ABOUT PEW RESEARCH CENTER Pew Research Center is a nonpartisan, nonadvocacy fact tank that informs the public about the issues, attitudes and trends shaping the world. It does not take policy positions. The Center conducts public opinion polling, demographic research, computational social science research and other data-driven research. Pew Research Center is a subsidiary of The Pew Charitable Trusts, its primary funder. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Retrieval-augmented generation (RAG) has become a popular method for grounding large language models (LLMs) in external knowledge. RAG systems typically use an embedding model to encode documents in a knowledge corpus and select those that are most relevant to the user’s query. However, standard retrieval methods often fail to account for context-specific details that can make a big difference in application-specific datasets. In a new paper, researchers at Cornell University introduce “contextual document embeddings,” a technique that improves the performance of embedding models by making them aware of the context in which documents are retrieved. The limitations of bi-encoders The most common approach for document retrieval in RAG is to use “bi-encoders,” where an embedding model creates a fixed representation of each document and stores it in a vector database. During inference, the embedding of the query is calculated and compared to the stored embeddings to find the most relevant documents. Bi-encoders have become a popular choice for document retrieval in RAG systems due to their efficiency and scalability. However, bi-encoders often struggle with nuanced, application-specific datasets because they are trained on generic data. In fact, when it comes to specialized knowledge corpora, they can fall short of classic statistical methods such as BM25 in certain tasks. “Our project started with the study of BM25, an old-school algorithm for text retrieval,” John (Jack) Morris, a doctoral student at Cornell Tech and co-author of the paper, told VentureBeat. “We performed a little analysis and saw that the more out-of-domain the dataset is, the more BM25 outperforms neural networks.” BM25 achieves its flexibility by calculating the weight of each word in the context of the corpus it is indexing. For example, if a word appears in many documents in the knowledge corpus, its weight will be reduced, even if it is an important keyword in other contexts. This allows BM25 to adapt to the specific characteristics of different datasets. “Traditional neural network-based dense retrieval models can’t do this because they just set weights once, based on the training data,” Morris said. “We tried to design an approach that could fix this.” Contextual document embeddings Contextual document embeddings Credit: arXiv The Cornell researchers propose two complementary methods to improve the performance of bi-encoders by adding the notion of context to document embeddings. “If you think about retrieval as a ‘competition’ between documents to see which is most relevant to a given search query, we use ‘context’ to inform the encoder about the other documents that will be in the competition,” Morris said. The first method modifies the training process of the embedding model. The researchers use a technique that groups similar documents before training the embedding model. They then use contrastive learning to train the encoder on distinguishing documents within each cluster.  Contrastive learning is an unsupervised technique where the model is trained to tell the difference between positive and negative examples. By being forced to distinguish between similar documents, the model becomes more sensitive to subtle differences that are important in specific contexts. The second method modifies the architecture of the bi-encoder. The researchers augment the encoder with a mechanism that gives it access to the corpus during the embedding process. This allows the encoder to take into account the context of the document when generating its embedding. The augmented architecture works in two stages. First, it calculates a shared embedding for the cluster to which the document belongs. Then, it combines this shared embedding with the document’s unique features to create a contextualized embedding. This approach enables the model to capture both the general context of the document’s cluster and the specific details that make it unique. The output is still an embedding of the same size as a regular bi-encoder, so it does not require any changes to the retrieval process. The impact of contextual document embeddings The researchers evaluated their method on various benchmarks and found that it consistently outperformed standard bi-encoders of similar sizes, especially in out-of-domain settings where the training and test datasets are significantly different. “Our model should be useful for any domain that’s materially different from the training data, and can be thought of as a cheap replacement for finetuning domain-specific embedding models,” Morris said. The contextual embeddings can be used to improve the performance of RAG systems in different domains. For example, if all of your documents share a structure or context, a normal embedding model would waste space in its embeddings by storing this redundant structure or information.  “Contextual embeddings, on the other hand, can see from the surrounding context that this shared information isn’t useful, and throw it away before deciding exactly what to store in the embedding,” Morris said. The researchers have released a small version of their contextual document embedding model (cde-small-v1). It can be used as a drop-in replacement for popular open-source tools such as HuggingFace and SentenceTransformers to create custom embeddings for different applications. Morris says that contextual embeddings are not limited to text-based models can be extended to other modalities, such as text-to-image architectures. There is also room to improve them with more advanced clustering algorithms and evaluate the effectiveness of the technique at larger scales. source

