GPT Reviews

Microsoft plans to sell a new version of Databricks software that helps customers make AI apps for their businesses, potentially hurting OpenAI's business. Businesses should prioritize customer experience over cost reduction when implementing AI, according to an article titled "How NOT to apply Artificial Intelligence in your business". Three AI research papers were discussed, including a multi-agent debate framework for language model evaluation, a curricular subgoal-based framework for inverse reinforcement learning, and a parameter-efficient module operation approach for deficiency unlearning in large language models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:32 Microsoft Plans AI Service With Databricks That Could Hurt OpenAI

02:46 AI news are dire this august

04:12 How NOT to apply Artificial Intelligence in your business

05:37 Fake sponsor

07:37 ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate

09:20 Curricular Subgoals for Inverse Reinforcement Learning

11:08 Separate the Wheat from the Chaff: Model Deficiency Unlearning via Parameter-Efficient Module Operation

12:51 Outro

Google's AI unit, DeepMind, is developing AI tools for life advice, planning, and tutoring, but AI safety experts have concerns about users taking life advice from AI tools. Former Google CEO Eric Schmidt is launching an AI-science moonshot, building a new nonprofit organization to tackle scientific challenges with the help of AI. "Transformers in Reinforcement Learning: A Survey" explores how transformers can be used in reinforcement learning to address unique challenges and improve applications. "Better Zero-Shot Reasoning with Role-Play Prompting" investigates the influence of role-playing on large language models' reasoning abilities and introduces a new methodology called "role-play prompting" that consistently outperforms standard zero-shot approaches across diverse reasoning benchmarks.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:32 Google reportedly building A.I. that offers life advice

03:04 Ex-Google CEO Eric Schmidt to launch AI-science moonshot

04:35 The Batch Newsletter August 16

06:11 Fake sponsor

07:59 Transformers in Reinforcement Learning: A Survey

09:38 Probabilistic Constraint for Safety-Critical Reinforcement Learning

11:04 Better Zero-Shot Reasoning with Role-Play Prompting

13:03 Outro

IBM has unveiled a new prototype of an analog AI chip that works like a human brain, promising to be more efficient and less battery-draining for computers and smartphones. Google has rolled out new search AI features, including the ability to see definitions within AI-generated responses and color-coded syntax highlighting for coding. The paper "Learning to Identify Critical States for Reinforcement Learning from Videos" explores how videos can be used to extract implicit information about rewarding action sequences in deep reinforcement learning, with potential applications in robotics. "Link-Context Learning for Multimodal LLMs" proposes a new approach called Link-Context Learning (LCL) that emphasizes "reasoning from cause and effect" to augment the learning capabilities of Multimodal Large Language Models (MLLMs), with the potential to significantly improve their performance.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:46 IBM unveils an analog AI chip that works like a human brain

03:06 Google Rolls Out New Search AI features

04:53 The Mathematics of Training LLMs — with Quentin Anthony of Eleuther AI

05:54 Fake sponsor

07:38 Learning to Identify Critical States for Reinforcement Learning from Videos

09:27 RAVEN: In-Context Learning with Retrieval Augmented Encoder-Decoder Language Models

10:55 Link-Context Learning for Multimodal LLMs

13:01 Outro

Microsoft's new enterprise spin-off of ChatGPT, the race for scarce Nvidia chips, the remarkable performance of GPT-4 Code Interpreter on challenging math datasets, and a new paper from Google Research comparing the performance of prefixLM and causalLM for in-context learning.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:54 Microsoft Azure ChatGPT allows enterprises to run ChatGPT within their network

03:32 Saudi Arabia, UAE join Elon Musk and Chinese tech titans in the race for scarce Nvidia chips

05:14 AI Town

06:19 Fake sponsor

07:59 Solving Challenging Math Word Problems Using GPT-4 Code Interpreter with Code-based Self-Verification

09:41 OctoPack: Instruction Tuning Code Large Language Models

11:20 CausalLM is not optimal for in-context learning

13:25 Outro

The Federal Election Commission is considering regulating the use of AI-generated content in political ads ahead of the 2024 elections. Zoom has updated its policies to clarify that user data, such as videos, won't be used to train AI models. OpenAI is facing financial struggles and may have to file for bankruptcy by the end of 2024. Three exciting papers were discussed, including a method for building a high-quality instruction-following language model, a comparison of LLM-augmented autonomous agent architectures, and a large dataset of over 16 million multiple sequence alignments for protein structure prediction.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:57 FEC could limit AI in political ads ahead of 2024 elections

03:38 Zoom rewrites its policies to make clear that your videos aren’t used to train AI tools

