

Discover more from Last Week in AI
Last Week in AI #238: OpenAI's DALL-E 3, DeepMind's AI self-optimization breakthrough π, LLMs struggle with logical symmetry π, Teachers embrace generative AI in classrooms π«, and more!
OpenAI to release DALL-E 3, DeepMind's algorithm that optimizes LLM prompts, LLMs seem unable to perform basic logical deductions, teachers are embracing generative AI tools for lesson plans and more
Top News
OpenAI releases third version of DALL-E
OpenAI announced DALL-E 3, the new version of its text-to-image AI model. The new version of DALL-E is much better at following text instructions, and it is also much better at rendering text in the image (the above generated picture had a prompt of βAn illustration of an avocado sitting in a therapist's chair, saying 'I just feel so empty inside' with a pit-sized hole in its center. The therapist, a spoon, scribbles notes.β) This new version of DALL-E is integrated with ChatGPT, which allows users to leverage the chatbot to create more detailed prompts. To avoid potential legal issues, OpenAI is also allowing artists to opt-out their artwork in future image-generating models.
DeepMind discovers that AI large language models can optimize their own prompts
DeepMind researchers have proposed a new method called Optimization by PROmpting (OPRO) that uses LLMs as optimizers. Unlike traditional methods that rely on formal mathematical definitions, OPRO defines optimization tasks in natural language. The process begins with a "meta-prompt" that includes a task description and examples of problems and solutions. The LLM then generates candidate solutions based on this information, which are evaluated and assigned a quality score. The best solutions are added to the meta-prompt for the next round of solution generation. This method allows for a wide array of problems to be solved by simply modifying the problem description or adding specific instructions. The researchers found that OPRO can sometimes match or even surpass the performance of expert-designed heuristic algorithms on small-scale optimization problems. Amazingly, OPRO found one of the most performant prompts to improve LLM reasoning is βTake a deep breath and work on this problem step-by-stepβ, which significantly outperforms just βLetβs think step-by-stepβ
LLMs trained on βA is Bβ fail to learn βB is Aβ
In a recent paper, researchers discovered a surprising failure of LLMs: their inability to learn the concept of symmetry, specifically in the context of "A is B" not necessarily implying "B is A". The paper, titled "LLMs trained on 'A is B' fail to learn 'B is A'", explores this phenomenon, referred to as the "reversal curse". The study reveals that despite being trained on a large amount of data, LLMs struggle to understand this basic logical concept. Experiments with GPT-4 found that while the model can correctly answer 79% of βforwardβ questions (e.g. Who is Tom Cruiseβs mother?), it only answered 33% correctly for the βreverseβ questions (e.g. WHo is Mary Lee Pfeifferβs son?). This raises questions about LLMsβ reasoning capabilities and the ability to perform logical deductions.
Teachers Are Going All In on Generative AI
Recent surveys show that teachers actually use generative AI tools more than students (e.g. for creating lesson plans and interesting word problems). The AI Education Project, funded by Google, Intel, and OpenAI, has trained over 7,000 teachers in AI usage, with generative AI commonly used for lesson planning and parent communication. However, some teachers worry about the ethical implications of automating parts of their job and fear job displacement due to AI. Despite skepticism due to past disappointments with edtech, experts like Joseph South from the International Society for Technology in Education (ISTE) insist that AI in education is not a passing fad. The Charter School Growth Fund has formed working groups to advise on AI policy, emphasizing the need to monitor the quality of AI-generated content.
Other News
Business
5 lessons from 139 YC AI startups (S23) - A record-breaking 139 AI startups were showcased at YC's Demo Day, with key takeaways including the diverse industries being impacted by AI, the rise of AI copilots for businesses, the emergence of AI Ops as a critical sector, the question of model ownership, and the realization that AI is not a magical solution but requires a strong business foundation.
Writer nets $100M for its enterprise-focused generative AI platform - Writer, an enterprise-focused generative AI platform, has raised $100 million in a Series B funding round to develop industry-specific text-generating AI models, with customers including Accenture and Vanguard.
Intel says newest laptop chips, software will handle generative AI - Intel's new laptop chip, set to be released in December, will allow generative AI chatbots to run on laptops without relying on cloud data centers, enabling businesses and consumers to test AI technologies without sending sensitive data off their own computer.
Midjourney vs. Ideogram, ML product companies, preventing AI winter - Midjourney and Ideogram are two ML-powered startups competing in the image generation space, with Midjourney pioneering a unique approach to building an ML company and Ideogram offering a text feature that sets it apart from its competitor.
Microsoft Bing to gain more personalized answers, support for DALLE-E 3, and watermarked AI images - Microsoft Bing is gaining AI improvements, including support for OpenAI's DALLE-E 3 model, more personalized answers in search and chat, and tools to watermark AI-generated images.
Microsoftβs mobile keyboard app SwiftKey gains new AI-powered features - Microsoft's SwiftKey mobile keyboard app for iOS and Android is gaining new AI-powered features, including AI camera lenses, AI stickers, an AI-powered editor, and the ability to create AI images from the app.
The AI Hype Train Has Stalled in China - Chinese startups in the AI sector are facing challenges in competing with internet giants due to the high cost of building large language models (LLMs), limited supplies of GPUs, and a challenging economic environment, leading to a shift towards using AI for small efficiency improvements rather than revolutionary applications.
