June 19 - 24th: Safe Superintelligence Inc (SSI), Claude 3.5 Sonnet, ∇^2 DFT: A Universal Quantum Chemistry Dataset, Gen-3 Alpha (Runway), The Devil is in the Details: StyleFeatureEditor, Florence-2
+ Bot Mimicry, A Reality Check, AI & Education, NotebookLM goes global, a ‘Virtual Rodent’ powered by AI, data-driven crop-growth simulation, Instruction Pre-Training, etc...
June 19 - 24th: Safe Superintelligence Inc (SSI), Claude 3.5 Sonnet, ∇^2 DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules, Gen-3 Alpha (Runway), The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing, Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks (Azure AI, Microsoft), ChatGPT is bullshit (Ethics and Information Technology), Elephants call each other by name, Why artists are becoming less scared of AI (MIT Technology Review), Is the Intelligence-Explosion Near? (A Reality Check: Sabine Hossenfelder), An AI Bot Is (Sort of) Running for Mayor in Wyoming (Wired), When misogyny intersects with AI and elitism, girls get hurt (The Guardian), Bot-mimicry in Digital Literary Culture (Cambridge), Revenge Of The Humanities (Steven Johnson), Review of Carl Öhman's “The Afterlife of Data” (Inside HigherEd), CFP: Contesting Artificial Intelligence, Artificial Intelligence and Education: A critical view through the lens of human rights, democracy and the rule of law, NotebookLM goes global with Slides support and better ways to fact-check (Google), Opening up ChatGPT: LLM openness leaderboard, Slack Combines Abstract Syntax Tree (ASTs) with LLMs to Automatically Convert 80% of 15,000 Unit Tests, Rice Farming Gets an AI Upgrade, TopCVPR-2024-papers, a ‘Virtual Rodent’ powered by AI to understand the brain (DeepMind), Data-driven crop growth simulation, AI enables faster, more effective antibiotic treatment of sepsis, Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models (Nature Machine Intelligence), Instruction Pre-Training: Language Models are Supervised Multitask Learners, Stable Diffusion 3 Medium, Luma Dream Machine, Kling AI, Poetics of Encryption (Kunsthal Charlottenborg), Glaze, Junie Lau (AIGC), Laurie Anderson on making an AI chatbot of Lou Reed: ‘I’m totally, 100%, sadly addicted’, Art or Artifice? Large Language Models and the False Promise of Creativity (CHI Conference)
TL;DR Ilya Sutskever is back, proposing a crack straight-shot to Safe Superintelligence Inc (SSI); Anthropic releases Claude 3.5 Sonnet “Sonnet now outperforms competitor models and Claude 3 Opus on key evaluations, at twice the speed” for free; open-sourced: ∇^2 DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules a paper which begins: “Solving the many-particle Schrödniger equation (SE) for electrons makes it possible to describe the electronic structure of matter” ; Microsoft releases: Florence-2 “a strong vision foundation model contender with unprecedented zero-shot and fine-tuning capabilities.”; Runway releases: Gen-3 Alpha: A New Frontier for Video Generation; DeepMind and Harvard built a virtual rodent that helps explain the brain; it is claimed that ChatGPT is bullshit; “Keynome gAST, or genomic Antimicrobial Susceptibility Test, bypasses the need for culture growth by analyzing bacterial whole genomes” and enables faster, more effective antibiotic treatment of sepsis; predictably, Elephants call each other by name; sensibly, Is the Intelligence-Explosion Near? (A Reality Check: Sabine Hossenfelder); helpfully, NotebookLM goes global with Slides support and better ways to fact-check (Google); and a sota image-editing style transfer model emerges The Devil is in the Details: StyleFeatureEditor
🏓 Observations: ChatGPT is bullshit (Ethics and Information Technology), Elephants call each other by name, Why artists are becoming less scared of AI (MIT Technology Review), An AI Bot Is (Sort of) Running for Mayor in Wyoming (Wired), When misogyny intersects with AI and elitism, girls get hurt (The Guardian), Bot-mimicry in Digital Literary Culture (Cambridge), Revenge Of The Humanities (Steven Johnson), Review of Carl Öhman's “The Afterlife of Data” (Inside HigherEd), CFP: Contesting Artificial Intelligence, Artificial Intelligence and Education: A critical view through the lens of human rights, democracy and the rule of law
✭ChatGPT is bullshit | Ethics and Information Technology “Recently, there has been considerable interest in large language models: machine learning systems which produce human-like text and dialogue. Applications of these systems have been plagued by persistent inaccuracies in their output; these are often called “AI hallucinations”. We argue that these falsehoods, and the overall activity of large language models, is better understood as bullshit in the sense explored by Frankfurt (On Bullshit, Princeton, 2005): the models are in an important way indifferent to the truth of their outputs. We distinguish two ways in which the models can be said to be bullshitters, and argue that they clearly meet at least one of these definitions. We further argue that describing AI misrepresentations as bullshit is both a more useful and more accurate way of predicting and discussing the behaviour of these systems.”
