GPT-4 API Available to All!

Plus, see this video of the world's most advanced humanoid robot

Author

AI Breakfast
July 07, 2023

Good morning. It’s Friday, July 7th.

Did you know: We wrote a book about AI? Decoding AI: A Non-Technical Explanation of Artificial Intelligence is now available on Gumroad.

In today’s email:

  • OpenAI releases GPT-4 API to all

  • Lightning Labs introduces Bitcoin tools for AI

  • DigitalOcean acquires Paperspace for $111M

  • AI excels in creative thinking tests

  • AI robots as care home companions

  • VA study predicts aggressive prostate cancer with AI

  • Stanford enhances AI adaptability through self-reflection

  • Shutterstock protects users with generative AI

  • Japan prioritizes AI in school guidelines

  • Pentagon tests AI in all-out war planning

  • OpenAI forms Superalignment team for controlling AI

  • Mount Sinai opens Center for Ophthalmic AI and Human Health

  • ChatGPT experiences traffic decline

  • NYC regulates AI in hiring tools

  • AI detection industry grows with AI cheating

  • 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:

Attending Ai4 this year? We will be!

Join the world's premier AI for business conference at Ai4, uniting the global AI ecosystem through events, media, and resources​.

Attend multi-day, networking-enabled conferences featuring the world’s most innovative speakers, attendees, and AI companies​.

Today’s trending AI news stories

OpenAI has officially released the GPT-4 API to all users enrolled in the API program: The release also includes the availability of GPT 3.4 Turbo, DALL-E, and Whisper. As part of this update, OpenAI has announced its plan to retire older models in the Completions API, starting early 2024. The focus will now shift to the Chat Completions API, which offers a more structured prompt interface, enabling developers to create conversational experiences. You can signup for API access here.

Lightning Labs Introduces Bitcoin Tools for AI Applications: Lightning Labs has unveiled a suite of tools that enable ChatGPT to leverage the Lightning Network for holding, sending, and receiving Bitcoin. By integrating high-volume Bitcoin micropayments via Lightning with large language models, this could enable chatbots to send and receive bitcoin payments over the Lightning network.

DigitalOcean Acquires Paperspace for $111M: DigitalOcean has acquired Paperspace, a startup specializing in cloud computing and AI development, for $111M in cash. Paperspace customers will gain access to DigitalOcean's cloud services, while DigitalOcean expands its offerings tailored to the needs of small and medium-sized businesses and startups in the AI and machine learning domain.

AI Achieves Top 1% Performance in Creative Thinking Tests: According to research from the University of Montana, AI has demonstrated the ability to match the top 1% of human thinkers in a standard test for creativity. Using the Torrance Tests of Creative Thinking, the AI application ChatGPT performed exceptionally well in terms of fluency and originality. This study suggests that AI is rapidly developing creative abilities that are on par with or even surpassing human capabilities.

AI Robots Could Serve as Companions in Care Homes: Nadia Magnenat Thalmann, a robotics expert from the University of Geneva, suggests that social robots like Nadine could be valuable assets in caring for the sick and elderly. These robots have the potential to “outperform” human caregivers due to their constant availability, though acceptance from the elderly will likely be met with fierce resistance.

VA Study Utilizes AI for Aggressive Prostate Cancer Prediction: The Department of Veteran Affairs is conducting a study to develop an AI tool capable of predicting aggressive prostate cancer. Researchers from five VA medical centers will analyze data from over 5,000 veterans with high-risk prostate cancer, leveraging high-resolution scans, diagnostic imaging, and social determinants of health to identify early detection patterns of the cancer using AI.

Stanford Researchers Enhance AI Adaptability through Self-Reflection: Researchers at Stanford University have developed a training method called "curious replay," which significantly improves the performance of AI agents when adapting to changing environments. Inspired by animal behavior, this method encourages AI agents to self-reflect on novel and interesting experiences they encounter. The method demonstrates promising results in both simple tasks and complex problem-solving scenarios.

Shutterstock Provides Legal Protection for Enterprise Customers Using Generative AI: Shutterstock has announced that it will offer full indemnification to its enterprise customers for the license and use of generative AI images available on its platform, similar to what Adobe did last week. This indemnification will safeguard customers against potential claims associated with the use of these images.

Japan Prioritizes AI Comprehension in New School Guidelines: The latest school guidelines in Japan highlight the importance of integrating generative AI into educational settings. The guidelines encourage the use of AI in English learning and group activities to enhance students' comprehension and critical thinking, and will be implemented in both public and private educational institutions.

Pentagon Tests AI's Role in Planning Responses to All-Out War: The Pentagon is conducting tests to determine if generative AI can assist in planning responses to global conflicts and provide faster access to internal information. The LLMs have already been given classified information to generate responses on real-world matters. The tests see if the models could help plan a response to a potential escalation of the already tense military situation with China.

OpenAI Forms Superalignment Team to Control "Superintelligent" AI: OpenAI has established a new team called the Superalignment team, led by Ilya Sutskever, with the aim of developing methods to control and steer "superintelligent" AI systems. The team will focus on addressing technical challenges related to controlling such advanced AI and will have access to significant computing resources. The post predicts that we could see an AI Superintelligence (surpassing the capabilities of AGI, in their words) before the and of the decade.

