Netflix Wants to Pay AI Engineers $900k/yr

...and why OpenAI quietly removed their AI detector

Author

AI Breakfast
July 26, 2023

Good morning. It’s Wednesday, July 26th.

Did you know: Researchers may have just achieved room-temperature superconductivity? If true, this could rival AI for tech breakthrough of the decade. Read the full paper here.

In today’s email:

  • Netflix posts $900k AI job

  • Microsoft's $20.1B profit with AI shift

  • Palantir CEO supports AI weapons

  • Cohere's Coral AI targets enterprise use

  • OpenAI discontinues flawed AI detection tool

  • Spotify's CEO teases AI features

  • STEVE-1 AI plays Minecraft

  • Worldcoin rallies on trading debut

  • Stability AI launches eco-friendly language models

  • Bing Chat tests on Chrome and Safari

  • Google co-founder aids in AI project

  • MIT's 'PhotoGuard' fights AI image edits

  • 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.

This week’s editions are brought to you by:

MinervaAI: Empowering AML Compliance with AI-driven Solutions

Streamline and automate regulatory obligations, KYC, EDD, and ongoing monitoring, with auditable reports and reduced manual work.

Trusted by leading companies, MinervaAI offers a cloud-based risk assessment platform for real-time financial crime analysis and decreased risk.

Today’s trending AI news stories

As Actors Strike for AI Protections, Netflix Lists $900,000 AI Job: As Hollywood actors’ union have gone on a strike to demand AI protection as entertainment firms make significant investments in AI programs, Netflix is offering a staggering $900,000 for an AI product manager role.

Microsoft reports $20.1B quarterly profit as it promises to lead 'the new AI platform shift’: Microsoft reported a 20% surge in quarterly profits, reaching $20.1 billion, surpassing analyst expectations. The software giant’s revenue for the April-June period was $56.2 billion, an 8% increase from the previous year. The company’s focus on leading the new AI platform is evident through its AI features integrated into cloud computing services, workplace software, and the search engine Bing. Microsoft’s cloud business segment contributed significantly to the quarterly sales with its flagship Azure platform and the other cloud services revenue.

Palantir CEO Argues US Must Develop AI Weapons Despite Ethical Concerns: In a thought-provoking New York Times op-ed, Palantir CEO Alex Karp boldly argues that the value of software in defense contexts lies in its ability to eliminate enemies. This morally complex stand highlights the deep divisions surrounding the use of AI for lethal purposes. Despite ethical debates, Karp firmly believes that advanced AI development is crucial to “constrain adversaries” and foster lasting peace.

Cohere releases Coral, AI assistant designed for enterprise business use: AI company Cohere has launched Coral, a “knowledge assistant” tailored for enterprise use. Coral leverages natural language processing (NLP) powered by large language models (LLMs) to provide knowledge workers with sector-specific responses based on proprietary company data. The platform goes beyond publicly available generative AI tools, offering advantages like mitigating hallucination problems by providing citations to proprietary data and maintaining privacy and security through cloud-agnostic data storage.

OpenAI Quietly Shuts Down Its AI Detection Tool: OpenAI quietly discontinues its AI detection tool, the AI Classifier, due to poor accuracy in distinguishing content created using generative AI tools. The tool’s limitations, including mislabeling human-written text as AI-written and poor performance with shorter texts, led to its shutdown.

Spotify CEO teases potential AI-powered capabilities surrounding personalization, ads: During Spotify’s second-quarter earnings call, CEO Daniel Ek discussed potential AI-powered features for the streaming service. Ek mentioned using AI for personalized experiences, summarizing podcasts, and generating audio ads. The company recently launched a DJ feature that delivers curated music with AI-powered commentary, and Ek hinted at more contextual and personalized AI-powered features. Spotify is seeking a patent for an AI-powered “text-to-speech synthesis” system, indicating a focus on AI voice technology. The streaming service reported strong user growth and announced a price hike for its premium plans.

STEVE-1 is a chatbot that plays Minecraft: STEVE-1, an innovative generative AI model, has emerged as a formidable player in the world of Minecraft, showcasing its ability to perform tasks through text instructions. Leveraging the power of two existing AI models, VPT and MineCLIP, STEVE-1, surpasses its predecessors in short-term tasks, gathering resources and exploring more efficiently. Chaining text prompts, it demonstrates impressive results in longer-term tasks, hinting at its potential as an interactive assistant. The approach employed STEVE-1’s development holds promise for creating instructable agents beyond the world of Minecraft, with future efforts focused on handling more complex instructions seamlessly.

Open AI CEO Sam Altman's Worldcoin rallies on first day; around $145 million worth of token traded: The cryptocurrency venture, “Worldcoin,” spearheaded by Sam Altman, the CEO of OpenAI’s ChatGPT, witnessed a surge in its token’s value. Initially starting at $1.70, Worldcoin soared to $3.58 before settling at $2.52, as reported by CoinMarketCap. This unique project aims to establish individual digital identities termed “World IDs” through advanced eye-scanning technology, solidifying one’s authenticity in the digital realm. However, due to regulatory restrictions, Worldcoin will not be available for trading in the United States.

Stability AI unveils new FreeWilly language models trained using minimal — and highly synthetic — data: Stability AI unveils its latest breakthrough in language models with the introduction of FreeWilly1 and FreeWilly2. These models, inspired by the ‘90s classic film, boast exceptional capabilities in intricate reasoning and linguistic nuances, particularly in domains such as law and mathematics. What sets them apart is their training on a smaller dataset, including synthetic data, making them more environmentally friendly and cost-effective. Released under a non-commercial license, the models are aimed at advancing research and fostering open access within the AI community.