There can be little doubt that the FAA is paving the way for a framework governing the widespread operation of commercial drones in the U.S. In advance of a definitive ruling on whether commercial drones can operate beyond visual line of sight (BVLOS), the FAA has been busily granting case-by-case permission to drone operators for exactly that. One recent example, just announced, drone company American Robotics has added seven additional sites of operation approved by the FAA for its automated BVLOS drone technology, the Scout System. American Robotics has 10 operational sights across eight U.S. states. “American Robotics is excited to have seven additional sites of operation approved by the FAA. As we continue to build upon our offerings, we look forward to providing current and future customers with the tools needed to unlock scalable, autonomous drone operations that will help propel their businesses and critical industries forward,” says Reese Mozer, co-founder and CEO of American Robotics. “Not only is this a milestone for American Robotics, but it is also another signal that we have reached an inflection point in commercial drone operations in the United States, and American Robotics is proud to be at the forefront of these industry advancements.” Elsewhere in the commercial drone sector, Percepto, which offers autonomous inspection by industrial robotics, recently announced it will deploy autonomous drones to monitor Florida Power & Light’s substations and power distribution grids across the state. The deployment represents the largest commercial autonomous drone project in the world, a staggering feat given the relatively slow pace with which the FAA has moved to adopt a framework. Not surprisingly, both Percepto and American Robotics are on the FAA’s industry-focused BVLOS rulemaking committee, which is tasked with helping the FAA adopt a regulatory framework for wider commercial drone adoption. The companies that invested early in that process are reaping early benefits from their friendly stance toward the FAA.  The FAA previously issued a nationwide waiver for Florida Power & Light to fly Percepto drones for surveillance and inspection purposes at sites owned and serviced by FPL. The Beyond Visual Line of Sight (BVLOS) Aviation Rulemaking Committee (ARC) has provided its regulatory recommendations to fully incorporate highly automated BVLOS operations flights in US national airspace, a process that is expected to take place in the months ahead. For players like Percepto and American Robotics, eventual approval will open the floodgates. The FAA has been very deliberate in its progress (slow in the eyes of some in the sector).  “Every step by American Robotics toward full autonomy is significant: autonomous drones provide continuous, real-time information,” says David Boardman, CEO of Stockpile Reports. “With zero touch, high frequency automated data collection, the bulk materials supply chain will be transformed as we can provide answers to enable real-time decisions at any site. This approval is a critical turning point in addressing the market demand for continuous information.” source

Asian immigrants come from many cultures and origins. Their migration stories are also diverse. The Vietnam War and other conflicts in Southeast Asia brought Vietnamese and other Southeast Asian refugees to the United States. More recently, flows of Asian immigrants, particularly highly skilled immigrants from India and China, came to study and work in the U.S. This chapter explores Asian American immigrants’ backgrounds and their experiences adjusting to life in the U.S. A demographic profile of Asian American immigrants Some 13 million Asian Americans are immigrants, making up 54% of the Asian American population in 2022, according to a Pew Research Center analysis of the Census Bureau’s American Community Survey. The five largest Asian immigrant groups – Indian, Chinese, Filipino, Vietnamese and Korean Americans – make up about 80% of the Asian American immigrant adult population. Additionally, 14% of Asian immigrants are of another single Asian ethnicity and 6% identify with two or more Asian ethnicities. Asian American immigrants’ experiences in the United States are impacted by their diverse demographic backgrounds. Years in the U.S.: About half (51%) of Asian immigrant adults have lived in the country more than 20 years. Some 22% have been in the country for 11 to 20 years, and another 27% have lived in the U.S. for a decade or less. Education: Among those ages 25 and older, 57% have a bachelor’s degree or higher, while 28% have a high school diploma or less. Income: 52% of Asian immigrant adults live in families with a yearly income of $100,000 or more, while 14% are in families with yearly incomes of less than $30,000. Why do Asian immigrants come to the U.S.? Asian immigrants have different motivations for coming to the U.S. According to the Center survey, 28% say they immigrated to the U.S. to be with family, 27% immigrated for economic opportunities, 26% immigrated for educational opportunities, and 7% immigrated due to conflict or persecution in their origin country. Some 4% cited other reasons for coming to the U.S. Asian immigrants’ main reason for coming to the U.S. differs significantly by ethnic groups and by how long ago they arrived in the country. By ethnicity Among Chinese immigrants, 38% say they came for educational opportunities while 31% immigrated to be with family. Among Filipino immigrants, 41% immigrated for economic opportunities