• AI Breakfast
  • Posts
  • GPT-4 Passes Harvard Experiment with 3.57 GPA

GPT-4 Passes Harvard Experiment with 3.57 GPA

Plus, DeepReel helps you create your custom AI video avatar

Author

AI Breakfast
July 31, 2023

Good afternoon. It’s Monday, July 31st.

Did you know: MIT produced one of the first artificial intelligence reports in 1972? Take a look.

In today’s email:

  • Google's RT-2 translates vision to robotic action

  • AI creates virtual influencers in fashion

  • OSI disputes Meta's LLaMa 2 open source claim

  • GPT-4 earns 3.57 GPA at Harvard

  • LinkedIn developing AI chatbot for jobs

  • AI identifies Neanderthal proteins for antibiotics

  • DoorDash tests AI chatbot for food ordering

  • Worldcoin under European scrutiny for biometric data

  • AI replicates artist Rutkowski's style amid debate

  • 5 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think of this edition by replying to this email, or DM us on Twitter.

Today’s edition is brought to you by:

Imagine if making videos was as simple as writing emails.

Enter DeepReel – the game-changer in video creation. With DeepReel, you can bring your ideas to life by creating a lifelike avatar from just text. No more time-consuming and expensive processes, just pure creativity!

Join the waitlist today to be among the first to experience the magic of DeepReel.

As an AI Breakfast reader, you'll even get an exclusive 50% off on your avatar creation cost. Unleash your potential and unlock the power of Gen-AI with DeepReel.

Today’s trending AI news stories

Google DeepMind's RT-2 Transforms Vision and Language into Action: Google DeepMind has unveiled Robotic Transformer 2 (RT-2), a pioneering vision-language-action model that equips robots with exceptional capabilities. By translating web and robotics data into general instructions, RT-2 offers groundbreaking improvements in generalization and semantic understanding, setting a new milestone for robotic intelligence.

AI Creates Virtual Influencers Earning Millions in Fashion Industry: Virtual influencers like Lil Miquela and Shudu, crafted using AI technology, are engaging millions with their fabricated lifestyles and high-profile fashion deals. Despite being entirely virtual, these influencers contribute to an industry estimated to grow 26% by 2025, worth £3.5 billion.

Open Source Watchdog Challenges Meta's LLaMa 2 Licensing: The Open Source Initiative (OSI) has accused Meta of misusing the term “open source” for its LLaMa 2 models. OSI argues the license contradicts open source principles by restricting commercial use, and is working on clarifying the definition, with a candidate release set for October 17, 2023.

GPT-4 Passes Harvard Experiment with 3.57 GPA: AI model GPT-4 successfully completed first-year humanities and social sciences essays at Harvard, earning a 3.57 GPA in an experiment conducted by student Maya Bodnick. The results highlight the evolving relationship between AI and human intelligence, with implications for the future of education.

LinkedIn's AI Chatbot Coach Aids in Job Applications: LinkedIn is developing an AI-powered chatbot, “LinkedIn Coach,” to assist users in job applications and networking. The tool could provide application support and skill training, reflecting Microsoft's broader integration of chatbots into its products.

AI Resurrects Neanderthal Proteins for Antibiotics: Researchers have employed AI to identify extinct Neanderthal bacteria-fighting proteins for potential drug development. The AI-driven method significantly shortened the identification process, although some experts call for more accurate peptide prediction before anticipating major impacts on drug discovery.

DoorDash Tests AI Chatbot DashAI to Streamline Food Ordering: DoorDash is testing an AI chatbot, DashAI, to speed up food ordering and enhance customer experience in select markets. The move follows statements from the company's CEO about experimenting with technology to improve services.

Worldcoin's Biometric Data Collection Faces Privacy Scrutiny in Europe: Worldcoin, offering crypto tokens in exchange for biometric data, is being investigated by European data protection authorities. Concerns regarding Worldcoin’s compliance with GDPR and privacy laws have been raised, though the company asserts its adherence to all relevant regulations.

Digital Artist Greg Rutkowski's Style Mimicked by AI Despite Controversy: After digital artist Greg Rutkowski's opposition to AI-generated art led to the removal of his work from Stability AI's dataset, the AI art community created a model to replicate his style. While some argue this was unethical, others see it as a mark of innovation in the ever-changing world of AI and art.

