Bits Technology is the powerhouse for customer and business identity needs. With a single integration to our API, you can easily connect to the best service providers and data sources for Know Your Customer and Know Your Business processes.

Bits Technology

An AI Article Not Really About AI “You’re soaking in it.” That’s the sense THE innovation conversation of 2023 gives me about AI (aka generative computing, , , et al). It’s everywhere. The topic of seemingly every current and upcoming tech conference panel? Yes. In daily headlines and non-stop LinkedIn comments? Sure. Core to every pitch deck presently? Of course. Heck, AI will probably be the rage at non-tech trade shows soon enough ( anyone?), and kindergarten kids will declare wanting to be prompt engineers when they grow up. machine/deep learning LLMs Coat Hanger Innovation Summit ‘23 If you listen to the conversations, AI promises a new age that will disrupt our lives in ways we are only beginning to fathom, with impacts beyond the public’s initial forays into ChatGPT-powered term papers and other outputs from its rivals and relatives. Whether this new age is one of machine-driven empowerment and progress, one where dystopian fears are realized and truths are eroded, or some combination thereof is now speculated upon widely and loudly. What’s struck me as much as the predictive scenarios and prospective use cases for AI is the pace at which it and its disruptive potential has entered the . It feels as if this has mainstreamed faster than any prior innovation, including the Internet. It’s one thing to observe folk in the tech world ret-con their creds (and keywords) to claim long-held expertise in and intent to incorporate AI into their platforms, but when late-night talk shows center bits around the technology, elderly, tech-averse relatives speak knowingly of the topic, students come home counseled about ChatGPT and academic integrity, and .ai domains proliferate, we’re seemingly up (or down) whatever curve consultants use to describe their innovation frameworks. Kudos to AI’s PR lobby. zeitgeist We Believe How is it that we have so quickly accepted an AI-infused future as a fait accompli? Recall that past disruptive technologies and services – such as the Web, mobile, e-commerce, and social media – certainly had their boosters and made news about their impact potential when they emerged, but the foreboding sense of how each would profoundly change our personal and professional lives simply wasn’t omnipresent. Instead, prior innovations needed to prove their transformative effects. This required some mix of accomplishments, including but not limited to HW/SW innovation and evolution, infrastructure deployment, acquisition of trust, and user scaling to achieve network effects. As with many technologies that seem to suddenly break through in the public consciousness, foundational work on AI has been underway for decades. Still, AI in all its flavors is still quite early in its mass commercialization. In Internet-equivalent evolutionary time, we are perhaps at the stage when one would be using an NCSA Mosaic browser on a 9600 baud dial-up modem and wait hours to download a Quicktime movie trailer. Sure, it was very cool for its time, but not a capability set that was widely recognized for its game-changing potential. Yet AI is already heralded as the catalyst for a new age. On the surface this is understandable. Today’s AI engines are accessible to the masses and produce highly demonstrable and, equally important, relatable deliverables. Via our phone or laptop, we can readily generate that term paper (though we shouldn’t), home repair instructions, or an article like this [ed. note: no Large Language Models were harmed in the making of this essay]. We can use a few words to prompt images like the ones in this article or generate almost-Hollywood-level VFX for our media creations. And we are entertained by “new” works emulating our favorite musicians, leaving labels and lawyers grappling with a new era of sampling and copyright challenges. For some, AI is already touching their lives directly in more profound ways, as AI models drawing upon massive data sets diagnose otherwise missed cancers and guide physicians’ treatment options toward improved efficacy and away from unnecessary therapies. Further, these AI use cases are enabled by and benefit greatly from prior waves of disruption – such as the Web, mobile, and distributed compute and social communications infrastructures. These technologies enable the amplification of AI, fueling its mainstreaming by disseminating the news, hype, and speculation about our AI future instantly to all corners of the globe. In ways unmatched by the announcements of earlier developments, a crescendo of all things AI dominates our screens. That may explain our mass awareness, but what’s to account for our forgone conclusion about AI’s coming transformative impact? Surely we must have a reason to believe it will be more than ChatGPT-powered scripts and faux-Oasis tracks. I believe we do, and it's simply this: we have normalized disruption. It starts with our technology experiences of the past few decades. After experiencing how the Web, mobile, et al have transformed so many aspects of our lives, they have seasoned us to expect disruption and step-level evolution. Whether we are old enough to remember payphones or have been born into the age of screens and instant everything, when it comes to technology, we have become conditioned to believe. Life, Interrupted Recent, non-technological disruptions have also contributed to our acclimation, demonstrating that change is not only possible, but it can be fast, massive, and tremendously impactful on our daily lives. Today’s politics of course, whatever your beliefs, have brought to the fore stresses and divisions at levels never before felt so strongly or amplified and perpetuated so greatly. While one can readily make the case that these realities have always been present in the fabric of America, it’s hard to deny that the disruption brought about by Trump’s election was stark and impactful and entirely unexpected as the outcome of a Presidential election. No longer. Above all though is COVID. The pandemic was and remains the uber-disruptor on every level of society, a wave that crashed upon each of us directly, laying bare the precariousness of our norms. More than anything it shook our perception of society, safety, and our future in ways with which we are still grappling. In our (barely) post-pandemic state, we have come to understand and expect that disruption is to be expected and its impacts can be immense and potentially permanent, whether we’re speaking of the shift to remote work, an erosion of trust in assumed “solids” such as science and supply chains, or an overall instillment of fear in the new and unknown. Accept The Unexpected Today, whether we’ve seen a wave or three of disruption break upon our shores, or have only lived through the world-shaking events of recent years, I suggest that we have come to internalize an understanding that in our connected world, every successive disruption – technological, political, medical or otherwise – holds the potential to more greatly, strongly, and quickly flip our scripts and change our lives. Through this lens, I can better understand why we are so bullish on AI and have essentially priced its proliferation into our societal P&L. It’s a broad set of disruptive technologies. They appear to work. We expect they will only get better as we have seen with other technologies. And we are now freshly conditioned to expect their disruption to be massive, if unpredictable.