and another 41% say they came to be with family. Among Indian immigrants, the most common reasons given were economic opportunities (42%) and educational opportunities (29%). Among Korean immigrants, 38% immigrated to be with family, while 28% and 26% say they came for educational or economic opportunities respectively. Among Vietnamese immigrants, 32% say they came mainly to avoid conflict or persecution in their origin countries, and 29% say they immigrated to be with family. Among immigrants of less populous ethnic groups, about a quarter each say they came for educational opportunities (26%) or to be with family (25%). Some 21% say they came for economic opportunities, and 13% say they came due to conflict or persecution in origin countries. By years in the U.S. Nearly half of those who have been in the country 10 years or less say they came to the U.S. for educational opportunities (44%). Smaller shares of those who have been in the U.S. for 11 to 20 years (32%) and those living in the country more than 20 years (19%) say the same. Those who have been living in the country more than 20 years are most likely to say their main reason for coming was to be with family (36%). Smaller shares of those who have been in the U.S. for 11 to 20 years (25%) or 10 years or less (22%) say the same. Where do Asian immigrants find support when they arrive in the U.S.? Many immigrants face financial challenges when they first arrive in the U.S. due to various factors, such as language barriers and a lack of credit history in the country. In the first six months of living in the U.S., a majority of Asian immigrants (58%) say they received financial assistance in some form. This includes: 52% who say they received financial assistance from family or friends. 15% who received assistance from federal, state or local governments. 10% who received assistance from religious organizations, such as churches and temples. 5% who received assistance from Asian community organizations. 10% who say they received assistance from some other group or person. Still, about a third of Asian immigrants (35%) say they did not receive financial assistance during their first six months living in the U.S. from any of the sources asked about in the survey. By main reason for immigrating     About two-thirds of immigrants who came to the U.S. due to conflict or persecution in origin countries (64%) say they received assistance from the government during their first six months in the U.S. By comparison, only about one-in-ten immigrants who came for other reasons say the same. Immigrants who came to the U.S. for economic opportunities are the least likely to say they received help from any source. About half say this (50%), compared with 85% among those who came to escape conflict, 70% who came for educational opportunities and 60% of those who came to be with family. By ethnicity Vietnamese immigrants and immigrants of less populous ethnic groups are more likely than Chinese, Filipino, Indian and Korean immigrants to say they received financial assistance from the government, religious organizations and Asian community organizations within six months of immigrating. Notably, these are also the groups most likely to cite escaping conflict as one of the main reasons for immigrating to the U.S. 48% of Vietnamese immigrants say they received financial assistance from governments. A quarter of immigrants from less populous immigrant ethnic groups say the same. 28% of Vietnamese immigrants, 19% of immigrants from less populous ethnic groups and 9% of Korean immigrants received assistance from religious organizations when they arrived in the country, compared with fewer than 5% among other ethnic groups.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Third Dimension AI raised $7 million to enable game developers to build 3D game worlds using generative AI. The capital will be used to expand the Third Dimension team, further train generative AI models that convert 2D (images/video) to 3D and to bring to life Third Dimension’s vision of becoming the leading 3D generation company, said Tolga Kart, CEO of Third Dimension, in an interview with GamesBeat. Felicis led the funding round, with participation from Abstract Ventures, MVP, Soma Capital and the Salt Fund. The founders are autonomous vehicle and gaming experts: Kart, Piotr Sokolski and Özgun Pelvan. They created Third Dimension to simplify creation of large-scale 3D environments, whether real or imaginary, and make it as easy as pressing a button. Third Dimension can build worlds for drone piloting. Prior to founding Third Dimension, Kart was vice president of engineering for a synthetic data company, Parallel Domain. Before joining Tesla as technical program management & simulation for autopilot for over two years, Kart was a senior director on Call of Duty at Activision for over seven years. Kart worked on Call of Duty: WWII and Call of Duty: Advanced Warfare while working for Activision’s Sledgehammer Games studio. Sokolski spent four years at Wayve building photorealistic neural simulators for self-driving cars, as well as over three years at Google; and Pelvan is a published machine learning engineer with five published papers about neural networks and data imputation. Third Dimension provides immersive quality, rendering engine ready content that is ready to use by professionals across multiple industries. Third Dimension’s target customers range from The U.S. Military to video game developers to autonomous vehicle companies. “The ability to simulate the real world is one of the last frontiers in solving some very difficult engineering problems”, said Kart. “Video game engines combined with new generative AI technology will not only make creative industries’ content more robust, but will also enable all simulation efforts to represent