its problem

05:52 Fake sponsor

08:00 WebArena: A Realistic Web Environment for Building Autonomous Agents

09:24 Evaluating the Moral Beliefs Encoded in LLMs

11:02 PanGu-Coder2: Boosting Large Language Models for Code with Ranking Feedback

12:37 Outro

A new partnership promoting responsible AI, the release of a new text-to-image model with ethical safeguards, a framework for detecting factual errors in generated text, and a framework for high-quality object tracking in videos.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:32 A new partnership to promote responsible AI

03:18 Stability AI releases its latest image-generating model, Stable Diffusion XL 1.0

05:27 No One Wants To Talk To Your Chatbot

06:28 Fake sponsor

08:45 FacTool: Factuality Detection in Generative AI -- A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios

10:09 Tracking Anything in High Quality

11:59 Towards Generalist Biomedical AI

13:59 Outro

Japan's Ministry of Economy, Trade, and Industry is investing heavily in AI development by building a new supercomputer to accelerate progress in AI and reduce Japan's dependence on foreign countries. GitHub and other companies are calling for more open-source support in EU AI law to promote a more open ecosystem for AI innovation. The ARB benchmark presents a more challenging test for large language models, and current models score well below 50% on more demanding tasks. "Predicting Code Coverage without Execution" proposes using Machine Learning to predict code coverage without actual execution, which could lower the cost of code coverage. State-of-the-art LLMs were able to predict code coverage with reasonable accuracy, demonstrating the potential of using Machine Learning for this task.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:40 Japan's METI to build new supercomputer to help develop AI at home

03:28 GitHub and others call for more open-source support in EU AI law

05:06 ChatGPT broke the Turing test — the race is on for new ways to assess AI

06:26 Fake sponsor

08:30 ARB: Advanced Reasoning Benchmark for Large Language Models

09:55 Contrastive Example-Based Control

11:16 Predicting Code Coverage without Execution

13:14 Outro

The launch of Worldcoin at the intersection of AI, identity, and finance, to the Android release of OpenAI's ChatGPT, this episode explores the cutting-edge of AI-powered conversational tools. The paper "Evaluating the Ripple Effects of Knowledge Editing in Language Models" proposes a new evaluation benchmark that could have important implications for improving the accuracy of language models. Finally, the Retentive Network, a new architecture proposed as a successor to the Transformer architecture for large language models, achieves favorable scaling results, parallel training, low-cost deployment, and efficient inference.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:45 OpenAI CEO Sam Altman Launches Worldcoin: A Bold Crypto Experiment At The Intersection Of AI, Identity And Finance

03:13 ChatGPT for Android launches next week

04:52 Michaël Benesty on Speeding up LLAMA v2 inference

06:30 Fake sponsor

08:19 Evaluating the Ripple Effects of Knowledge Editing in Language Models

09:50 ZeroQuant-FP: A Leap Forward in LLMs Post-Training W4A8 Quantization Using Floating-Point Formats

11:28 Retentive Network: A Successor to Transformer for Large Language Models

13:15 Outro

two new open-source Large Language Models, the departure of OpenAI's head of trust and safety, the introduction of STEVE-1, a model that can follow a wide range of instructions in Minecraft, and a paper exploring ways to protect the copyright of Neural Radiance Fields.

Contact:  [email protected]

Timestamps:

00:34 Introduction

02:03 Stability AI Releases 2 Language Models

03:20 OpenAI’s head of trust and safety Dave Willner steps down

04:59 Twitter Thread by Anthropic on Faithfulness in Chain of thought reasoning

06:12 Fake sponsor

08:19 Invalid Logic, Equivalent Gains: The Bizarreness of Reasoning in Language Model Prompting

09:54 STEVE-1: A Generative Model for Text-to-Behavior in Minecraft

11:36 CopyRNeRF: Protecting the CopyRight of Neural Radiance Fields

13:01 Outro