Have We Normalized Disruption?

What Is Quantum Computing and How to Harness It? For starters, quantum computers use subatomic particles as the basic unit for doing computations. To be more exact, they use quantum bits; also known as qubits. “Nature isn’t classical, dammit, and if you want to make a simulation of nature, you’d better make it quantum mechanical, and by golly, it’s a wonderful problem because it doesn’t look so easy.” — Richard Feynman Two of the most important aspects of qubits are and . Entanglement Superposition “One key factor in this development is how quantum mechanics allows two or more particles to exist in what is called an state. What happens to one of the particles in an entangled pair determines what happens to the other particle, even if they are far apart.” — Entangled states from theory to technology. entangled The more entangled bits, the more calculations quantum computers can do. refers to the ability of a quantum system to be in multiple states at the same time . Superposition until it is measured One of the biggest misconceptions concerning superposition is that it applies only to two states, but in reality, superposition refers to two or more states and is not always uniform meaning that the states can have different probabilities. Naturally, the act of measurement will disturb the coherent superposition of states. How Does It Work? You might wonder how qubits are manipulated in quantum computers; the answer is microwaves. The quantum bits are actually quantum particles like electrons or photons which are operated at low temperatures using microwave pulses. Naturally, quantum computers don’t work at room temperature; they need ultracold temperature ( near zero Kelvin, or -273.15 Celsius) Can I Use It? The most popular SDK is which provides the medium for end-users to interface with systems hardware and perform complex computing operations like simulation, optimization, and artificial intelligence. Qiskit At its core Qiskit is a quantum circuit construction, optimization, and execution engine, and by quantum circuit I mean a . Basically, you can mimic the execution of a quantum circuit, and the workflow will consist of four steps: computational routine Build the quantum circuit that models the problem you are considering. Compile the circuit for the desired quantum service: quantum system- the online quantum processors are the systems or local classical simulators. IBM Quantum Run the compiled circuits. Analyze the summary statistics and visualize the results of the experiments. How It Will Evolve and How It Will Impact Society Of course, some of the big players are tech giants ala IBM, Microsoft, Amazon, and Google are already involved. IBM has a rich history of quantum computers; David P. DiVincenzo (former IBM) proposed the minimal requirements for creating a quantum computer, also known as the . DiVincenzo criteria Nonetheless, there are some well-funded startups like Rigetti, IQM, and Quantinuumm that are researching this domain, and in its , IBM has projected that a 1000+ qubit quantum computer will be available in 2023. roadmap Even though the computing power increases exponentially with the number of qubits, quantum computers have fairly high error rates, but the purpose of a quantum computer isn’t classical computing problems; rather, it is more suited for simulations or artificial intelligence tasks. Just think about simulating a chemical reaction; it’s impossible even for the most powerful supercomputer to simulate molecular interaction, i.e., . molecule modeling As a closing note, I would like to share the following quote which I feel encompasses the very essence of quantum computing. “Using quantum principles to compute is as different from classical computing as a classical supercomputer is from an abacus” — William Phillips