✭Elephants call each other by name, study finds | Zoology | The Guardian “Researchers used artificial intelligence algorithm to analyse calls by two herds of African savanna elephants in Kenya.”
✭ Why artists are becoming less scared of AI | MIT Technology Review “I’ve just come back from Hamburg, which hosted one of the largest events for creatives in Europe, and the message I got from those I spoke to was that AI is too glitchy and unreliable to fully replace humans and is best used instead as a tool to augment human creativity. ~ Right now, we are in a moment where we are deciding how much creative power we are comfortable giving AI companies and tools. After the boom first started in 2022, when DALL-E 2 and Stable Diffusion first entered the scene, many artists raised concerns that AI companies were scraping their copyrighted work without consent or compensation. Tech companies argue that anything on the public internet falls under fair use, a legal doctrine that allows the reuse of copyrighted-protected material in certain circumstances. Artists, writers, image companies, and the New York Times have filed lawsuits against these companies, and it will likely take years until we have a clear-cut answer as to who is right. ~ Meanwhile, the court of public opinion has shifted a lot in the past two years. Artists I have interviewed recently say they were harassed and ridiculed for protesting AI companies’ data-scraping practices two years ago. Now, the general public is more aware of the harms associated with AI. In just two years, the public has gone from being blown away by AI-generated images to sharing viral social media posts about how to opt out of AI scraping—a concept that was alien to most laypeople until very recently.”
✭ An AI Bot Is (Sort of) Running for Mayor in Wyomin (Wired) “Wyoming’s secretary of state wants the county to reject its candidacy, but the AI bot’s human “meat puppet” says everything is in order.”
✭ Vomit-inducing deepfake nudes show yet again that when misogyny intersects with AI and elitism, girls get hurt (The Guardian) “A teenage private schoolboy has been arrested for allegedly distributing “incredibly graphic” deepfake images of 50 girls from Bacchus Marsh grammar in Victoria. The story represents a triskelion knot of technolibertarianism, exclusivity and misogyny we are disastrously failing to untie. ~ The girls’ recognisable features were apparently scraped from social media photos, then an AI “nudifying” app did the rest. The boy allegedly shared the resulting composite images on social media. He was arrested but released without charge. But the girls saw the images. Friends saw them. Parents saw them. One parent described having to provide a bucket for her traumatised daughter to be sick into after seeing them – and her daughter wasn’t even one of the victims. One word the parent used in her description of the images was “mutilated”. ~ The Bacchus Marsh episode is shocking but not unprecedented. The wild west world of barely regulated AI technology has made schoolyard deepfake pornographers a global phenomenon. In February they struck in Beverly Hills. Before that, New Jersey. Last September a group of local girls aged 12 to 14 found themselves the victims of a similar attack from generational peers in a small town in Spain.”
✭ Bot-mimicry in Digital Literary Culture (Cambridge) “This Element traverses the concept and practice of bot mimicry, defined as the imitation of imitative software, specifically the practice of writing in the style of social bots. Working as both an inquiry into and an extended definition of the concept, the Element argues that bot mimicry engenders a new mode of knowing about and relating to imitative software – as well as a distinctly literary approach to rendering and negotiating artificial intelligence imaginaries. The Element presents a software-oriented mode of understanding Internet culture, a novel reading of Alan Turing's imitation game, and the first substantial integration of Walter Benjamin's theory of the mimetic faculty into the study of digital culture, thus offering multiple unique lines of inquiry. Ultimately, the Element illuminates the value of mimicry – to the understanding of an emerging practice of digital literary culture, to practices of research, and to our very conceptions of artificial intelligence.”