the real world in the most accurate fashion.” How it works Third Dimension generates worlds for games. Third Dimension wants to build a one-stop tool that accelerates companies’ abilities to create worlds, allows artists to have a baseline to get to quality faster and allows engineers to build fantastical or entirely accurate representations of the real world in high-fidelity. This technology will accelerate workflows of developers from months of time to days or hours and help save millions in expensive graphics development budgets, the company said. In the world that Kart conceives, game artists in the future will stil be creating concepts. They can create a video or a 2D image or a 3D image. The can feed that into the AI engine, which will create a version of the world that is basedon the inspiration. “Now a level designer or a concept artist can block out a world and figure out what this world’s going to look like. They draw a concept. Maybe it takes a couple days. They feed out into a generation system that allows them to generate the world by Third Dimension, and within a day or two, they have a fully playable world,” Kart said. “It’s not only like pitching ideas, but it’s also like supercharging the production process. The core goal here is that it shouldn’t cost a billion dollars to make video games. We will just accelerate the process and therefore make it less expensive to create content at large scale.” Third Dimension generates worlds for autonomous vehicle testing. It won’t be just for video games. It can also be used for autonomous vehicle training, simulation, virtual backgrounds for film and other applications. Kart chose to focus on world generation because it is a large-scale production problem that consumers a ton of resources and makes games very expensive. He showed a demo of taking a 3D image and converting into a simulated game world with 3D geometry. Third Dimension foresees virtual set applications for film and TV. “The goal here is to make it look like real life,” Kart said. “We are going to go after the actual final pixels that are going to end up on the screen. It has to save time and money. Otherwise, it’s not useful for anybody.” Kart said the company is in the midst of research still in terms of how it converts videos into 3D mesh. “We’re creating a large-scale 3D world reconstruction pipeline,” he said. “Third Dimension is going to redefine 3D creation.” says Aydin Senkut, managing partner at Felicis, in a statement. “Their groundbreaking technology, which seamlessly generates precise 3D environments with a single click, is set to transform how engineers and artists create and simulate both real and imagined environments. By accelerating creativity and enhancing the detail in simulations, Third Dimension opens up new possibilities across various industries from defense to video games, and more. We’re so excited to be partnering with this incredibly experienced, veteran team.” Origins Two of Third Dimension’s founders. The company got started earlier this year. The aims is to enable developers to create large scale, usable environments, whether they’re true digital twins of real places, or completely virtual worlds. They wanted to create 3D versions of real world locations like San Francisco. These worlds can be used for games, simulation, military applications and geospatial work. It gets there through a combination of reconstruction, and generation. The work is a combination of Radiance Fields and Diffusion (Nerf, Gaussian Splatting and Image/Video Generation). With the workflow in the future, Third Dimension envisions the creator starts with an image, either drawn by the creator generated. The creator can also have a 3D block of the world, using the creator’s own inspiration or source image. That gets converted directly to 3D using Third Dimension’s tech, and it’s ready to be loaded into a game engine. Kart started working on this in early 2023. Third Dimension will also focus on military

Chipotle Some of the most interesting tech development is happening in an unusual space: Fast food. From burger-flipping robots to drone delivery and Amazon-level logistics, the face of fast food is changing quickly.  The latest proof? Chipotle has announced a new venture fund called Cultivate Next, which will make early-stage investments into strategically aligned companies. Why does a quick-serve chain need a venture fund? I put the question to Chipotle CTO Curt Garner. “Cultivate Next aims to support seed to Series B stage companies that can accelerate our strategic priorities such as running great restaurants,” says Garner, “amplifying technology and innovation, further advancing our Food With Integrity mission, and expanding access and convenience for our consumers.” The subtext is that competition is fierce in fast food, and it pays to be in on the ground floor of technological innovation — a lesson the sector might well have gleaned from Amazon’s ambitious takeover of Kiva Robotics, which was a big key in unlocking Amazon’s logistical competitive advantage. “Cultivate Next allows us to meet consumer and employee preferences that have evolved over the last two years,” says Garner. “We have an aggressive goal of achieving 7,000 restaurants, and technology is the key to accelerating these growth plans.” Also: Are ghost kitchens here to stay? Chipotle is already making headway, teaming up with companies like Miso Robotics on a tortilla chip-making robot, which Chipotle is piloting at select locations. The chain is also testing RFID technology for backend management, which is critical to maintaining quality in a high throughput kitchen.  “Chipotle is testing radio-frequency identification (RFID) technology to enhance its traceability program and inventory management systems,” says Garner. “Ingredients arrive at Chipotle restaurants affixed with RFID enabled case labels and are scanned by RFID readers. Our RFID program is designed to allow the company to act on food safety and quality concerns swiftly, efficiently, and precisely.” The new venture fund will have an initial size of $50 million and will be financed solely by Chipotle. As funds go, it’s not the biggest, but this is surely a sign of growing competition and white-hot development in an industry largely aided by pandemic-influenced consumer trends but also reeling from an extremely competitive labor market and rising wages. In that regard, it’s a bellwether and sign of lively development in fast food tech. “We are looking to support a wide range of forward-thinking ventures, including those focused on farming, supply chain, employee experience, and advanced robotics.” source