Quantum Computing: From Zero to Hero

If you want to understand how humans will eventually treat artificial intelligence now or in the near future, then this 1957 article says it all. "You'll own slaves by 1965," claimed the headline. Granted, this is 60 years old, but in a few years’ time, AI will force us to face our own ethics and morality and in doing so force us to face questions we'd rather not answer. There are people at the moment advocating we need to discuss this in the wake of ChatGPT, arguing that they’re more than a tool. We can’t even agree on a hard definition of (one which requires sentience and experience of the physical world beyond just a screen). AGI We can certainly discuss "robot rights" at the same time as human rights, but we have a woeful history of the latter to suggest we'd not take the former seriously. So do we need to think about or discuss the idea of "rights" right now? It's a good question, and I'll revert to Shelley's Frankenstein or the Star Trek TNG episode, "The Measure of a Man,” as guidance — where the android was put on trial to determine whether he was a thing and property to be dealt with however we choose or something that deserved rights to be protected. In fact, the movie Bicentennial Man also dealt with this question, too — to be recognized as a new type of lifeform in itself with its own sets of rights. But we are not there yet with AI today. Data Will they be the same rights as humans? Possibly not, because of their very nature and how they are constructed. They would be artificial beings, and so we need to define what those criteria are too. We've only recently recognized that Octopii are sentient lifeforms, and it's taken us hundreds of years to do this, so the debate on whether an AI needs rights might also take just as long. The grey area is whether what may emerge from early AI attempts can be construed as something more than a thing someday. What we're used to with our devices like Alexa or Siri or Tesla autopilot is a form of Weak or Narrow AI, something that has been developed to perform one or few tasks but cannot go beyond this or learn more. OpenAI has created something that is more complex and is swiftly moving towards AGI or Strong AI — an artificial intelligence that can handle complex tasks, and multiple instructions, and draw on larger sets of information to complete these tasks, making suggestions based on that information and more. These AI tend to be more incorporeal; they have no physical form, so it’s hard to imagine how something with an apparent intelligence is a slave or needs rights. But as they become more and more advanced, they may express particular desires or needs or appear individualistic beyond their algorithms. We need to be ready to recognize this when it happens. Another question is, does the form factor make something more deserving of rights? Does an AGI hoover deserve more respect than a humanoid-shaped robot with AGI because we find the hoover worthless or comical in comparison to something that reminds us of ourselves? There are videos of children attacking an airport security robot, so that tells us a lot already. Will we anthropomorphize and attach emotions where none exist in the same way we do with other inanimate objects? We need to be objective about how we approach this for sure. And then there's the issue of transhumanists and others who advocate merging minds with machines. Does that make future humans more than humans, with more or fewer rights? A further thing to consider is the Ship of Theseus - at what point, if you continually replace bits of yourself with robotic or AI technology, are you no longer "you"? There is a limit, and we need to define it, but I'm not convinced we are able to on our own. Perhaps that's something an emergent AI can eventually help us with. In all, far too much to think about at a time when there's so much future economic uncertainty in the wake of what's happening today. But a future generation might have to deal with this if we don't. Also published . here

With Rise of ChatGPT, Is Now the Correct Time to Ask Whether AI Should Have Rights?