✭Revenge Of The Humanities (Steven Johnson) “... thanks to the AI revolution, we are entering a period where it will be a great time to be a humanities major with an interest in technology. The conventional story of course is that we're in the middle of a mass exodus from English or History as majors—sometimes blamed on the excesses of cultural theory, sometimes on the fact that all the money is in Computer Science and Engineering right now. And to be sure, the exodus is real. But there is a case to be made that college and grad students are over-indexing on the math and the programming, just as the technology is starting to demand a different set of skills. ~ The simple fact of the matter is that interacting with the most significant technology of our time—language models like GPT-4 and Gemini—is far closer to interacting with a human, compared to how we have historically interacted with machines. If you want the model to do something, you just tell it what you want it to do, in clear, persuasive prose. People who have command of clear and persuasive prose have a competitive advantage right now in the tech sector, or really in any sector that is starting to embrace AI. Communication skills have always been an asset, of course, but thanks to language models they are now a technical asset, like knowing C++ or understanding how to maintain a rack of servers. ~ This is, of course, a variation on Andrej Karpathy's quip from more than a year ago: "The hottest new programming language is English." ”
✭ Review of Carl Öhman's “The Afterlife of Data” (Inside HigherEd) ““Onlife” has become, for all practical purposes, a clumsy and unnecessary synonym for ordinary experience—especially since the turn of this decade, when social space and digital communication merged for an agonizingly long time to a degree that became normal. ~ Yet “onlife” may still have its uses. It certainly proves an essential concept for Carl Öhman in The Afterlife of Data: What Happens to Your Information When You Die and Why You Should Care (University of Chicago Press). The author is an assistant professor of political science at Uppsala University, in Sweden, though the present work belongs to the interdisciplinary field of information and communications technology (ICT). ~ Everyone online generates enormous quantities of personal information and tracking data, much of which is stored and will continue to exist after the person creating it has died. “The corpus of information left behind upon death,” Öhman writes, “is not just etymologically, but also conceptually analogous to the corpse.” As the population of users grows, so does the number of “information bodies” left by the deceased.”
CFP ✭Contesting Artificial Intelligence: Communicative Practices, Organizational Structures, and Enabling Technologies | Frontiers Research Topic “AI-enabled decision-making is becoming mainstream in many social and organizational contexts. Despite multiple benefits including efficiency, effectiveness, and objectivity, scholars from multiple academic disciplines have also identified various areas of concern related to AI-enabled decisions and mechanisms, including functional unreliability, epistemic opacity, privacy violations, a lack of accountability, and instances of algorithmic discrimination (Barocas and Selbst, 2016). A key instrument in achieving more trustworthy and ethical AI is to provide effective means for stakeholders to challenge the decisions and mechanisms of AI systems. As reflected in emerging AI regulation (e.g., EU AI Act), opportunities for contestation are especially crucial in high risk application areas (e.g., law enforcement, justice, medicine, recruitment, and autonomous driving) where problematic decisions of AI systems may have a direct negative impact on stakeholders. However, developing ways to effectively challenge aspects of AI can be equally important in cases where AI has potential indirect negative consequences for people, communities, and the environment (Crawford, 2021). While contestability of AI is implicitly addressed by various emerging norms (regulation and standards), technical mechanisms (explanations, user interfaces features, open source software), and reporting procedures (data sheets, model cards, system cards), it has so far rarely been investigated in a holistic fashion incorporating interdisciplinary perspectives on the subject.”