If one event demonstrated how vulnerable organisations and infrastructure around the world are to software vulnerabilities, it was Log4j. The critical zero-day vulnerability in the Java logging library Apache Log4j enabled attackers to remotely execute code to gain access to devices and networks. And because the open-source software was embedded in a vast array of applications, services and enterprise software tools, it had the potential for widespread and long-term disruption. No wonder director of US cybersecurity and infrastructure agency CISA Jen Easterly described the vulnerability as “one of the most serious that I’ve seen in my entire career, if not the most serious”. Security patches were quickly developed and organisations quickly moved to apply them, although the ubiquitous nature of Log4j’s open-source code means there will be software and applications out there which won’t receive the update, especially if nobody realises Log4j was part of the development process. Log4j is just one example of severe security vulnerabilities being uncovered in software that has been used for years – and it came 20 years on from when then-Microsoft boss Bill Gates issued his Trustworthy Computing memo, which urged Microsoft’s developers to produce more secure software after various bugs and security holes were uncovered in its operating systems and products. “Eventually, our software should be so fundamentally secure that customers never even worry about it,” wrote Gates. Two decades on, and while Microsoft Windows is generally regarded as a pretty secure operating system, when used correctly and security updates are applied, even Microsoft can’t escape critical vulnerabilities in the code. And more broadly there is still far too much insecure software around.  Software has always shipped with bugs, but software and services have become ever more important to our everyday lives, making the potential impact of security vulnerabilities even more damaging. In many ways, software development hasn’t evolved to face this new reality: products are still rolled out, only for vulnerabilities — sometimes major ones — to be discovered much later. And when it involves a somewhat obscure component like Log4j, organisations might not even be certain if they’re affected or not. “Inherently, the way in which we do software development just lends itself towards bugs and defects,” says Rob Juncker, CTO and head of software development teams at Code42, a software security company. “The accelerated pace of work that we live in contradicts most security teams’ best practices”. Cybersecurity wants to make software secure, a process that needs investment, personnel and time. That often flies in the face of what companies who build software require: they want to make sure the code is functional and to get it out there as soon as possible, especially if new products or features are depending on it. SEE: A winning strategy for cybersecurity (ZDNet special report)  The state of security is massively uneven across the industry, with pretty good security at some of the top vendors, but the vast majority — even ones that are very well funded — lacking basic security investments, says Katie Moussouris, CEO of Luta Security. “Unfortunately we’ve seen an under investment in cybersecurity over the last 20 to 30 years,” she says. What companies need to do is ensure that cybersecurity is baked in from the very start and features as the building blocks of a software development program at every step of the way — that way all the risks and potential risks can be considered and acted upon before they become problems down the line. “If you think about how software is made and deployed and maintained, it’s a whole supply chain. And it starts out with when you’re designing software or you’re thinking about new features,” says Jonathan Knudsen, senior security strategist at Synopsys, a software security firm. “In the design phase, you have to be thinking about security, you have to do threat modelling or architectural risk assessments, so before you write any code you’re just thinking about how it’s going to work, and what it’s going to do — and how it could be attacked,” he added. SEE: Cybersecurity: Let’s get tactical (ZDNet special report) Bosses might be reluctant to spend the extra time and resources on ensuring code gets delivered securely, but in the long run, it should be the most effective approach, both in terms of cost and reputation. It’s safer to ensure the code is secure before it’s pushed out, rather than having to deliver a critical update later on, which might not even be applied by users. The problem is that many organisations are so used to a development model where speed is key, and the risks to them of producing poor code are seen as relatively low. That could mean more hands-on intervention is needed in order to encourage secure code — and penalise those who wilfully ignore security issues. “In other industries where we have such a critical dependence we regulated