A lot of software and tools appear on the market pursuing the objective to fulfill a modification of an ordinary learning curve. The Edtech sector is exploding with current modern solutions, which apply different ways how to improve the educational form. In this list, there are . mLearning, adaptive learning, (a)synchronous or hybrid courses, and so on However, the micro-learning concept distinguishes itself from many others. It utilizes the form of bite-sized material, which creates an engaging training experience. Microlearning is a type of eLearning software that works well with mobile learning and gamification. Thus, it is a great learning approach to consider as a part of your application. But let’s thoroughly review this concept. What Is Microlearning? So, what is microlearning? This term is pretty easy to sort out because it is as straightforward as it sounds. And it sounds like this concept envisages the separation of learning blocks into smaller pieces for better material absorption. Microlearning is a notion that entails tiny, concentrated, and condensed learning exercises that may be carried out on a variety of devices. Microlearning has proven to be an excellent method for improving long-term memory and the overall quality of the learning experience. The combination of microlearning with behaviorism, constructivism, cognitivism and experientialism (psychological theories) impacts the way people comprehend fresh knowledge. Surely, such an approach, with its beneficial compact nature, wouldn’t be omitted in the online education industry. As a result, microlearning apps appeared to provide a full-fledged competitive advantage to the other eLearning software with their hybrid courses, adaptive learning, and so on and so forth. They come in handy because they are accessible and convenient to use for the majority of users, whether it is a pupil or an adult learner. One of the most common ways to use microlearning apps is for corporate online education. As a matter of fact, many businesses use micro-content for performance management. In terms of training and development, you might rapidly witness a significant boost in employee engagement thanks to the use of a microlearning course. Given that, workers are better to learn smaller junks of information, which won’t interrupt practical studying of particular responsibilities needed for the job. Whether you decide to develop some microlearning app for onboarding or eLearning software for other target audiences, either way, it will be applied as an LMS (Learning Management System). Yet, out of the range of different LMSes, the ones with the microlearning approach will definitely stand out. But how? Let’s find out together. Launching Microlearning Apps: What will be different from any other eLearning software?
Indeed, the launch and development of microlearning apps will be the same as any other application. But still, take into account some slight nuances that are needed for a proper microlearning approach. The first and the most important part is the of your courses. Because creating the material which is already divided into small sections depends on the content selection. content Then, when you have junks of materials, generate for them. For each piece of information, there should be a question afterward. This way, learners will summarize if they understood or remembered offered information. tests and quizzes Such a method of online education could be full of entertainment and . Don't ignore the incorporation of various types of (flashcards, videos, contests, images, puzzles, etc.). visual materials gamification Additionally, simplify users' access to the micro-courses by developing a of an app. As we have already mentioned, microlearning is greatly combined with mobile learning. mobile version Considering that, learners can get the needed pieces of info with a click, even on the go. Although we defined crucial criteria for microlearning apps’ development, now we can review the features of the top microlearning platforms. What Features Are Common For Top Microlearning Platforms? Indeed, the market is already overwhelmed with great solutions, but to have a competitive advantage, you should combine basic features with unique ones. That is why we outlined a few key characteristics that distinguish the mediocre from the outstanding when it comes to microlearning platforms. The following are the main features to look for: Intuitive design for User Experience Easy-to-use User interface Extensive possibilities for adding extra technologies or tools Availability for team interactions Advanced analytics and reporting tools to help you get the most out of your data Quicker recording, creation, and sharing of training programs, resources, and information mLearning for mainly iOS and Android Well-thought-out content which is clearly separated Listed features we have confirmed in all of the tools that we examine in this post. List Of Top Microlearning Platforms Now, when we overviewed the features that are required for top microlearning platforms, we can name the applications. After all, you should know whom you should be looking up to when developing your own solution. Qualee Technology It's an employee experience system that keeps employees engaged during their employment, involving onboarding, online education, micro-surveys, and offboarding. There is no coding necessary with Qualee: simply drag & drop. As a consequence, even those with little technical skills may find it straightforward to install and use this application. The dashboard has a really simple style, and the simplicity makes it easy for users to follow their progress. This adds to the enjoyment of the learning process. Because Qualee provides real-time data and analytics, businesses can keep a continual eye on employee sentiment. From onboarding through offboarding, this solution is comprehensive in terms of both compliance and employee experience. BrainStorm QuickHelp It's a one-of-a-kind learning solution that combines content expertise with a desire to change the way people engage with Microsoft technology. Pre-packaged instruction in tiny bits to accommodate everyone's schedule. Beneficial is that training duration is for as little as 2 minutes or as much as an hour. It may be used as a learning management system, but you must upload each piece of information separately. You can have automatic orientation skillsets that are allocated to new employees automatically. Bites The Bites platform allows businesses to produce and share professional information with their mobile and frontline workers. It delivers an experience, which is close to social media, making it easy for modern users. With the help of AI, Bites enables to produce of a short "story-like" film and makes it sophisticated. The material may then be shared with workers via current instant messenger platforms (SMS, WhatsApp, MS Teams, and so on) to attain them where they are most engaged. As a result, complex dashboards may be used to track and improve employee engagement, feedback, and performance. UMU UMU's AI-enabled eLearning software, which is based on learning science, helps to drive performance and outcomes. UMU makes creating and disseminating instructional information to a large audience a breeze. Users' buy-in is embedded into the UMU system, which generates excitement around the learning process. New employee training is scalable with UMU. Its platform is easy to get started with, deploy, and configure. Users may access training from any device, at any time. Additionally, the UMU team can assist in optimizing platform usage and improving the learning experience. EdApp It is a free online training platform that offers microlearning software for a different number of users and types of work (office/remote). The microlearning feature in EdApp with its bite-sized courses focuses on only the most important aspects. Its microlearning technique is paired with gamification aspects, allowing you to turn boring courses into fun activities. Moreover, users may also use a system-automated quiz called Brain Boost to reinforce important topics or a quiz creator called Rapid refresh to test their knowledge. GoSkills GoSkills is a customizable learning management system and microlearning software that helps small businesses increase employee engagement. Companies accomplish this by dividing courses into short-form modules and employing gamification approaches. It includes a free authoring tool that allows you to effortlessly build and personalize the corporate identity of the courses. A library of ready-to-use courses is also available on this platform. 7taps This web solution makes it easy to create mobile-first microlearning experiences in seven taps. Anyone may easily develop courses on the platform by using templates. You may also enhance interaction by using GIFs, soundbites, and quizzes. You may then easily share courses by sending a link, SMS, QR code, or email. To begin their course, learners do not need to download any apps or register. Even if you don't take these procedures, you'll be able to measure and assess company employees' learning progress. While this product isn't really a microlearning app, the courses are mobile-friendly and adaptable to a desktop setting. Summarizing Developing an LMS with a microlearning approach requires a deep understanding of this method in order to make it bring just the right outcome and quality learning experience. Content is what makes it really valuable for users, and the correct division of the information into smaller pieces of material. But this should be reinforced by a set of features that will make such an application easy to use, convenient for team usage, and with multiple educational formats. These three steps appeared all the way to this small discovery. And they are definitely worth considering because they all create user engagement, which depends on the whole popularity and success. Now, you can either take these platforms as an example and develop something similar or come up with your idea. First published . here

