 |
|
|
|||||
Its hard to believe that just a few short years ago a video of Will Smith eating spaghetti generated by ModelScope, a text-to-video AI model, was the peak of AI slop. Fast-forward to today and our trust for CCTV footage of cute animals has been eroded, slop is showing up across marketing and music playlists, and Sora 2 deepfakes are fooling both grandparents and politicians nationwide. A number of artist projects are fighting back against the deluge of slop polluting the shared waters of the internet (or at least poking fun at those who willingly consume it). Steve Nasopoulos and Peter Henningsen, both freelance copywriters, recently created the Slop Trough in their spare time. Its a digital feeding trough that serves up endless slop, so long as you turn on your webcam and get down on all fours like a good little piggy. Are you a little piggy who needs your slop? the homepage asks. Click yes and it tells you to get on the ground on all fours oink oink. We just wanted to capture the degrading feeling of having someone put this horrible content in front of us and actually expect us to consume it. It feels, how shall we say, a little dehumanizing? the creators told Fast Company. The internet was once a magical place, because it was full of weirdos making bizarre websites and stupid art projects. Slop and AI content are diametrically opposed to that because its mass-produced garbage made by robots. Other online art projects imagine an internet untouched by generative AI. 404 media recently reported on Slop Evader, a browser tool created by artist and researcher Tega Brain that filters web searches to include only results from before November 30, 2022the day ChatGPT was released to (or, rather, unleashed on) the public. The term AI slop itself emerged around 2023, when platforms like ChatGPT and DALL-E became publicly available and more widely adopted, according to Google Trends. Yet concerns about AI among U.S. adults have grown exponentially since 2021, according to the Pew Research Center, so much so that slurs for robots now exist. But for every new AI slop video created, there will always be those resisting it with human-made projects. As Nasopoulos and Henningsen put it: We think humans making stuff and putting it on the internet is what the internet was designed for, so the more of that the better.
Category:
E-Commerce
From the latest skyscraper in a Chinese megalopolis to a sixfoottall yurt in Inner Mongolia, researchers at the Technical University of Munich claim they have created a map of all buildings worldwide: 2.75 billion building models set in highresolution 3D with a level of precision never before recorded. Made from years of satellite data analysis by machinelearning algorithms, the model reflects a sustained effort to capture the built world in three dimensions. The result now provides a crucial basis for climate research and for tracking progress toward global sustainable development goals, according to the scientists behind it. Professor Xiaoxiang Zhu, who leads the project and is the chair of data science in Earth observation at TUM, says the real achievement is that the new map is a threedimensional picture of how much space people actually inhabit. 3D building information provides a much more accurate picture of urbanization and poverty than traditional 2D maps, she explains. With 3D models we see not only the footprint but also the volume of each building. [Screenshot: FC] At the heart of this work is the GlobalBuildingAtlas, an open dataset that describes individual buildings across the planet both as 2D outlines and as simple 3D objects. In total, it contains 2.75 billion building footprintspolygons tracing the edges of each structurecovering every building the satellites could detect in satellite imagery from 2019. [Screenshot: courtesy of the author] At first glance, there are some interesting takeaways from the map, like the distribution of building volume clusters around major metropolitan regionswith particularly dense concentrations in East Asia, Europe, and North America. Meanwhile, many parts of the Global South show vast numbers of buildings that are small and lowrise, especially in Africa, which has more buildings than Europe and North America, but far less total built area and volume. [Screenshot: FC] The ability to map building height and volume reveals disparities that conventional 2D maps tend to hide: A dense informal settlement and a carefully planned neighborhood of multistory buildings can look similar in a flat, areabased statistic. But if you have accurate 3D buildings, experts can understand that they offer radically different housing conditions and require different infrastructure. Their proposed metric of building volume per capita turns the GlobalBuildingAtlas into a lens for spotting where housing and infrastructure lag behind population and, therefore, where urban policy and investment should concentrate. [Screenshot: FC] How they made it The scientists used machine learning algorithms to identify one billion more buildings than any previous global database, creating simplified 3D “shoebox” models for 97% of them. That’s 2.68 billion 3D buildings, compared to Google Open Buildings, which has 1.8 billion building outlines. The team started with daily satellite images from the PlanetScope constellation, which photographs the Earth at roughly 9.8 feet per pixel. Then they stitched together about 800,000 cloud-free scenes from 2019 into a seamless global mosaic, and taught a neural network to recognize buildings by training it on known building outlines from OpenStreetMap and other sources. To add height to these flat building outlines, the team used laser measurements (LiDAR) from airborne surveys in developed countries to train an AI that can estimate how tall a building is just by looking at a single satellite photosimilar to how a person can judge a skyscraper’s height from its appearance and shadow. This height-prediction model scans the entire global image and assigns a height value to every pixel, even calculating its own margin of error.