🎧 Did you know AI Breakfast has a podcast read by a human? Join AI Breakfast team member Luke (an actual AI researcher!) as he breaks down the week’s AI news, tools, and research: Listen here

5 new AI-powered tools from around the web

Rewind is a productivity-enhancing AI tool empowering users to browse, search, and inquire about phone content. Effortlessly access past information, find tweets, or revisit web pages. With Optical Character Recognition (OCR) for instant searchability and summarization for easy digestion, Rewind delivers a private, personalized AI experience.

Dubverse is an AI-powered video dubbing tool enabling effortless multilingual content creation. Leverage text-to-speech, machine translation, and generative AI for quick, cost-effective videos. Benefit from script editor and human-like AI voices with multi-language dubbing.

Insumo AI’s ADHD brain planner streamlines task management and time tracking, featuring the unique Magic Planner AI assistant. With seamless integration with Google Calendar, Asana, and Todoist, Insumo enhances productivity and motivates users to achieve their goals.

Object Remover is an AI-powered photo editing tool enabling users to easily remove unwanted elements from images. It supports removing people, cars, products, text, and backgrounds utilizing deep learning algorithms for precise results. With a simple interface and watermark-free exports, it offers a free and efficient editing experience.

Synthical simplifies research with AI-powered article processing, collaboration, and a vast repository of science articles. Designed for researchers, it covers various disciplines and offers a free option.

arXiv is a free online library where scientists share their research papers before they are published. Here are the top AI papers for today.

The paper presents a scaling rule for optimization in the presence of model Exponential Moving Averages (EMA) to preserve training dynamics across batch sizes in machine learning. EMA is a model copy that follows its target model with momentum and contributes to optimization. The authors demonstrate the validity of their EMA scaling rule across various architectures, optimizers, and data types. The EMA Scaling Rule allows for consistent training dynamics across different batch sizes, enabling the training of EMA-based techniques like pseudo-labeling and Self-Supervised Learning (SSL) at both small and large batch sizes. Empirical experiments support the effectiveness of the proposed rule.

The paper explores the challenges and limitations of using RLHF to train AI systems. Authored by researchers from institutions like MIT, Harvard, and UC Berkeley, the paper addresses issues like sensitive information disclosure, biased outputs, and undesirable preferences in AI models. RLHF has become crucial for training large language models, but the authors stress the need for a multi-faceted approach to AI system development, combining RLHF with other safety measures and transparency standards. They call for research contributions to address RLHF’s challenges and advocate for transparency in safety approaches and risk disclosure.

MiDaS v3.1, a Model Zoo for Monocular Relative Depth Estimation, introduces a diverse range of new models based on different encoder backbones for depth estimation. This release includes transformer-based encoders, which have shown promising results in a computer vision tasks, as well as convolutional encoders. MiDaS v3.1 improves depth estimation quality by 28% compared to its predecessor. MiDaS v3.0, while offering efficient models suitable for high frame rates. By factoring out metric scale, MiDaS enables robust depth estimation and generalizability across environments. The release covers a variety of transformer-based backbones, such as BEiT, Swin, SwinV2, Next-ViT, and LeViT, alongside a general guide for integrating future backbones into MiDaS.

NeRF-Det, an advanced AI-system, revolutionizes indoor 3D object detection using RGB images. Unlike previous methods that struggle with scene geometry, NeRF-Det incorporates Neural Radiance Field (NeRF) in an end-to-end manner to explicitly estimate 3D geometry, vastly improving detection performance. Through a shared MLP, NeRF and detection branches synergize, enhancing geometry-aware volumetric representations for precise 3D detection. NeRF-Det outperforms current benchmarks by 3.0 mAP and 3.1 mAP on ScanNet and ARKITScenes datasets, respectively. Its generalizability to unseen scenes sets it apart, making it a breakthrough in RGB-based 3D detection.

TEDi is an AI-driven motion synthesis tool that extends De-noising Diffusion Probabilistic Methods (DDPM) for long-term motion sequences. By entangling the temporal-axis of motion with the diffusion time-axis, TEDi generates continuous streams of clean motion frames. The framework allows for diverse and controllable motion generation, making it valuable for character animation and other applications. TEDi’s auto-aggressive U-Net architecture produces high-quality motion sequences and can potentially be applied to other sequential data types. Future improvements aim to reduce latency and explore text-conditioned motion generation with user guidance.

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.

Attending Ai4 this year? We will be!