A Quick Guide to Microlearning Apps

A brief summary of my PhD research into designing better online and offline social networks. [This is post one of three, looking at how important the shape of your social network is. The second looks at , and the third a the politics of .] design strategies for social networks social network data The following is one of the most yawn-inducing statements in the universe: this is what I learned from my PhD. If you’ll give me just two seconds before you close the browser tab, I promise I’m going to climb out of my tiny PhD niche and do what you aren’t allowed to in your thesis — cut to the interesting bits. My research looked at how digital technology is making it possible to measure our social network in unprecedented detail, how digital technology is shaping our social connections, and how we can make decisions with these facts in mind. design This post looks at some of the reasons you should care about social networks — by which I mean the network of social connections between people, whether digital or offline. Social connections powerfully shape our lives — happy or sad, fat or thin, even healthy or sick. I’ll also highlight the relatively low-quality data much historical social network research relies upon, and how things are starting to change. I’ll follow up with two more posts: a second post looking at research on how to design for better social networks; a third on the politics of social network data. One of the discoveries of my PhD is a humbling sense of the tiny amount of research a single person can undertake — even over four years ( ). What I thought would be the bulk of my research — the practical work of building and testing social network analysis software — in fact only functions as the , a plot device to power me on my journey through the huge and dispersed existing literature on social networks. 4 years! MacGuffin I’ll spend the next paragraph outlining my practical work, then move on to the wider picture, which, despite everything ( ), I still find exciting. 4 years! I built and tested software ( ) that analyses and visualises Twitter data to understand local community issues: people upset about when the bins are (not) collected, perceptions of crime, planning issues, etc. It also helped identify ‘community influencers’, civically active local individuals. Building a prototype served as a way into the wider political and ethical issues of how the public sector can use data. localnets.org social media Done. My practical work was about social media data in local government, but I want to zoom out to talk about the importance of social connections in general. In short, people’s social interactions are fantastically important. We know this intuitively, we say that there is an “old boys network”, that it’s “who you know”, and many similar aphorisms. We understand that similar people form networks — “birds of a feather…”. In Chinese culture the term reflects the same concept, referring to how well connected a particular individual is. Guanxi Digital technology makes it easier to measure social networks. It also modifies them. Perhaps the clearest example of technology modifying social connections is online dating — forming intimate connections between people who would never have met otherwise. Dating apps are fundamentally changing the way people meet their life partners, leading to increased marriages between — that is, in the long term, changing demographics. ethnicities If dating apps reinforce a pattern where wealthy people tend to date another, and less wealthy people also disproportionately date each other, it will increase household inequality as rich people couple off and form double-wealthy households. If dating apps causes people to meet outside of their income-based social circles, it could reduce income inequality. Changes in who ends up marrying whom could easily as headline topics like executive pay. So far as I can tell, there is no research looking at income and online dating. contribute as much to income inequality The history of social network research It’s not new to observe that social networks are both important and easily altered. Influential psychologist Leon Festinger published in 1950. 270 MIT students were randomly assigned to dorm rooms across 10 buildings. A follow-up survey asked the students who their three closest friends were. 65% came from the same building, 41% were in adjacent dorm rooms. Being even slightly physically closer to someone makes a social connection to them vastly more likely. This might be obvious, but the strength of the correlation is striking. Many people form their friendships for life at university. How many lives have been shaped by the accident of dorm room assignment? research that captures both these ideas Diagram of a dorm building used in Festinger’s research. In 1967, Stanley Milgram did an experiment to test the ‘6 degrees of separation’ concept, the idea that any random pair of humans can be linked by some small number of friends of friends. The only way to measure the average friend-of-friend path length in 1967 was to send out randomly addressed postcards, explaining the experiment and asking recipients to forward the letter to a designated target (name and approximate address were provided). In the (highly probable) case they didn’t know the target, the recipient was instructed to forward the letter to someone they knew who they thought would be more likely to know the target. By examining the letters that arrived at the target address, Milgram could see that the postcard had been forwarded, on average, 5.7 times, and so concluded that there were, on average, 5.7 social connections between any two Americans. There are problems with this approach. Most letters got lost, decreasing statistical validity of the results, and people don’t know the optimal person to forward the postcard to in order to get it to its destination. The structure of our social connections is a fundamental fact about human society, yet, until the era of the web, it remained almost completely opaque. In 2012, research demonstrated very precisely that the was 4.74 — and that’s across