GPT Reviews

Anthropic raises $450 million in Series C funding to scale their reliable AI products, while Waymo and Uber partner to bring Waymo's autonomous driving technology to the Uber platform. We also explore two fascinating papers: RWKV, which combines the efficiency of transformers with the inference of RNNs, and RecurrentGPT, which enables interactive generation of arbitrarily long text without forgetting.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:53 Anthropic Raises $450 Million in Series C Funding to Scale Reliable AI Products

03:11 Waymo and Uber partner to bring Waymo’s autonomous driving technology to the Uber platform

05:26 How Rogue AIs may Arise

06:28 Fake sponsor

08:56 RWKV: Reinventing RNNs for the Transformer Era

10:37 "According to ..." Prompting Language Models Improves Quoting from Pre-Training Data

11:58 RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text

14:11 Outro

Meta's €1.2 billion GDPR fine over US mass surveillance is a major blow for the company, while LIMA, a language model trained with only 1,000 prompts and responses, demonstrated strong performance. Composable Diffusion (CoDi) is a powerful generative model that can handle any combination of input and output modalities, and CRITIC is a framework that allows large language models to self-correct with external feedback.

Contact:  [email protected]

Timestamps:

00:34 Introduction

01:44 noyb win: € 1.2 billion fine against Meta over EU-US data transfers

04:21 GPT-4 didn't really score 90th percentile on the bar exam

05:43 Governance of superintelligence

07:08 Fake sponsor

09:42 LIMA: Less Is More for Alignment

11:31 Any-to-Any Generation via Composable Diffusion

13:18 CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing

14:48 Outro

Apple is limiting the internal use of some AI-powered tools, while Hippocratic AI has raised $50 million to develop healthcare chatbots. We also discuss a new approach to teaching AI called Guidance and a new large language model called Med-PaLM 2 that improves upon previous work in medical question answering.

Contact:  [email protected]

Timestamps:

566:54 Introduction

1905:32 Apple reportedly limits internal use of AI-powered tools like ChatGPT and GitHub Copilot

3376:42 Hippocratic AI Raises $50 Million To Power The Healthcare Bot Workforce

5500:37 Guidance: an alternative to prompting or chaining

6601:09 Fake sponsor

8624:46 A Generalist Dynamics Model for Control

9843:34 GPT detectors are biased against non-native English writers

11499:34 Towards Expert-Level Medical Question Answering with Large Language Models

13531:16 Outro

The ChatGPT app for iOS, Meta's breakthrough in AI accelerator technology with MTIA v1, and an article that explores the future of language and communication. We also discuss three fascinating AI research papers, including "TinyStories," "SLiC-HF," and "Bot or Human?" which presents a new way for online service providers to protect themselves against nefarious activities.

Contact: [email protected]

Timestamps:

00:34 Introduction

01:48 Introducing the ChatGPT app for iOS

03:13 MTIA v1: Meta’s first-generation AI inference accelerator

05:34 Life After Language

07:13 Fake sponsor

09:34 TinyStories: How Small Can Language Models Be and Still Speak Coherent English?

11:18 SLiC-HF: Sequence Likelihood Calibration with Human Feedback

12:54 Bot or Human? Detecting ChatGPT Imposters with A Single Question

14:53 Outro

Apple's AI-Driven Accessibility Updates and Google's PaLM 2. We also delve into the world of efficient audio generation with SoundStorm and explore the potential of active retrieval augmented generation. Finally, we discuss the importance of in-context learning in NLP and how Pre-Training to Learn in Context can enhance language models' abilities.

Contact: [email protected]

Timestamps:

00:34 Introduction

01:45 Apple's AI-Driven Accessibility Updates Include Text-to-Speech That Can Mimic Your Voice

03:10 PaLM 2 Details Leaked: it uses 5x the data the original PaLM trained on

05:32 Numbers every LLM Developer should know

06:48 Fake sponsor

08:38 SoundStorm: Efficient Parallel Audio Generation

09:58 Active Retrieval Augmented Generation

11:30 Pre-Training to Learn in Context

13:26 Outro

We discussed Sam Altman's proposed solutions for regulating AI's impact on elections, as well as the exciting partnership between Anthropic and Zoom to build customer-centric AI products. We also highlighted the LLM University by Cohere, a comprehensive learning resource for anyone interested in NLP using language models. Finally, we delved into three research papers, including Optimizing Memory Mapping Using Deep Reinforcement Learning, Professional Certification Benchmark Dataset, and Symbol Tuning for ICL in Large Language Models.