05:25 OpenAI Might Go Bankrupt by the End of 2024

06:55 Fake sponsor

08:54 Self-Alignment with Instruction Backtranslation

10:09 BOLAA: Benchmarking and Orchestrating LLM-augmented Autonomous Agents

11:39 OpenProteinSet: Training data for structural biology at scale

13:44 Outro

Anthropic has released Claude Instant 1.2, a faster and safer model that outperforms its previous version in math, coding, and safety. Media organizations are calling for regulations to protect copyright in data used to train generative AI models, as it undermines their business models and reduces media diversity. Researchers have made a breakthrough in detecting keystrokes over Zoom calls, using machine learning and microphones to interpret remote keystrokes based on sound profiles of individual keys. The papers discussed in this episode showcase advancements in reinforcement learning for complex games like StarCraft II, language models that critique and refine their own outputs, and metacognitive prompting to improve the understanding abilities of Large Language Models.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:30 Anthropic Releases Claude Instant 1.2

03:01 News outlets demand new rules for AI training data

04:47 AI researchers claim 93% accuracy in detecting keystrokes over Zoom audio

05:48 Fake sponsor

07:40 AlphaStar Unplugged: Large-Scale Offline Reinforcement Learning

09:16 Shepherd: A Critic for Language Model Generation

10:57 Metacognitive Prompting Improves Understanding in Large Language Models

12:44 Outro

Google's new project IDX, Nvidia's bet on AI, and two papers on language models are discussed. The first paper from Google DeepMind explores how simple synthetic data can reduce sycophancy in large language models, while the second paper from Stanford University proposes a new algorithm called staged speculative decoding to speed up the inference of large language models in small-batch, on-device scenarios.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:41 Google Unveils Project IDF

03:20 Nvidia CEO: "We bet the farm on AI and no one knew it"

04:53 Jason Wei Long Tweet on AI research (Researcher at OpenAI)

06:16 Fake sponsor

08:07 Simple synthetic data reduces sycophancy in large language models

09:34 Leveraging Few-Shot Data Augmentation and Waterfall Prompting for Response Generation

10:56 Accelerating LLM Inference with Staged Speculative Decoding

12:45 Outro

Nvidia's new AI chip, the GH200, promises to significantly reduce the cost of running large language models, making AI more accessible for smaller companies. Disney is exploring the use of AI to cut costs in movie and television production, as well as enhance customer support and create unique interactions within its theme parks. The paper "Skills-in-Context Prompting" proposes a novel prompting strategy that significantly improves the compositional generalization capabilities of large language models.  The paper "SILO Language Models" addresses the legal risks associated with training language models on copyrighted or restricted data, proposing a solution that mitigates these risks while maintaining high-quality performance.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:40 Nvidia reveals new A.I. chip, says costs of running LLMs will ‘drop significantly’

03:22 Disney explores cutting costs through AI use

05:06 AI hysteria is a distraction: algorithms already sow disinformation in Africa

06:26 Fake sponsor

08:41 Skills-in-Context Prompting: Unlocking Compositionality in Large Language Models

10:28 SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore

12:23 Current and Future Challenges in Knowledge Representation and Reasoning

14:22 Outro

Microsoft's AI-powered Bing Chat now available on all mobile browsers, Zoom's updated privacy policy, the introduction of AgentBench for evaluating LLMs as agents, and the Flows framework for modeling complex interactions between AI systems and humans. These developments have the potential to lead to more robust and reliable AI models that can perform well in complex, real-world scenarios.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:29 Microsoft’s AI-powered Bing Chat is coming to mobile browsers

02:50 Zoom says its new AI tools aren’t stealing ownership of your content

04:30 Kubernetes Exposed: One Yaml away from Disaster

05:39 Fake sponsor

07:30 AgentBench: Evaluating LLMs as Agents

09:03 Studying Large Language Model Generalization with Influence Functions

10:31 Flows: Building Blocks of Reasoning and Collaborating AI

12:21 Outro

Microsoft is shutting down Cortana and shifting its focus to modern-day AI advances, like its ChatGPT-like Bing Chat and other AI-powered productivity features across Windows and its web browser Edge. Researchers have used AI-powered brain implants to restore movement and sensation for a man who was paralyzed from the chest down, offering life-changing mobility and independence to many. A paper presents a categorical formulation of Predictive Processing and Active Inference using string diagrams, providing a graphical language for understanding these cognitive frameworks with potential implications for robotics, cognitive science, and machine learning. A new approach called "self-translate" leverages the few-shot translation capabilities of multilingual language models themselves, outperforming direct inference and demonstrating important implications for the development of multilingual language models and their use in diverse linguistic settings.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:26 Microsoft kills Cortana in Windows as it focuses on next-gen AI

02:58 Mind Over Paralysis: AI Helps Quadriplegic Man Move and Feel Again

04:48 Twitter thread on Python's Global Interpreter Lock (GIL)

06:05 Fake sponsor

08:04 Tool Documentation Enables Zero-Shot Tool-Usage with Large Language Models

09:30 Active Inference in String Diagrams: A Categorical Account of Predictive Processing and Free Energy

10:51 Do Multilingual Language Models Think Better in English?

12:49 Outro