the whole world. equivalent to Milgram’s number in the context of Facebook users The idea of ‘social capital’ — roughly, the aggregate benefit you get from your social connections — goes back to the . We now have a wide range of evidence about the structure of our social connections and correlations with other phenomena. I wanted to highlight some of the most compelling results, but also to point out they are mostly based on weak data. We haven’t come that far since Milgram’s study. Even though the big tech companies almost certainly have exquisitely detailed data, making it available is highly problematic. origin of sociology in the 1850s Contemporary social network research The evidence we do have shows how important social connections are. Loneliness is often said to be a chronic issue in developed countries, and it is . People who say they are lonely and have worse . Loneliness is, in some ways, the social network equivalent of poverty. increasing according to some analyses die younger mental health Social capital correlates and . Surveys also show that our subjective experience of happiness is anchored in our social connections: being richer doesn’t make you happy (past a certain point), but being . Studies have also used this data to test for a link between and found a small positive correlation. with health well being richer than your friends does social capital to GDP growth Negative properties are transmitted by the network too. Being friends with obese people . If your friend becomes obese they can transmit obesity to you almost like an infection, as though their weight gain gave you social permission to follow suit. makes you more likely to be fat But the empirical data around social networks is often weak. We only know about obesity ‘contagion’ because of the Framingham Heart Study, one of the largest medical studies ever conducted, running since 1948 and with over 5,000 participants. This kind of data is hard to come by. Research mostly depends on proxy measures of social capital, for example ‘civic participation’ — membership of sports clubs or participation rates in national elections. Alternatively, surveys which ask questions like “Generally speaking, would you say that most people can be trusted, or that you can’t be too careful in dealing with people?” are assumed to reflect people’s social connections. (This question from the popular World Values Survey). A sociometric badge, worn on a lanyard (from ) Ben Waber Sociometric badges are another way of gathering data about social networks. Sociometric badges are electronic devices worn on lanyards around the neck. They collect metrics such as speech patterns (through the microphone) and where and when those wearing them engage in face to face contact (through the IR transceiver). The badges can measure the offline, non-digital social networks in offices. Using this data researchers were able to identify high value social connections that were missing within the office network. Researchers then intervened, adding those social links by introducing people to one another. Their intervention reduced the and caused staff to describe their workplace as ‘better connected’. Alternatively, you can use the data to see , which could, in turn, inform interventions that might promote gender equality. These studies underline the benefits of high-quality data, and how new technologies are increasingly providing it. A sociometric badge might easily become a smartphone app. number of emails sent how men and women form different networks Which brings us to the enormous privacy issues, the complex ethical terrain, and the political difficulties inherent in studying social networks. At the same time, it’s not a subject that can be ignored — the more technology allows us to understand network formation, the more responsibility we have. The more technology changes social networks, the more responsibility accrues to those making design designing. Social network driven design Sunstein and Thaler’s influential book also argues for those making design decisions to take ethical responsibility, but in the context of behaviour change. They open their book with a thought experiment about a school canteen: where should the fruit go, where should the chocolate bars go? Throughout their book they provide empirical evidence showing how powerful ‘choice architectures’ can be — placing the apples next to the checkout and hiding that chocolate bars can drastically change health outcomes. Once there is empirical evidence about fruit placement, responsibility ensues. There is no neutral setting: merchandising decisions are driving behaviour. Again, as we get an increasingly detailed picture of how social networks form, responsibility ensures. Nudge There are, of course, risks, but there are also benefits. In my work on LocalNets, one unexpected discovery was the importance of local, non-chain shops in the social media ecology. A Costa coffee outlet won’t run it’s own Twitter account, but a local cafe often will, and when it does it often catalyses community social media activity around it. Twitter data gives us an empirical window on something we always suspected — Costa might be very efficient at pumping out coffee, but there is a social value to locally owned shops. The better the evidence for these social network structures, the stronger the argument against market forces which erode communities. Sunstein and Thaler’s nudges have been applied to hypothetical school canteens, , and providing . Social network formation has a similar diversity of applications, from the algorithm behind your dating app to the spread of fake news, from increasing political polarisation to fostering innovation — even if the social network perspective is a more contested and less developed field than the mature nudge ‘industry’. organ donation laws making you pay your tax clean drinking water in Morocco My next post . A final post looks at the politics of . I’ll leave you with Bill Withers telling us about the importance of social networks. will look at some design strategies that can help designers think about social network structures social network data