Microsoft’s Bing Chat comes to Chrome and Safari in tests for ‘select users’: The AI chatbot powered by OpenAI’s GPT-4 model, is expanding its reach beyond Microsoft products to non-Microsoft browsers such as Chrome and Safari. This move comes as Microsoft aims to make Bing Chat accessible to a wider user base. Although specific details about the expansion’s timeline and supported platforms remain undisclosed, this strategic move demonstrates Microsoft’s commitment to enhancing user engagement and accessibility with its AI chatbot technology. Bing Chat is being tested with a native dark theme.

Google Co-Founder Returns to Help With AI Efforts: Google co-founder Sergey Brin has returned to the company to aid in developing its AI system named Gemini, which aims to be more “human-like.” Sergey Brin stepped back from active roles in Google and its parent company Alphabet four years ago but is now working with researchers on the upcoming AI model. Gemini, introduced during Google’s I/O 2023 event, is designed to be a next-generation foundation model with various sizes and capabilities, mirroring the approach of PaLM 2.

MIT's 'PhotoGuard' protects your images from malicious AI edits: MIT CSAIL introduces “PhotoGuard,” a technique to protect images from malicious AI edits. The method involves nearly invisible “perturbations” that disrupt the AI’s understanding of the image. It employs both “encoder” and “diffusion” attack methods to prevent unauthorized image manipulation. While it shows promise, the technique may not be foolproof and requires collaborative efforts from model developers, social media platforms, and policymakers to combat potential threats.

🎧 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:

Cape Privacy: Keep sensitive data private while prompting LLMs like GPT-4 and GPT-3.5 Turbo by using : Anonymous prompting through APIs.

Eden AI offers a seamless API that combines multiple top AI engines, enabling users to access various generative AI APIs for content creation, translation, text, image, video, speech, and document parsing. The platform caters to all AI maturity levels, providing lightning-fast integration, no vendor lock-in, reduced costs, and the ability to switch between providers.

Slideoo is an AI-powered SaaS platform that allows users to create professional slide decks in just 1–2 minutes from lengthy text, PDFs, and website URLs. With a simple editor and a wide range of themes and templates, it offers a quick and efficient solution for creating unique presentations.

Mindstudio creates personalized context-aware AI apps without coding. Explore content generators, utilities, enterprise apps, agents, characters, bots, and more. Join their discord and get started for free.

Akstro offers AI-enabled search for static documentation. It utilizes text similarity search and large language models. Integrate the search UI and data plugin to allow users to search broader topics and ask specific questions. Early access is available.

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

Microsoft researchers propose a groundbreaking approach to predict code coverage without the need for code execution. Traditionally, measuring code coverage during testing is resource-intensive and requires executing the entire program. To alleviate this, they introduce a novel benchmark, “Code Coverage Prediction for Large Language Models (LLMs),” enabling models to understand code execution based on given test cases. The researchers curated a dataset, COVERAGEEVAL, comprising coverage-annotated methods. Evaluating prominent LLMs like GPT-4 and GPT 3.5, they found that models still struggle to accurately predict code coverage. This task offers a valuable metric to improve LLMs’ performance in software engineering.

In a bid to test the reasoning abilities of LLMs, researchers have introduced the Advanced Reasoning Benchmark (ARB). The benchmark comprises challenging problems in multiple fields, such as mathematics, physics, biology, chemistry, and law, sourced from advanced exams and professional resources. Current models, including GPT-4, have struggled to achieve high scores on the more demanding tasks. To enhance evaluation capabilities, a rubric-based self-evaluation approach allows LLMs to assess their own reasoning steps. ARB aims to spur progress in LLMs’ reasoning skills and enable more reliable assessments of complex model outputs.

Researchers from Tel Aviv University and Google Research examined the impact of knowledge editing in language models. They introduced a new evaluation approach, focusing on “ripple effects” resulting from factual edits, leading to the creation of the RIPPLEEDITS benchmark. Evaluating various LM models and knowledge editing methods, they found that existing methods were effective in changing specific facts, but often failed to address the associated ripple effects. However, a simple in-context editing approach showed promising results. The research underscores the significance of considering broader implications when updating factual knowledge in language models.

Researchers from Google DeepMind and the University of Tokyo, developed WebAgent, leveraging pre-trained language models to navigate real-world websites following natural language instructions. The WebAgent plans tasks by decomposing instructions into sub-instructions, summarizes long HTML documents, and acts on websites using Python programs. They introduced two language models, HTML-T5 and Flan-U-PaLM, for HTML summarization and program generation, respectively. Empirical results showed that the WebAgent significantly improved success rates on real websites by over 50% and achieved state-of-the-art results on web-based tasks. HTML-T5 alone outperformed other models on the MiniWoB++ benchmark.

The paper proposes 3D-LLMs, a new family of large language models that can take 3D representations as input and perform various 3D-related tasks such as captioning, question answering, and navigation. To overcome the scarcity of 3D-language data, the authors devise unique generation pipelines, using three prompting mechanisms to collect over 300k 3D-language pairs. The 3D-LLMs leverage 2D pre-trained vision-language models as backbones and introduce a 3D localization mechanism to capture 3D spatial information. Experimental results show that the model outperforms state-of-the-art baselines, demonstrating its capability in handling diverse 3D tasks.

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!