✭Artificial Intelligence and Education: A critical view through the lens of human rights, democracy and the rule of law “Artificial intelligence (Al) is increasingly having an impact on education, bringing opportunities as well as numerous challenges. These observations were noted by the Council of Europe’s Committee of Ministers in 2019 and led to the commissioning of this report, which sets out to examine the connections between Al and education (AI&ED). In particular, the report presents an overview of AI&ED seen through the lens of the Council of Europe values of human rights, democracy and the rule of law; and it provides a critical analysis of the academic evidence and the myths and hype. ~ The Covid-19 pandemic school shutdowns triggered a rushed adoption of educational technology, which increasingly includes AI-assisted classrooms tools (AIED). This AIED, which by definition is designed to influence child development, also impacts on critical issues such as privacy, agency and human dignity – all of which are yet to be fully explored and addressed. But AI&ED is not only about teaching and learning with AI, but also teaching and learning about AI (AI literacy), addressing both the technological dimension and the often-forgotten human dimension of AI.”
⛲Foundational Revelations: Safe Superintelligence Inc (SSI), Claude 3.5 Sonnet, Gen-3 Alpha (Runway), Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks
✭Safe Superintelligence Inc (SSI) “Superintelligence is within reach. Building safe superintelligence (SSI) is the most important technical problem of our time. We have started the world’s first straight-shot SSI lab, with one goal and one product: a safe superintelligence… We approach safety and capabilities in tandem, as technical problems to be solved through revolutionary engineering and scientific breakthroughs. We plan to advance capabilities as fast as possible while making sure our safety always remains ahead. This way, we can scale in peace… Ilya Sutskever, Daniel Gross, Daniel Levy”
✭Introducing Claude 3.5 Sonnet “Introducing Claude 3.5 Sonnet—our most intelligent model yet. Sonnet now outperforms competitor models and Claude 3 Opus on key evaluations, at twice the speed.”
✭ Introducing Gen-3 Alpha: A New Frontier for Video Generation (Runway) “Gen-3 Alpha is the first of an upcoming series of models trained by Runway on a new infrastructure built for large-scale multimodal training. It is a major improvement in fidelity, consistency, and motion over Gen-2, and a step towards building General World Models.”
✭Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks (Azure AI, Microsoft) “We introduce Florence-2, a novel vision foundation model with a unified, prompt-based representation for a variety of computer vision and vision-language tasks. While existing large vision models excel in transfer learning, they struggle to perform a diversity of tasks with simple instructions, a capability that implies handling the complexity of various spatial hierarchy and semantic granularity. Florence-2 was designed to take text-prompt as task instructions and generate desirable results in text forms, whether it be captioning, object detection, grounding or segmentation. This multi-task learning setup demands large-scale, high-quality annotated data. To this end, we co-developed FLD-5B that consists of 5.4 billion comprehensive visual annotations on 126 million images, using an iterative strategy of automated image annotation and model refinement. We adopted a sequence-to-sequence structure to train Florence-2 to perform versatile and comprehensive vision tasks. Extensive evaluations on numerous tasks demonstrated Florence-2 to be a strong vision foundation model contender with unprecedented zero-shot and fine-tuning capabilities.”
🛠️ Tech: NotebookLM goes global with Slides support and better ways to fact-check (Google), Opening up ChatGPT: LLM openness leaderboard, Slack Combines Abstract Syntax Tree (ASTs) with LLMs to Automatically Convert 80% of 15,000 Unit Tests, Rice Farming Gets an AI Upgrade, TopCVPR-2024-papers
✭NotebookLM goes global with Slides support and better ways to fact-check (Google) “Our AI-powered research and writing assistant is getting a big upgrade and expanding to over 200 countries and territories around the world.”
✭Opening up ChatGPT: LLM openness leaderboard
✭Slack Combines ASTs with Large Language Models to Automatically Convert 80% of 15,000 Unit Tests “Slack's engineering team recently published how it used a large language model (LLM) to automatically convert 15,000 unit and integration tests from Enzyme to React Testing Library (RTL). By combining Abstract Syntax Tree (AST) transformations and AI-powered automation, Slack's innovative approach resulted in an 80% conversion success rate, significantly reducing the manual effort required.
✭Rice Farming Gets an AI Upgrade | Hakai Magazine “Agricultural drones are transforming rice farming in the Mekong River delta, cutting down the amount of pesticides and fertilizers that wash into the ocean in the process.”