Technology is shaping our social connections. Designers need to take responsibility.

Every player wants to beat the odds. But few actually do. Have you ever wondered how slot machines in casinos can be beaten? Or how to win in Web3 games that rely on chance? In this article, we explore the role randomness plays in Web3, whether it’s possible to beat random number generators (RNGs), and how to go about cracking RNGs. In the vast expanse of Web3, randomness isn’t just a happenstance. It’s a pivotal element in various scenarios, influencing outcomes that range from gaming experiences to critical blockchain functionalities. Where Randomness Plays a Role in Web3 In Web3, randomness operates as a fundamental force influencing various aspects. One prominent domain is within the gaming sphere, where the chance of unveiling rare in-game items at specific locations is contingent upon the mechanisms of random number generators. These random encounters often dictate the rewarding experience of players upon defeating opponents. Lottery draws a quintessential example of randomness, requiring an unequivocally transparent and verifiably random process, especially within the Web3 ecosystem, where fairness and openness are paramount. If not, the lottery draw may potentially favor only specific groups or players. Additionally, the random selection of block producers through RNGs is crucial in certain layer-one blockchains. If this selection process lacks verifiable randomness and exhibits discernible patterns, block producers might exploit this to their advantage, strategically producing blocks that serve their interests. Such actions compromise the network's security, underlining the critical role of truly random and unbiased selection mechanisms in Web3 blockchain operations. While randomness plays such an important role in Web3 gaming and the Autonomous World, it's crucial to differentiate between pseudo-random number generators (PRNGs), which use algorithms to produce sequences of random numbers, and true random number generators (TRNGs) that use physical processes for higher security. Why Predicting Random Numbers Is Possible Predicting numbers generated by pseudo-random number generators (PRNGs) might be feasible due to specific vulnerabilities and patterns inherent in their algorithms. The generation process of PRNGs follows patterns that can be exploited if the initial seed (starting point) and state (current condition) of the generator are known. Let’s dive deeper into what a “seed” and “state” mean in this context. The seed, a critical input, initiates the random number generation process. It can be something as simple as the current time. The state refers to the current condition of the generator, which evolves with each number generated. Knowing the initial state allows for the prediction of subsequent numbers, presenting a significant security risk in scenarios where unpredictability is crucial. Moreover, the use of centralized sources for RNGs introduces additional vulnerabilities. Centralized systems become single points of failure, making them prime targets for malicious attacks. Historical incidents have demonstrated that brute force attacks can compromise these systems. Attackers can inundate the central server with numerous random guesses and make incremental adjustments to fine-tune their predictions, ultimately increasing their accuracy in guessing the next random number generated. These vulnerabilities emphasize the need for more secure and decentralized RNG solutions in various applications, especially within the Web3 landscape, where data integrity and randomness are of utmost importance. Exploiting RNG Vulnerabilities: Theoretical Methods Finding the Seed Through Hypothesis Testing The initial step in exploiting an RNG is to uncover the seed, which is the starting point for number generation. This process involves formulating and testing various hypotheses about potential seed sources. For example, if the hypothesis for the seed is the time of RNG operation, then generating numbers at consistent times and analyzing the outputs for patterns could confirm or refute this hypothesis. Advanced statistical techniques can be applied to detect subtle patterns that might not be immediately apparent, thereby increasing the accuracy of seed prediction. State Analysis and Reverse Engineering Beyond identifying the seed, understanding the state transitions within the RNG algorithm is crucial. By analyzing how the state changes with each number generated, it's possible to reverse-engineer the RNG algorithm. This method involves sophisticated mathematical and computational techniques, including algorithmic analysis and potentially cryptographic reverse engineering, depending on the complexity of the RNG. Leveraging Machine Learning for Pattern Recognition Machine learning, a subset of artificial intelligence (AI), excels at deciphering complex patterns in large datasets. This capability extends beyond simple pattern recognition, enabling the algorithm to learn and adapt without explicit programming. A prominent example of this is Google's AlphaGo. This AI system, trained on the board game Go, mastered the game by learning from vast quantities of game data, eventually outperforming world champions by anticipating their moves. This principle is directly applicable to analyzing RNG outputs. Machine learning algorithms can process extensive datasets of RNG-generated numbers to detect underlying patterns, including subtle irregularities and algorithmic resets that might not be apparent through traditional analysis methods. By training on a sufficiently large dataset, machine learning models can identify the cyclical nature of these RNGs, including their reset points and output patterns. Understanding these patterns allows for the prediction, and potentially the manipulation, of future RNG outputs. This capability is particularly crucial when RNGs are used in environments where predictability can lead to security vulnerabilities or unfair advantages, such as in gaming or cryptographic applications. Cryptographic Analysis Cryptographic analysis of RNGs is a multi-faceted approach that involves scrutinizing every aspect of their cryptographic design and implementation. This process starts with a thorough examination of the entropy sources, which are the raw materials for randomness. Entropy quality, its collection method, and the rate at which it's gathered are crucial factors; low-quality entropy or predictable collection methods can weaken the RNG's output. Next, the cryptographic algorithm itself is examined for resistance to known attacks. This includes statistical analysis to detect any biases or patterns that emerge over time, and testing the algorithm against known cryptanalytic techniques like differential cryptanalysis or linear cryptanalysis. The goal here is to ensure that the algorithm doesn’t produce outputs that are disproportionately influenced by certain bits of the seed or state. Moreover, the implementation of the RNG in software or hardware can introduce additional vulnerabilities. This part of the analysis involves code audits to find bugs or oversights that might compromise randomness. Flaws like insufficient mixing of entropy, predictable incrementation of state variables, or improper seeding can be exploited by attackers. Cryptographic analysis also extends to the operational environment of the RNG, checking for side-channel vulnerabilities like timing attacks or power analysis, where an attacker could infer internal states by observing indirect characteristics of the system. Verifiably Random and Tamper-Proof RNGs The process begins when the Randcast Adapter Contract receives a randomness request from a DApp or Web3 game. It then initiates a BLS-TSS (Boneh-Lynn-Shacham Threshold Signature Scheme) task, sending an on-chain event to the ARPA Network, which consists of multiple nodes capable of performing these tasks. The network acknowledges this event and completes the task, subsequently returning the signature as the random seed to the Randcast Adapter Contract. This seed is then transformed into the desired randomness type—be it a dice roll, shuffled array, or any other form—and is utilized in the subsequent DApp or Web3 game logic. How Randcast is different Randcast's approach is particularly vital in scenarios where reliable randomness is essential. In the deterministic environment of blockchain, where smart contract outcomes hinge on input, generating a random number can be easily manipulated. For example, using a block hash or timestamp as a randomness source could be exploited by miners for their benefit. Randcast circumvents this issue by generating random numbers through a decentralized network, ensuring transparency and verifiability in the process. A crucial aspect of Randcast's design is its resilience against tampering. Traditional single-entity randomness generation poses the risk of manipulation. Randcast circumvents this problem by utilizing the collective capabilities of the ARPA Network. Through BLS Threshold signature tasks, it ensures that no individual node can influence the final randomness outcome. This decentralized approach not only enhances security but also upholds the integrity and impartiality of the randomness generated, making Randcast an indispensable tool in the Web3 developer’s arsenal. Reference links https://www.roulettephysics.com/how-to-beat-rng-roulette-slot-machines-other-casino-games/ https://research.nccgroup.com/2021/10/15/cracking-random-number-generators-using-machine-learning-part-1-xorshift128/ https://lemire.me/blog/2017/08/22/cracking-random-number-generators-xoroshiro128/ https://crypto.stackexchange.com/questions/24646/can-machine-learning-analyze-random-number-generator https://www.reddit.com/r/askscience/comments/56r419/is_it_possible_to_find_the_algorithm_for_a_random/?rdt=57618 https://lcamtuf.coredump.cx/oldtcp/tcpseq/print.html#tcpseq