Category:
E-Commerce
Instacart just became the first company to offer an end-to-end integrated shopping experience with OpenAIs ChatGPT. Its yet another signal that AI is about to upend the way we shopand, maybe, the way we cook. The new partnership was announced by Instacart and OpenAI on December 8. To use the interface, ChatGPT users need to make an Instacart account and then surface Instacart within their chat thread using a prompt like, Instacart, help me shop for apple pie ingredients. From there, they can discuss recipes, ingredient swaps, and their preferred store with ChatGPT, which will help them order all of the items they need from Instacart without ever changing tabs or leaving the chat. [Image: Instacart] This partnership is a significant milestone in the race among tech companies to make AI an integral part of the shopping experience. Amazon, for example, now offers a suite of AI tools to help shoppers make decisions and point them toward future purchases. According to Adobe Digital Insights 2025 report on holiday season shopping, the company saw the first material surge in AI-directed traffic (users following links recommended by chatbots like ChatGPT and Google Gemini) to U.S. retail sites in 2024. This year, it expects AI traffic to rise by 520%. In all, Adobe found that over a third of shoppers in the U.S. have used AI to help with online shoppingand that number is bound to keep growing. Clearly, many shoppers are already turning to ChatGPT for advice on the best products to buy and where to get them. For OpenAI, then, it makes sense to bring the shopping itself directly onto its own platform. In all likelihood, this partnership with Instacart is only a trial run ahead of plenty more integrations to come. In a press release, Nick Turley, head of ChatGPT, said that the new collaboration will allow users to go from meal planning to checkout in a single, seamless conversation. I decided to put Turley’s promise to the test by using the new interface the way I predict that its target audience might: recreating a TikTok-viral recipe (Ina Gartens brownie pudding) from start to finish. Testing out ChatGPT’s recipe-generating chops Making a trendy recipe with the new Instacart integration starts with actually getting ChatGPT to accurately reproduce its ingredients and instructionswhich, as it turns out, can be a challenge. Based on my testing, ChatGPT is pretty good at regurgitating more general, nonspecific recipes from the open web. For example, a search for a popular, gooey chocolate chip cookie yields a standard recipe that ChatGPT describes as similar to The New York Times or Nestlé Toll House; while a search for green goddess salad yields a recipe that went viral in 2022 and has since resulted in dozens of publicly available articles, which ChatGPT is then able to pull from for its own summary. Things get a bit trickier when youre looking for one specific recipe, thoughespecially if it’s protected by a paywall or other blocker. When I asked ChatGPT to find the recipe and instructions for The New York Times Lemon-Tumeric Crinkle Cookies, it confidently provided a slightly inaccurate ingredient list and instructions, and attributed the recipe to the wrong author. I asked the question again, this time including the real author in the prompt, only to be met with the same response with the disclaimer, I cant reproduce the copyrighted article verbatim, but these ingredients + steps accurately reflect the recipe (they didnt). I moved on to attempting to recreate Ina Gartens brownie pudding, starting by asking ChatGPT to use popular TikTok videos to find the recipe. The resulting recipe was almost correct, but not quiteit substituted Gartens recommended framboise liquor for coffee. Next, I specifically requested that ChatGPT use the most-viewed TikTok video about the recipe in order to recreate it. The chatbot told me that it doesnt have access to TikToks live trending videos, so it couldnt pull exact instructions from the most-viewed clip, instead offering a TikTok-style version based on what it called popular adaptations. This version strayed even further from the original. [Screenshot: courtesy of the author] As a last-ditch effort, I asked ChatGPT to pull the brownie pudding recipe directly from Ina Gartens official website. ChatGPT then assured me that it was providing the exact recipe from her site (not an adaptation, not a TikTok version, but her real published recipe). This was, once again, not the real recipe. [Screenshot: courtesy of the author] For OpenAIs model, it seems, finding general recipes on the open web is simple, but accurately retrieving information from external apps, like TikTok, or paywalled websites, like The New York Times, is unpredictable at best. Following this slightly maddening exchange, I decided to bake both Gartens official recipe and ChatGPTs bootleg TikTok-style version in order to decide which reigns supreme. The battle of the brownie puddings After my frustrating back-and-forth with ChatGPT, I was ready to throw in the towel and place my Instacart order as quickly as possible. But the process of actually using the integration proved to be a bit of a rollercoaster. At first, everything was proceeding smoothly. I conducted several test runs using the activation word Instacart, and ChatGPT successfully added my requested ingredients to my cart directly through our chat. Mid-way through this experimentation, though, ChatGPT appeared to lose the plot, informing me, I dont have the ability to directly add items to Instacart or access your account. [Screenshot: courtesy of the author] After several troubleshooting questions, during which ChatGPT informed me that the Instacart connector wasnt active, I asked how to reactivate it. ChatGPT then said that I needed to be in a ChatGPT Plus or Pro plan session with Plugins enabled. In an email to Fast Company, though, an Instacart spokesperson clarified that the integration is available to all accounts, including free ones. [Screenshot: courtesy of the author] Its unclear to me exactly what went wrong, but when I tried again several hours later in a new chat, the connection was up and working again. Ordering the ingredients for the Tik-Tok style recipe was quick and straightforward, and everything arrived from my local Target within two hours (except the unsalted butter, which was substituted for salted due to a store shortage). The recipes themselves were a similar concept with notably different executions. The TikTok style version, for example, called for vanilla extract instead of Gartens seeds from one vanilla bean; likely a result of multiple TikTokers making the swap themselves at home and suggesting it to viewers (vanilla beans in this economy?). Gartens original version also required cocoa powder alone for the chocolate component, whereas ChatGPTs interpretation called for solid chocolate. And, in terms of the baking process, Gartens pudding needed to be suspended in a water bath and baked for an hour, while ChatGPT omitted the water step entirely and suggested just 30 minutes in the oven. Given its presumably crowd-sourced origins, the TikTok-style recipe was unsurprisingly cheaper, easier to make, and quicker. It had an extremely dark, almost bitter chocolate taste compared to the original recipe, which was mellower and sweeter. Both have their place, in my opinionthough Gartens was ever so slightly tastier. Right now, the Instacart integration feels built for people who are already regular users of both ChatGPT and Instcart. For that niche, it might save time when brainstorming for meal prep and troubleshooting general recipes. But for everyone else, Im not sold on the utility of this tool. If you have a specific recipe in mind, its probably easier (and less headache-inducing) to just make it the old-fashioned way. [Photo: courtesy of the author]
Category:
E-Commerce
All news |
||||||||||||||||||
|
||||||||||||||||||