🔎 Research: ∇^2 DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials, a ‘Virtual Rodent’ powered by AI to understand the brain, The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing, Data-driven crop growth simulation, AI enables faster, more effective antibiotic treatment of sepsis, Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models (Nature Machine Intelligence), Instruction Pre-Training: Language Models are Supervised Multitask Learners
✭ ∇^2 DFT: A Universal Quantum Chemistry Dataset of Drug-Like Molecules and a Benchmark for Neural Network Potentials “Methods of computational quantum chemistry provide accurate approximations of molecular properties crucial for computer-aided drug discovery and other areas of chemical science. However, high computational complexity limits the scalability of their applications. Neural network potentials (NNPs) are a promising alternative to quantum chemistry methods, but they require large and diverse datasets for training. This work presents a new dataset and benchmark called ∇^2DFT that is based on the nablaDFT. It contains twice as much molecular structures, three times more conformations, new data types and tasks, and state-of-the-art models. The dataset includes energies, forces, 17 molecular properties, Hamiltonian and overlap matrices, and a wavefunction object. All calculations were performed at the DFT level (omegaB97X-D/def2-SVP) for each conformation. Moreover, nabla^2DFT is the first dataset that contains relaxation trajectories for a substantial number of drug-like molecules. We also introduce a novel benchmark for evaluating NNPs in molecular property prediction, Hamiltonian prediction, and conformational optimization tasks. Finally, we propose an extendable framework for training NNPs and implement 10 models within it. ~ The dataset is open sourced and is available via https://github.com/AIRI-Institute/nablaDFT”
✭Google DeepMind on X: "With @Harvard, we built a ‘virtual rodent’ powered by AI to help us better understand how the brain controls movement. 🧠 With deep RL, it learned to operate a biomechanically accurate rat model ✭ A virtual rodent predicts the structure of neural activity across behaviors | Nature “Animals have exquisite control of their bodies, allowing them to perform a diverse range of behaviors. How such control is implemented by the brain, however, remains unclear. Advancing our understanding requires models that can relate principles of control to the structure of neural activity in behaving animals. To facilitate this, we built a ‘virtual rodent’, in which an artificial neural network actuates a biomechanically realistic model of the rat 1 in a physics simulator 2. We used deep reinforcement learning 3–5 to train the virtual agent to imitate the behavior of freely-moving rats, thus allowing us to compare neural activity recorded in real rats to the network activity of a virtual rodent mimicking their behavior. We found that neural activity in the sensorimotor striatum and motor cortex was better predicted by the virtual rodent’s network activity than by any features of the real rat’s movements, consistent with both regions implementing inverse dynamics 6. Furthermore, the network’s latent variability predicted the structure of neural variability across behaviors and afforded robustness in a way consistent with the minimal intervention principle of optimal feedback control 7. These results demonstrate how physical simulation of biomechanically realistic virtual animals can help interpret the structure of neural activity across behavior and relate it to theoretical principles of motor control.”
✭ [2406.10601] The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing “The task of manipulating real image attributes through StyleGAN inversion has been extensively researched. This process involves searching latent variables from a well-trained StyleGAN generator that can synthesize a real image, modifying these latent variables, and then synthesizing an image with the desired edits. A balance must be struck between the quality of the reconstruction and the ability to edit. Earlier studies utilized the low-dimensional W-space for latent search, which facilitated effective editing but struggled with reconstructing intricate details. More recent research has turned to the high-dimensional feature space F, which successfully inverses the input image but loses much of the detail during editing. In this paper, we introduce StyleFeatureEditor -- a novel method that enables editing in both w-latents and F-latents. This technique not only allows for the reconstruction of finer image details but also ensures their preservation during editing. We also present a new training pipeline specifically designed to train our model to accurately edit F-latents. Our method is compared with state-of-the-art encoding approaches, demonstrating that our model excels in terms of reconstruction quality and is capable of editing even challenging out-of-domain examples. Code is available at https://github.com/AIRI-Institute/StyleFeatureEditor?tab=readme-ov-file .”
✭Data-driven crop growth simulation on time-varying generated images using multi-conditional generative adversarial networks | Plant Methods “The realistic generation and simulation of future plant appearances is adequately feasible by multi-conditional CWGAN. The presented framework complements process-based models and overcomes their limitations, such as the reliance on assumptions and the low exact field-localization specificity, by realistic visualizations of the spatial crop development that directly lead to a high explainability of the model predictions.”
