Altman Reinstated at OpenAI

Author

AI Breakfast
November 22, 2023

Sponsored by

Good morning. It’s Wednesday, November 22nd.

In today’s email:

  • Altman, Brockman Return to OpenAI

  • Advancements in AI Technology

  • AI in Business and Workforce Development

  • 8 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Support this newsletter by becoming a Premium Subscriber for just $9/month

Today’s trending AI news stories

Altman Reinstated at OpenAI

In a striking reversal, OpenAI has announced the reinstatement of Sam Altman as its CEO, ending a period of intense internal conflict. This change follows a series of dramatic events, including a failed attempt by the previous board to retain control and widespread employee protests advocating for Altman's return.

The reappointment is accompanied by the introduction of a new board with distinguished members such as Bret Taylor (CEO of Salesforce), Larry Summers (former US Secretary of the Treasury), and Adam D’Angelo (CEO of Quora), marking a significant shift in the organization's leadership landscape.

Microsoft, a key investor and close collaborator of OpenAI, has expressed its interest in having representation on this newly formed board, highlighting the intertwined futures of the two influential entities. Additionally, an independent investigation is set to be launched to examine the circumstances that led to these recent upheavals.

OpenAI Co-founder Greg Brockman also posted a photo of himself and large gathering of employees with the title “We are so back” in the early hours of the morning, signaling his return to the company.

Yesterday, Elon Musk tweeted a link to a letter that appeared to be an insider’s summary of the reasons behind Altman’s termination, citing hostile work work environments and a clash in mission over the for-profit/non-profit model, though the letter has since been deleted.

In partnership with

Building an iconic brand may not be as hard as you think. With Scott Galloway’s Brand Strategy sprint, you’ll get an unfair advantage over your competition.

In this sprint, Scott breaks down practical insights from companies like Apple, Nike, & Uber to help you apply the same frameworks to your own brand.

For 30 days only, you can get The Brand Strategy sprint, normally $179, for FREE, when you use the discount code: SCOTT100.

Advancements in AI Technology

> Anthropic unveiled Claude 2.1 as OpenAI drama captured the media spotlight. Claude 2.1 boasts major improvements in accuracy, honesty, and technical capabilities, including a 200,000 token context window for processing up to 500-page documents. It reduces hallucinations and false claims, integrates with internal systems, and has enhanced summarization and comprehension abilities.

> MIT researchers have developed a system called StableRep, which employs synthetic images generated by text-to-image models for machine learning training. This method has shown to outperform traditional training using real images. StableRep utilizes a unique "multi-positive contrastive learning" strategy, enabling models to understand high-level concepts more deeply. The approach is efficient and highlights the potential of synthetic imagery in AI training, though it still begins with real data.

> UC Berkeley Researchers Introduce Ghostbuster: A SOTA AI Method for Detecting LLM-Generated Text. Ghostbuster utilizes a three-stage process involving probability computation, feature selection, and classifier training. It significantly outperforms existing models in detecting AI-generated content, achieving a 97.0 F1 score across various conditions. This tool addresses concerns about the authenticity of AI-generated text, particularly in academic and journalistic contexts.

> Stability AI introduced Stable Video Diffusion, an AI model for animating images into videos, based on its Stable Diffusion model. Available in "research preview," it's a rare open-source video-generating tool. The model offers two versions: SVD for 14-frame videos and SVD-XT for 24 frames, both with variable frame rates. Amid financial strains, this launch marks Stability AI's strategic move into the AI video generation arena.

AI in Business and Workforce Development

> Amazon has unveiled its "AI Ready" program, pledging to deliver free AI skills training to two million people worldwide by 2025. The initiative, responding to a growing demand for AI expertise in the job market, aims to democratize access to AI education, enabling a wider pool of people to gain skills in this transformative technology and enhance their earning potential.

> A think tank study suggests that by 2033, AI could enable a four-day workweek for 28% of workers in Britain and the US. The report by Autonomy indicates that AI-driven productivity gains could reduce weekly working hours from 40 to 32 while maintaining pay and performance. Large Language Models like ChatGPT could be used to augment worker roles, potentially avoiding mass unemployment and reducing mental and physical illnesses.