A flood of new AI products just arrived β whether weβre ready or not - Big Tech companies are rushing to release new AI products, such as chatbots and image generators, before they are fully ready, resulting in mistakes and unproven technology.
Concerns
Deepfakes of Chinese influencers are livestreaming 24/7 - AI-generated livestreamers are being used in China to replace mid-tier human influencers, leading to a decrease in job opportunities and salaries for human livestream hosts.
Nearly half of CEOs believe AI could replace their own jobs, says new poll: 'This is probably bigger than the internet' - Nearly half of CEOs believe AI could replace their own jobs, with 49% saying AI could effectively replace "most," or even "all," of their own roles, according to a survey from online education platform edX.
38TB of data accidentally exposed by Microsoft AI researchers - Microsoft AI researchers accidentally exposed 38 terabytes of private data, including personal computer backups, passwords, and internal messages, while publishing open-source training data on GitHub, highlighting the risks organizations face when working with massive amounts of training data in AI projects.
How generative AI is accelerating disinformation - Generative AI tools are making it easier and cheaper to generate and distribute disinformation on a massive scale, posing a threat to democracy and the spread of objective truth.
Research
Two New Papers Analyze in Detail the Protein Universe Unveiled by AlphaFold 2βs 200 Million Models - The recent release of over 200 million predicted protein structures by DeepMind's AlphaFold 2 has led to two groundbreaking papers that analyze the structural diversity, evolutionary relationships, and functional potential of proteins at an unprecedented scale.
2023 Innovator of the Year: As AI models are released into the wild, Sharon Li wants to ensure theyβre safe - Sharon Li, MIT Technology Review's 2023 Innovator of the Year, is pioneering an AI safety feature called out-of-distribution (OOD) detection to prevent AI models from breaking in unexpected and catastrophic ways when faced with unfamiliar scenarios in the real world.
Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality - The study examines the performance implications of AI on realistic, complex, and knowledge-intensive tasks, finding that consultants using AI were more productive and produced higher quality results, but struggled with tasks outside the current capability of AI.
How Google taught AI to doubt itself - Google's AI language model, Bard, now includes a feature that allows users to double-check the information it provides by evaluating whether there is content across the web to substantiate it, helping to address the problem of chatbots making things up.
AI just beat a human test for creativity. What does that even mean? - AI chatbots have achieved higher average scores than humans in a test designed to measure creativity, raising questions about the unique characteristics of humans and the limitations of AI.
How to Tell if Your A.I. Is Conscious - Researchers are exploring different theories and features that could potentially determine if an AI is conscious, but the complexity and interpretability of deep neural networks pose challenges in applying these theories to machines.
starter bibliography for Yann LeCun: 2000-09 - Yann LeCun's research from 2000-2009 covers a wide range of topics in machine learning, computer vision, and artificial intelligence, including deep learning, convolutional neural networks, energy-based models, object recognition, and autonomous navigation.
MindAgent: Emergent Gaming Interaction - A new infrastructure called MindAgent is proposed to evaluate planning and coordination emergent capabilities for gaming interaction, leveraging existing gaming frameworks to collaborate with human players and establish in-context learning, with the goal of shedding light on how such skills can be obtained by learning from large language corpora.
Analysis
Silicon Valleyβs vision for AI? Itβs religion, repackaged. - Silicon Valley's excitement about building superintelligent machines is rooted in recycled religious ideas, with many technologists proposing that AI can offer eternal life and perfect knowledge, similar to Christian eschatology, and drawing a straight line from Christian theologians in the Middle Ages to the futurist Ray Kurzweil to the tech heavyweights he's influenced in Silicon Valley.
Policy
Biden plans to work with world leaders to ensure AI's use as a tool of 'opportunity' - Biden plans to collaborate with world leaders to ensure that AI is used as a tool of opportunity and not as a weapon of oppression, working to strengthen rules and policies to govern the technology.
AI in Focus: Machine learning goes to Congress - US lawmakers met with tech leaders to discuss AI accountability and the working conditions for data labelers, while companies like Salesforce are hiring more employees to focus on AI and Adobe's sales outlook disappoints investors.
China and EU hold talks on AI, cross-border data flow amid renewed tensions - China and the European Union held talks on artificial intelligence and cross-border data flows in Beijing, amid tensions over an EU probe into China's electric vehicle subsidies.
How Silicon Valley doomers are shaping Rishi Sunakβs AI plans - Silicon Valley doomers associated with the Effective Altruism movement are shaping UK Prime Minister Rishi Sunak's AI plans, with concerns raised about regulatory capture and the downgrading of more immediate AI risks.
TikTok now lets creators label AI-generated content - TikTok is introducing labeling tools that allow users to identify posts that have been created or enhanced by AI, in an effort to increase transparency and prevent the spread of misleading content, with plans to develop an algorithm that automatically labels AI-generated posts in the future.
EU considering whether to attend Britain's AI summit, spokesperson says - The European Union is considering whether to send officials to Britain's upcoming artificial intelligence safety summit as it nears completion of wide-ranging AI legislation, while the British government hopes to position itself as a global leader in AI regulation.
Fun
Funky AI-generated spiraling medieval village captivates social media - An AI-generated image of a spiral-shaped medieval village gained attention on social media for its remarkable geometric qualities, with some praising it as a breakthrough in AI-generated art while others criticized its compositional elements.
Copyright Β© 2023 Skynet Today, All rights reserved.