✭AI enables faster, more effective antibiotic treatment of sepsis “Keynome gAST, or genomic Antimicrobial Susceptibility Test, bypasses the need for culture growth by analyzing bacterial whole genomes extracted directly from patient blood samples. The interim findings are based on studies that collected samples from four Boston-area hospitals. Unlike traditional methods that rely on known resistance genes, the machine learning algorithms autonomously identify drivers of resistance and susceptibility based on data from a continuously growing large-scale database of more than 75,000 bacterial genomes and 800,000 susceptibility test results (48,000 bacterial genomes and 450,000 susceptibility test results at the time of this study). This allows for rapid and accurate predictions of antimicrobial resistance, revolutionizing sepsis diagnosis and treatment. "The result is a first-of-its-kind demonstration of comprehensive and high-accuracy antimicrobial susceptibility and resistance predictions on direct-from-blood clinical samples."”
✭ Interpreting cis-regulatory mechanisms from genomic deep neural networks using surrogate models | Nature Machine Intelligence “Deep neural networks (DNNs) have greatly advanced the ability to predict genome function from sequence. However, elucidating underlying biological mechanisms from genomic DNNs remains challenging. Existing interpretability methods, such as attribution maps, have their origins in non-biological machine learning applications and therefore have the potential to be improved by incorporating domain-specific interpretation strategies. Here we introduce SQUID (Surrogate Quantitative Interpretability for Deepnets), a genomic DNN interpretability framework based on domain-specific surrogate modelling. SQUID approximates genomic DNNs in user-specified regions of sequence space using surrogate models—simpler quantitative models that have inherently interpretable mathematical forms. SQUID leverages domain knowledge to model cis-regulatory mechanisms in genomic DNNs, in particular by removing the confounding effects that nonlinearities and heteroscedastic noise in functional genomics data can have on model interpretation. Benchmarking analysis on multiple genomic DNNs shows that SQUID, when compared to established interpretability methods, identifies motifs that are more consistent across genomic loci and yields improved single-nucleotide variant-effect predictions. SQUID also supports surrogate models that quantify epistatic interactions within and between cis-regulatory elements, as well as global explanations of cis-regulatory mechanisms across sequence contexts. SQUID thus advances the ability to mechanistically interpret genomic DNNs.” ✭ New computational tool helps interpret AI models in genomics (Phys.org) “SQUID, short for Surrogate Quantitative Interpretability for Deepnets, is a computational tool created by Cold Spring Harbor Laboratory (CSHL) scientists. It's designed to help interpret how AI models analyze the genome. Compared with other analysis tools, SQUID is more consistent, reduces background noise, and can lead to more accurate predictions about the effects of genetic mutations. How does it work so much better? The key, CSHL Assistant Professor Peter Koo says, lies in SQUID's specialized training. "The tools that people use to try to understand these models have been largely coming from other fields like computer vision or natural language processing. While they can be useful, they're not optimal for genomics. What we did with SQUID was leverage decades of quantitative genetics knowledge to help us understand what these deep neural networks are learning,"”
✭Instruction Pre-Training: Language Models are Supervised Multitask Learners “Unsupervised multitask pre-training has been the critical method behind the recent success of language models (LMs). However, supervised multitask learning still holds significant promise, as scaling it in the post-training stage trends towards better generalization. In this paper, we explore supervised multitask pre-training by proposing Instruction Pre-Training, a framework that scalably augments massive raw corpora with instruction-response pairs to pre-train LMs. The instruction-response pairs are generated by an efficient instruction synthesizer built on open-source models. In our experiments, we synthesize 200M instruction-response pairs covering 40+ task categories to verify the effectiveness of Instruction Pre-Training. In pre-training from scratch, Instruction Pre-Training not only consistently enhances pre-trained base models but also benefits more from further instruction tuning. In continual pre-training, Instruction Pre-Training enables Llama3-8B to be comparable to or even outperform Llama3-70B. Our model, code, and data are available at https://github.com/microsoft/LMOps.”