> Baidu's AI chatbot, Ernie Bot, reached 70 million users in three months, boosting Baidu's finances. Integrated into Baidu's main products, it requires an $8 monthly fee. However, it trails OpenAI's ChatGPT, which had 1.7 billion visits in October 2023. Baidu's founder, Robin Li, criticizes China's focus on developing numerous AI models, advocating instead for AI applications with widespread user adoption.

> Nvidia has reported a record Q3 revenue of $18.2 billion, a 206% year-over-year increase, driven predominantly by its data center segment which grew 279% annually. This growth is attributed to the industry's shift from general-purpose to accelerated computing and generative AI. Nvidia's GAAP earnings per share were $3.71, showing significant growth from the previous year. The company's outlook for Q4 is optimistic, with expected revenue around $20 billion.

8 new AI-powered tools from around the web

Askpot, an automated competitive analysis tool, swiftly extracts key business insights and generates SWOT tables. It focuses on qualitative analysis, offering customization, structured results, and history tracking.

Twip AI, a content creation tool, enhances YouTube workflows using AI for personalized scripts and text-to-voice conversion. It suggests high-performing content ideas, manages ideas backlog, and optimizes video metadata for improved YouTube visibility.

SubMagic Pro is a free, online video caption generator supporting 99 languages, offering customizable fonts, colors, and sizes for captions. It handles large files, real-time editing, and provides project saving and downloading features.

ShipGPT AI provides a ready-to-use AI development platform, enabling easy integration and creation of AI applications without the need for specialized developers, featuring conversational chatbots, time-saving tools, and a continually updated open-source stack.

LLM Spark is a development platform designed for building production-ready Large Language Model applications, offering features like prompt testing, data source connectivity, version control, team collaboration, instant deployment, and observability with advanced analytics.

MyLens is a free AI tool that generates detailed timelines for any topic, allowing users to combine multiple topics and view overlapping events in a single, cohesive timeline.

ScrapetoAI is a user-friendly tool for extracting data from websites, converting it into JSON or CSV formats. It integrates with OpenAI, offers a Chrome extension for easy scraping, and supports saving data to platforms like Google Sheets, Airtable, and Zapier.

VIZ-X Landing Page Optimizer is a tool offering automated analysis and actionable design insights for websites, focusing on enhancing design, usability, and user experience. It aims to improve user engagement, conversion rates, elevate brand image, and create an intuitive user journey.

arXiv is a free online library where researchers share pre-publication papers.

AutoStory is an automated system for generating diverse, high-quality storytelling images with minimal human effort. It combines large language models for layout planning and text-to-image models for generating detailed story scenes. The system efficiently handles complex story visualizations with minimal user interaction, offering user-friendly features like sparse control for layout and dense control for image quality.

Orca 2, developed by Microsoft Research, enhances the reasoning capabilities of smaller language models. It teaches these models diverse reasoning strategies, such as step-by-step analysis and recall-then-generate. Outperforming its predecessors, Orca 2 shows significant gains in complex tasks, matching or even surpassing models 5-10 times its size in zero-shot settings.

FastBERT, a BERT variant, significantly speeds up language modeling by utilizing only 0.3% of its neurons (12 out of 4095 per layer) during inference. This efficiency is achieved by replacing traditional feedforward networks with fast feedforward networks (FFFs). FastBERT matches the performance of similar BERT models while demonstrating a 78x speedup over optimized CPU baselines and a 40x speedup on PyTorch GPU implementations.

ToolTalk is a benchmark for evaluating tool-augmented large language models (LLMs) in complex, multi-step tasks specified through dialogue. It comprises 28 tools across 7 categories with simulated implementations for automated evaluation. Testing GPT-3.5 and GPT-4 on ToolTalk yielded success rates of 26% and 50%, respectively, highlighting improvement areas for LLMs with tool integration.

GPT4Motion presents a novel training-free framework for text-to-video synthesis, integrating GPT-4's planning capabilities with Blender's physical simulation and Stable Diffusion's image generation. It transforms textual prompts into Blender scripts to simulate basic physical motions, then uses these simulations to create videos with coherent motions.

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, apply here.