Mount Sinai Establishes Center for Ophthalmic AI and Human Health: Mount Sinai has launched the Center for Ophthalmic Artificial Intelligence and Human Health in New York. The center aims to enhance patient care by leveraging AI for timely diagnosis of eye diseases and innovative risk assessment of systemic health conditions. Mount Sinai will integrate AI into education, research, and clinical settings to improve the diagnosis and treatment of various eye conditions.

ChatGPT Sees First Decline in Traffic Since Launch: ChatGPT experienced a decline in monthly traffic and unique visitors in June, marking its first decrease since its November launch. The website observed a 9.7% decrease in worldwide traffic and a 5.7% drop in unique visitors. Despite the decline, ChatGPT remains among the top 20 websites globally, with over 1.5 billion monthly visits. The release of the ChatGPT app on iOS in May likely contributed to the decrease in web traffic.

New York City Regulates AI in Hiring Tools: New York City is implementing a new law, NYC 144, to regulate the use of automation and AI in hiring decisions. The law requires employers to annually audit AI tools used in hiring and promotion decisions to identify potential biases based on race and gender, and publish the results. Employers failing to comply may face penalties. NYC 144 is being closely watched as a potential precedent for future AI regulation in other jurisdictions.

AI Detection Industry Expands Alongside AI Cheating: The surge in “AI cheating”has fueled the growth of companies specializing in AI detection. Companies such as Winston AI, Content at Scale, and Turnitin offer subscription services that attempt to identify AI involvement in student work. These detection methods typically involve searching for distinctive features that distinguish AI-generated content from human-authored content, though we’ve yet to come across an AI detector that is reliable enough to endorse.

Watch:

World’s most advanced humanoid robot, Ameca, draws a cat:

🎧 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

Featured Tool:

BeforeSunsetAI: An Ai powered daily planner that plans your day in a matter of seconds and frees you from wasting time and getting stressed. #1 Product of the Month on ProductHunt.

Knowledg.io: An AI automation tool that boosts work efficiency by providing real-time AI tools. It allows grouping prompts, automating repetitive tasks, integrating with third-party tools, and promoting collaboration in a unified workspace.

Redact | Pangea: This service simplifies sensitive data handling and compliance with regulations. It offers a single API for redacting sensitive information, streamlining data protection and compliance frameworks. Users can easily customize rules for flexibility in redacting sensitive data.

TakeTurns: A collaboration platform that brings structure, security, and transparency to document collaboration. It offers features like turn-based collaboration, file revisions, and participant roles, ensuring efficient collaboration while meeting confidentiality and security requirements.

BrewNote offers instant, high-quality notes for user interviews powered by AI. Skip rewatching interview recordings and simply drag and drop an audio file to will notes generated by AI. Designed for UX and product folks.

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

This research introduces KNOWNO, a framework that addresses the challenges of false confidence and ambiguity in large language model planners for robots. KNOWNO utilizes conformal prediction to align uncertainty, enabling LLM-based planners to know when to ask for help. It provides statistically guaranteed task success while minimizing the amount of human assistance required. The framework is evaluated in various simulated and real-world robot setups, showcasing improved efficiency and autonomy compared to baseline approaches. KNOWNO offers a promising solution for enhancing the reliability and performance of language-instructed robots in unstructured environments.

This paper presents a comprehensive study on training GPT4-style models with multimodal inputs. It explores network structures, training data, and diversified prompts to achieve high-performance multi-modal language models. The authors propose Lynx, a model that exhibits accurate multi-modal understanding and superior multi-modal generation ability compared to existing open-sourced models. They also introduce a benchmark for evaluating and comparing multi-modal LLMs, including both image and video tasks.

This paper presents LONGNET, a variant of the Transformer model that enables scaling sequence length to more than 1 billion(!) tokens. To address the challenge of computational complexity and model expressivity, LONGNET introduces dilated attention, which expands the attentive filed exponentially with distance. The proposed model offers linear computation complexity, supports distributed training, and can be seamlessly integrated with existing Transformer-based optimization techniques. This work opens up possibilities for modeling very long sequences, such as treating an entire corpus or the entire Internet as a single prompt.

This research paper by Meta AI explores the use of current language models to generate, recognize, and reframe unhelpful thoughts in the context of cognitive behavioral therapy (CBT). The authors propose PATTERNREFRAME, a new dataset consisting of examples of unhelpful thought patterns conditioned in specific personas, along with positive reframing proposals. They demonstrate that existing language models can be powerful tools to generate tailored practice material and hypotheses, aiding in the adoption of cognitive reframing techniques.

The research paper introduces Whisper-AT, a unified model that combines automatic speech recognition (ASR) and audio tagging. The authors first analyze Whisper, an ASR model trained on a large and diverse speech corpus, and find that it is robust against real-world background sounds but not noise-invariant. They leverage this finding to build Whisper-AT, which recognizes both spoken text and general audio events in a single pass. The model achieves high performance in audio tagging while maintaining computational efficiency.

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.