👀Watching: Ilya Sutskever (The birth of AGI will subvert everything | AI can help humans but also cause trouble), Is the Intelligence-Explosion Near? (A Reality Check: Sabine Hossenfelder), Reverse Turing Test Experiment with AIs (Tamalur)
✭AI Won't Be AGI, Until It Can At Least Do This (plus 6 key ways LLMs are being upgraded) (AIExplained on YouTube) “The clearest demonstration yet of why current LLMs are not just ‘scale’ away from general intelligence. First, I’ll go over the dozen ways AI is getting murky, from dodgy marketing to tragedy-of-the-commons slop, Recall privacy violations to bad or delayed demos. But then I’ll touch on over a dozen papers - and real-life deployments, of LLMs, CNNs and more - making the case that we shouldn't throw out the baby with the bathwater. Crucially, I'll cover 6 key approaches that are being developed to drag LLMs toward AGI. this video will hopefully, at the very least, leave you much better informed on the current landscape of AI.”
Ilya Sutskever | The birth of AGI will subvert everything |AI can help humans but also cause trouble
Is the Intelligence-Explosion Near? A Reality Check. Sabine Hossenfelder: “high-compute queries will require security clearance … People engaged in frontier research tend to vastly overestimate the rate at the world can be changed.”
Reverse Turing Test Experiment with AIs
🖲️AI Art-Research: Stable Diffusion 3 Medium, Luma Dream Machine, Kling AI, Poetics of Encryption (Kunsthal Charlottenborg), Glaze, Junie Lau (AIGC), Laurie Anderson on making an AI chatbot of Lou Reed: ‘I’m totally, 100%, sadly addicted’
✭ Stable Diffusion 3 Medium — Stability AI “Stable Diffusion 3 Medium is Stability AI’s most advanced text-to-image open model yet. The small size of this model makes it perfect for running on consumer PCs and laptops as well as enterprise-tier GPUs. It is suitably sized to become the next standard in text-to-image models. The weights are now available under an open non-commercial license and a low-cost Creator License. For large-scale commercial use, please contact us for licensing details. To try Stable Diffusion 3 models, try using the API on the Stability Platform, sign up for a free three-day trial on Stable Assistant, and try Stable Artisan via Discord.”
✭ Luma Dream Machine “Luma has released Dream Machine, an AI model designed to create high-quality, realistic, and fantastical videos from text instructions and images. Built on a scalable, efficient, and multimodal transformer architecture, Dream Machine has been trained directly on videos, enabling it to generate physically accurate, consistent, and action-packed scenes. Starting today, it is available to everyone for free at https://lumalabs.ai/dream-machine.”
✭ Kling AI - Kuaishou Kling Video Model Make Imagination Alive “Kling AI - Make Imagination Alive ~ similar technical route as Sora”
✭Glaze - What is Glaze “Glaze is a system designed to protect human artists by disrupting style mimicry. At a high level, Glaze works by understanding the AI models that are training on human art, and using machine learning algorithms, computing a set of minimal changes to artworks, such that it appears unchanged to human eyes, but appears to AI models like a dramatically different art style. For example, human eyes might find a glazed charcoal portrait with a realism style to be unchanged, but an AI model might see the glazed version as a modern abstract style, a la Jackson Pollock. So when someone then prompts the model to generate art mimicking the charcoal artist, they will get something quite different from what they expected. ~ But you ask, why does this work? Why can't someone just get rid of Glaze's effects by 1) taking a screenshot/photo of the art, 2) cropping the art, 3) filtering for noise/artifacts, 4) reformat/resize/resample the image, 5) compress, 6) smooth out the pixels, 7) add noise to break the pattern? None of those things break Glaze, because it is not a watermark or hidden message (steganography), and it is not brittle. Instead, think of Glaze like a new dimension of the art, one that AI models see but humans do not (like UV light or ultrasonic frequencies), except the dimension itself is hard to locate/compute/reverse engineer. Unless an attack knows exactly the dimension Glaze operates on (it changes and is different on each art piece), it will find it difficult to disrupt Glaze's effects. Read on for more details of how Glaze works and samples of Glazed artwork.”