those industries, but software has remained largely unregulated, so there’s no software liability laws,” says Moussouris. There has been some movement in this area: for example, the UK government has proposed legislation that will require Internet of Things device manufacturers to follow a set of software security rules before the products can be sold. However, government moves at a slower pace than the industry and even if the rules are enforced, there’s already plenty of IoT software out there that wouldn’t meet the requirements. But as organisations and individuals become more aware of cybersecurity issues, it could be the case that the market forces organisations to take software more seriously — leaving software developers who don’t think about security left behind. “Globally we’re getting more aware about software security, and so I think this is going to translate into buyers asking tougher questions from their builders,” says Knudsen. It’s, therefore, vital for software developers, their customers and even society as a whole, that software security is taken seriously. Perhaps ‘move fast and fix things’ could be a new motto for developers to aspire to. MORE ON CYBERSECURITY source

Getty Images I sometimes wonder how often managers in tech look at their direct reports and bet on who will quit first. And next. To be a tech employee is to be coveted and cosseted. To quit, however, is to be shunned. You are, after all, causing a problem for your bosses — and offering a reflection of their management skills.  That, at least, was always my impression. Yet I was recently assaulted by a curiously uplifting tale from, of all curious uplifting companies: Microsoft. Ben Armstrong, group program manager of Microsoft’s Azure Kubernetes Service, was so proud of his company that he had to deposit the story in everyone’s most pride-filled arena: Twitter. He told of an employee who quit to go to a rival. Tech companies would much rather you created a freakish startup — in which they can invest — rather than go to some dreaded enemy. Armstrong says he told the employee that as they were going to a rival company, their departure would likely be fast-tracked. On the employee’s last day, however, there was a family emergency. The employee needed to get on a plane immediately and fly to another country. Here was the problem: the employee was on an H1-B visa. So, if they flew without effective employment, America wouldn’t let them back in. The employee asked if there was anything Microsoft could do to help. I can think of one or two companies that would say: “Sorry, see ya, wouldn’t wanna be ya.” Armstrong himself wasn’t optimistic: “I told them we could try, but I was not hopeful.” Oh, Ben. You didn’t think Microsoft is a compassionate company? Bill Gates doesn’t work there anymore. Armstrong seems to have been surprised at the company’s alertness to another human’s plight. He said: “Two hours later we were in a call with an HR Director at MSFT; who immediately agreed that while this was against MSFT policy, that was not important here. What was important was that this person needed to be with their family.” And so Microsoft agreed, despite the paperwork of departure having already been signed, to keep their departing employee for another week. I confess I found this oddly heartening. To put aside any potential resentment and to consider the simple human situation was surprisingly commendable. Naturally, there were various Twittered positions. Many praised Microsoft’s readiness to break its own rules for the sake of a departing employee. Some suggested it was a fine way to make that employee feel they could return to Microsoft one day. One offered that this behavior may not have always been associated with Microsoft in the past. Scott Rich, the senior security engineer for Sentinel One Partnerships, mused: “When I announced my intent to leave MS, the CISO and Security Director stopped talking to me overnight. 2 weeks later, I ran into one of them where I was told in a spiteful voice ‘good luck’. “ He added: “2 years later we had the most successful cybersecurity IPO in history.” I was especially moved, though, by a comment from a rival. Massimo Re Ferrè, who styles himself as chief psychology officer, container team at AWS Cloud, offered this wise perspective: “My comment is not directed to MS but it’s sad that we live in times where it would have been ‘normal’ to do nothing and ‘amazing’ to do what it is mere common sense and minimum level of humanity.” I fear some might want to remind him that AWS is part of Amazon. More importantly, he’s right. The very fact that this act was, in some way, extraordinary does offer a dim view of where the corporate world has sunk. Too often, tech companies utter rote platitudes about their human focus, yet think nothing of instantly firing employees. Collectively, on a Zoom call. Too often, they have policies that say: “Hey, you quit, so too bad.” Which, I suppose, Microsoft still does. source