Contact: [email protected]

Timestamps:

00:34 Introduction

01:46 Sam Altman is concerned about AI being used to compromise elections

03:03 Anthropic Announces Partnership and Investment from Zoom

05:43 The LLM University by Cohere — Your Go-To Learning Resource for NLP🎓

06:59 A repo to build a small junior AI developer

08:43 Fake sponsor: AccenTrick Energy Drink

10:57 Optimizing Memory Mapping Using Deep Reinforcement Learning

12:37 Professional Certification Benchmark Dataset: The First 500 Jobs For Large Language Models

14:47 Symbol tuning improves in-context learning in language models

17:10 Outro

The EU AI Act's potential impact on American tech companies, Meta's impressive new model architecture for better ad experiences, a multi-scale decoder architecture for modeling long sequences, and the need for best practices and regulations for AGI labs.

Contact: [email protected]

Timestamps:

00:34 Introduction

01:36 EU AI Act To Target US Open Source Software

03:16 New AI advancements drive Meta’s ads system performance and efficiency

05:29 Brex's Prompt Engineering Guide

06:41 Why we need uncensored models and how to train them

07:45 Fake sponsor: Scrub Busters

09:51 MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers

11:21 Towards best practices in AGI safety and governance: A survey of expert opinion

12:49 A Suite of Generative Tasks for Multi-Level Multimodal Webpage Understanding

14:28 Outro

Meta introduces generative AI features for advertisers, while Yuval Noah Harari and Yann LeCun debate the promises and dangers of AI. The WebLLMs project allows users to run large-language models and LLM-based chatbots directly in their browsers, and the ImageBind project learns a joint embedding across six different modalities.

Contact: [email protected]

Timestamps:

00:34 Introduction

02:07 Meta announces generative AI features for advertisers

03:28 Debate highlights: Yuval Noah Harari (Sapiens) versus Yann Le Cun (Meta) on artificial intelligence

04:38 WebLLMs project: running Language Models in your browser

06:17 Language models can explain neurons in language models

07:35 Fake sponsor: Sobstopper Tissues from Super Soft Inc.

09:44 ImageBind: One Embedding Space To Bind Them All

11:29 An Inverse Scaling Law for CLIP Training

13:07 Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision

15:33 Outro

Anthropic's Claude introduces 100k tokens context windows for businesses to analyze complex documents quickly, while Huggingface releases Transformers Agents, an experimental API for natural language processing and task completion. Stability AI also releases Stable Animation SDK, a powerful text-to-animation tool for artists and developers. Additionally, researchers propose new AI models and frameworks, including Federated Instruction Tuning, Evaluating Embedding APIs for Information Retrieval, and Pretraining Without Attention.

Contact: [email protected]

Timestamps:

00:34 Introduction

01:26 Anthropic's Claude introduces 100k tokens context Windows

02:54 Huggingface Releases Transformers Agents

04:10 Stability AI releases Stable Animation SDK, a powerful text-to-animation tool for developers

05:39 Microsoft makes strategic investment into Builder.ai, integrates its services into Teams

07:16 Fake sponsor: SlimDown

09:19 Towards Building the Federated GPT: Federated Instruction Tuning

10:52 Evaluating Embedding APIs for Information Retrieval

12:24 Pretraining Without Attention

14:00 Outro

GPT Reviews Artificial Intelligence on Thursday, May 11, 2023

Google's use of generative AI in core products, a controversial AI version of a Snapchat influencer, advancements in language and video understanding with Google's PaLM 2 language model, VideoChat, the potential for unfaithful explanations in large language models revealed by NYU, and more.

Contact: [email protected]

Timestamps:

00:34 Introduction

01:22 Google IO is held with heavy focus on AI

03:00 Google finally demos generative AI in Search, with a waitlist starting today

04:08 A 23-year-old Snapchat influencer used OpenAI’s technology to create an A.I. version of herself that will be your girlfriend for $1 per minute

05:33 A closer look at Claude's Constitution

07:23 Fake sponsor

09:24 PaLM 2 Technical Report

11:03 VideoChat: Chat-Centric Video Understanding

12:49 Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting

14:55 Outro