✭ Poetics of Encryption | Kunsthal Charlottenborg | Udstillingsted for samtidskunst “New technologies, such as artificial intelligence, are often difficult to understand and the subject of much speculation. The exhibition explores the fascinating possibilities and concerns surrounding AI and other emerging tools. The featured artworks address the human image today—its limits and potential in an age of machine learning. ~ In the spring of 2023, more than 1,000 international scholars, scientists and business leaders signed a joint declaration urging a stop to the development of the major AI technologies. But what are they afraid of? And how do some of the leading artists of our time respond to the rapid development of artificial intelligence and its impact on contemporary art and society? ~ Through thought-provoking artworks, the exhibition Poetics of Encryption will present an imaginary landscape containing artistic statements about intelligence, ignorance, information, power and life in the 21st century.”
✭Ideogram: Image Generation for Everyone “Ideogram is a free-to-use AI tool that generates realistic images, posters, logos and more.”
JUNIE LAU - AIGC “Junie Lau is a multidisciplinary creative; she lives and works in London, UK, and Shanghai, China. Her artworks have been exhibited at the Royal Academy of Arts in the UK and featured in "British Vogue' magazine. Junie's music video "The Court" won the global championship in the Stable Diffusion official #Diffuse Together Al short film contest, earning high praise from Peter Gabriel, a six-time Grammy Award winner, and a wide audience. She passionately created the music video for the theme song "Difficult to Determine Between Good and Evil' from the film "I Did It My Way,' a collaboration between Mandarin pop king Andy Lau and music legend George Lam. With her exceptional skills and AI technology, she has crafted music videos that are unique in artistic creation and have pushed AI visual imagery to new heights, receiving high praise from extensive Asian media coverage. As a star artist in the Stability Al official community, Junie also maintains close connections with mainstream Al platforms such as Luma Al and Pika AI. Her exploratory experiments with Nerf and Gaussian Splat have been officially recognized by Luma Al. Additionally, she directed the production of the "e^(i*m) + 1 - 0* AI short film for the Sohu 25th Fashion Awards, which has been shortlisted as one of the top ten finalists at the 2nd Annual International Al Film Festival by Runway. She collaborated with the top 50 AI filmmakers globally to create the world's first 90-minute Al independent feature-length movie, 'Our T2 Remake,' which premiered in Hollywood, USA on March 6, 2024. “Work includes: e^(i*π) + 1 = 0 — Junie Lau & Peter Gabriel - The Court (Dark-Side Mix) (Junie Lau Official Video)
✭ Laurie Anderson on making an AI chatbot of Lou Reed: ‘I’m totally, 100%, sadly addicted’ (The Guardian) “The last time Anderson was in Australia, in March 2020, she spent a week working with the University of Adelaide’s Australian Institute for Machine Learning. Before the pandemic forced her to catch one of the last flights home, they had been exploring language-based AI models and their artistic possibilities, drawing on Anderson’s body of written work. ~ In one experiment, they fed a vast cache of Reed’s writing, songs and interviews into the machine. A decade after his death, the resulting algorithm lets Anderson type in prompts before an AI Reed begins “riffing” written responses back to her, in prose and verse.”
📚Retroactive Readings: Art or Artifice? Large Language Models and the False Promise of Creativity (CHI Conference)
✭ Art or Artifice? Large Language Models and the False Promise of Creativity | Proceedings of the CHI Conference on Human Factors in Computing Systems “Researchers have argued that large language models (LLMs) exhibit high-quality writing capabilities from blogs to stories. However, evaluating objectively the creativity of a piece of writing is challenging. Inspired by the Torrance Test of Creative Thinking (TTCT) [64], which measures creativity as a process, we use the Consensual Assessment Technique [3] and propose Torrance Test of Creative Writing (TTCW) to evaluate creativity as product. TTCW consists of 14 binary tests organized into the original dimensions of Fluency, Flexibility, Originality, and Elaboration. We recruit 10 creative writers and implement a human assessment of 48 stories written either by professional authors or LLMs using TTCW. Our analysis shows that LLM-generated stories pass 3-10X less TTCW tests than stories written by professionals. In addition, we explore the use of LLMs as assessors to automate the TTCW evaluation, revealing that none of the LLMs positively correlate with the expert assessments.”