Adult TikTok users in the U.S. use the platform to follow pop culture and entertainment accounts much more than news and politics (Michael M. Santiago/Getty Images) This study seeks to better understand the accounts that U.S. adults choose to follow on TikTok. The TikTok user experience happens largely within the site’s For You page, a feed that is “unique and tailored to each specific individual” based on many factors, such as the behavior they display while on the site. The Following page, by contrast, is constructed directly from the contents of accounts the user follows. However, user interactions with posts from the accounts they follow play a nontrivial role in shaping their For You page. And studying these followed accounts can give us a better understanding on the content that users actively choose to look for on the platform. To conduct this analysis, we started with a representative sample of U.S. adult TikTok users who gave us a valid account handle (their unique account name preceded by an “@” sign) for research purposes. All of these users are members of the Center’s American Trends Panel (ATP). For the 664 such users whose profile publicly displays the accounts they follow, we collected a list of all their followed accounts. That produced an initial list of 227,946 unique accounts. We then collected any available profile information for those accounts, such as their display name, bio and the number of followers they have. This collection occurred April 8-16, 2024. We also collected up to five of their most recent posts (if available). This content data collection was conducted June 14-20, 2024. Using a combination of human coding and machine classification with Large Language Models (LLMs), we then categorized those accounts into categories based on whom the account belongs to and the type of content they post. For more details on the categories we included in the analysis and how this data collection and classification was conducted, refer to the methodology of this report. There are a variety of ways to categorize the different types of prominent accounts present on a social media platform. Here are some of the terms and definitions we have adopted for this study: Followed accounts – Any TikTok account followed by any given user. The accounts a user follows on TikTok appear in that user’s “Following” list. Following page – A content feed on TikTok that consists solely of the content posted by the accounts that a given user follows. For You page (FYP) – A content feed on TikTok that is algorithmically curated to each user, based on their interests and behaviors on the platform. The For You page is the default feed that is served to users as they visit the platform. The FYP may include posts from accounts that a given user follows, but it typically contains recommended content from other accounts beyond the user’s following list. Influencers and content creators – Used interchangeably to refer to accounts with at least 5,000 followers on TikTok who attained their popularity primarily due to their presence on the internet, especially on social media (often described as “internet-native”), as opposed to those with a significant level of public awareness outside of social media (such as movie stars, professional athletes or politicians). Mega influencers and internet celebrities – Refers to influencer or creator accounts with at least 1,000,000 followers on TikTok. Mid-tier individual influencers and creators – Refers to influencer or creator accounts that appear to belong to an individual who have between 5,000 to 1,000,000 followers on TikTok. Small accounts – Refers to accounts with fewer than 5,000 followers on TikTok. These accounts are typically maintained by individual users as a personal profile on the site, are more often set to private, or often have not posted very much content. Entertainers, celebrities and other pop culture personalities – Refers to accounts that appear to belong to traditional celebrities and notable figures from pop culture or the entertainment industry, including movie stars, bands or musicians, professional athletes, comedians and more, who attained their fame primarily outside of social media. Journalists, pundits and media outlets – Refers to accounts that belong to professionals in the news media, either the official accounts of specific outlets or shows, or individual journalists, commentators or pundits. For this study, “journalists” are individuals with a current affiliation to a news organization that is listed in their account bio. Politicians, public officials and government agencies – Refers to accounts that belong to either government agencies, or individual politicians or officials. For this study, “politicians” are individuals who have ever held elected office or are currently running for elected office. Accounts that post about … – Topic labels in this study, such as “news,” “politics” or “pop culture and entertainment,” refer to accounts that were observed mentioning a given topic in their most recent posts as of the content collection period for this study (June 14-20, 2024). Accounts do not need to primarily post about a given topic to receive any of the content topic categorizations used in this study. Any mention is sufficient. For further discussion of the account type and topic categories used in the study, refer to the report methodology. A new Pew Research Center analysis of the accounts Americans follow on TikTok highlights the centrality of internet-native content creators, prominent influencers and traditional celebrities on the popular short-form video platform. It also finds that users choose to follow far more accounts that post about pop culture and entertainment than those posting about news or politics. To conduct this analysis, we surveyed a nationally representative group of U.S. adults who gave us access to their TikTok handles and identified all the accounts those users follow. We then categorized all of those followed accounts based on who they are and what sorts of things they tend to post about. These are some of the main findings: What types of accounts do U.S. adults follow on TikTok? Broadly, they follow lots of creators and influencers who

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Black Forest Labs (BFL), a startup founded by the creators of the popular Stable Diffusion AI image generation model that underpins many AI image generation apps and services (such as Midjourney), has announced the release of a new, faster text-to-image model called Flux 1.1 Pro, and with it, a paid application programming interface (API) on which developers can build third-party apps powered by the model (or incorporate it into their existing apps). This means that a company that offers creative tools can add Flux as an option to their offerings, if they (and by extension, their end users) are willing to pay the API costs. Individual users can access the new Flux 1.1 Pro model not through Black Forest Labs’s site, but rather, through partners together.ai, Replicate, fal.ai, and Freepik. Some of these services refer to the model under a different name, such as “Flux Fast.” No details were immediately provided about Flux 1.1 Pro’s training dataset, an issue of contention for generative AI companies with the original Stability AI and rival Midjourney being sued by artists who accuse the firms and others of violating their copyright by scraping and training en masse without consent or compensation on human-created images posted to the web. One key class action lawsuit against Stability AI and Midjourney remains in court. The news comes following the success of Flux’s initial open source text-to-image AI model which powers Elon Musk’s Grok 2 chatbot from xAI and available to subscribers of his social network X. Unlike its earlier model Flux.1, which was open source and free for anyone to download, fine-tune, customize, and otherwise use for all commercial or personal uses as they saw fit, the new Flux 1.1 Pro model appears to be, like Flux 1.0 Pro, a paid proprietary offering only. However, it is still available for commercial and enterprise usage. BFL sees the launch of its API and Flux 1.1 Pro as major steps in its growth as a company, offering both developers and enterprises access to powerful and customizable tools for image generation. Codenamed “Blueberry,” Flux 1.1 Pro takes the new top spot on the Artificial Analysis image arena leaderboard Flux 1.1 Pro improves on the earlier Flux 1.0 Pro model by delivering six times faster generation speeds, while also enhancing image quality, prompt adherence, and diversity. It enables workflows that prioritize speed without sacrificing quality, generating output three times faster than its predecessor. Additionally, BFL announced an update for the original Flux 1.0 Pro, doubling its generation speed to improve efficiency across the board. The performance of Flux 1.1 Pro has been validated through its secret debut on Artificial Analysis, an independent third party benchmark platform for comparing AI model performance, where the model was tested in the days prior to today’s announcement under the code name “blueberry.” (Some erroneously speculated on X that this was OpenAI testing Sora following its tests of the o1 LLM as “strawberry.”) As of October 1, 2024, Flux 1.1 Pro holds the highest ELO score on the platform at 1153, surpassing other generative models in terms of visual fidelity and prompt accuracy, including Midjourney 6.1 (ELO score of 1100) and Ideogram v2 (score of 1108). The ELO third-party benchmark was established earlier this summer of 2024 by Artificial Analysis co-founder and CEO Micah Hill-Smith and co-founder and Product Lead George Cameron, and uses human ratings of pairs of images to derive its scores. For users demanding high-resolution outputs, Flux 1.1 Pro will soon support ultra-high-resolution images (up to 2k), maintaining its precision and speed through upcoming API updates. BFL API offers developers AI image generation starting at 4 cents per image Complementing the Flux 1.1 Pro release is the BFL API in beta, which brings BFL’s generative capabilities directly to businesses and developers looking to integrate state-of-the-art image generation into their own applications. The API offers advanced customization, enabling users to adjust model choice, resolution, and content moderation to meet their specific needs. It also promises scalability, making it suitable for projects ranging from small-scale to enterprise-level. BFL’s API comes with competitive pricing, making it attractive for users seeking high-quality outputs without excessive costs. For example, the Flux 1.1 Pro image generation is priced at USD $0.04 per image, while the older Flux 1.0 Pro is available at $0.05 per image. Developers can begin integrating the API today, and BFL promises ongoing improvements as the beta progresses. The company envisions its API opening the door to countless creative applications, especially in industries like design, advertising, and entertainment, where demand for high-quality AI-generated media continues to grow. Building on initial strong success Black Forest Labs is no stranger to the spotlight. Just two months earlier, the company secured $31 million in seed funding, led by Andreessen Horowitz (a16z), with backing from high-profile investors such as Brendan Iribe, Michael Ovitz, and Garry Tan. As reported by VentureBeat, the launch of BFL and its earlier Flux 1.0 model was widely seen as a milestone in the AI community. BFL co-founders Robin Rombach, Patrick Esser, and Andreas Blattmann brought their expertise from Stability AI, the team behind Stable Diffusion, into this new venture, with a vision for more accessible, open-source generative AI tools. Flux 1.0, which came in three variants (Flux 1.0 Pro, Flux 1.0 Dev, and Flux 1.0 Schnell), gained early praise for its 12-billion parameter architecture and its ability to match or even surpass the output quality of competing models like MidJourney and DALL-E. The open-source nature of these models, especially Flux 1.0 Dev and Flux 1.0 Schnell, positioned BFL as a critical player in the debate over open-source versus proprietary AI. Industry context and competition Black Forest Labs’ move to launch Flux 1.1 Pro comes at a time of heightened competition in the generative AI media space, with many creators looking to harness text-to-image AI models alongside image-to-video models such as those from Pika, Runway, and Luma. Midjourney and