Securing the Unpredictable: Random Number Technology In Web3 Games

অপ্রত্যাশিত সুরক্ষিত করা: ওয়েব 3 গেমগুলিতে র্যান্ডম নম্বর প্রযুক্তি

Das Unvorhersehbare sichern: Zufallszahlentechnologie in Web3-Spielen

Asegurar lo impredecible: tecnología de números aleatorios en juegos Web3

Sécuriser l'imprévisible : la technologie des nombres aléatoires dans les jeux Web3

अप्रत्याशित को सुरक्षित करना: वेब3 गेम्स में यादृच्छिक संख्या प्रौद्योगिकी

予測不可能な状況を保護する: Web3 ゲームの乱数テクノロジー

예측 불가능한 보안: Web3 게임의 난수 기술

Protegendo o imprevisível: tecnologia de números aleatórios em jogos Web3

Защита непредсказуемого: технология случайных чисел в играх Web3

Tahmin Edilemeyenin Güvenliğini Sağlama: Web3 Oyunlarında Rastgele Sayı Teknolojisi

Đảm bảo điều không thể đoán trước: Công nghệ số ngẫu nhiên trong trò chơi Web3

确保不可预测的情况：Web3 游戏中的随机数技术

Quantum Computers are the most expensive technology on Earth. I suffice, that their success is a direct measure of how rich a technological civilization is. With the riches measured in usable, carbon-neutralized energy. (If we don’t count the carbon now we shall still have to count it later). Moore’s law will not work for quantum computers This idea came to me after watching a clip from with the quantum physicist, applied mathematician and entrepreneur - Guillaume Verdon. Lex Fridman’s podcast Classical computers were first invented in ancient times before Christ, to be honest. But for practical purposes they truly emerged on the global scene with the invention of the transistor in 1947 at Bell Labs. With the invention of the transistor emerged Moore’s law. “ ” – Wikipedia. This law was posited by Gordon Moore in 1965 and has been accurate since 1975 to date. Moore's law is the observation that the number of transistors in an integrated circuit (IC) doubles about every two years It is a law indeed. But for classical computers. Why? Just like artificial intelligence, a classical computer with a human user is a system trying to encode key details of entropic processes in the universe with as small an entropic model as possible. To compute is to represent states in the universe (be they naturally occurring or occurring in abstract mathematical worlds or wherever) and run one of their process changes in a simulation. The less energy we can use to do this, the better. Smaller transistors use less energy and thankfully, Moore’s law has seen these basic units of computation, transistors, shrink tremendously from the light-bulb-sized setup below. To this highly magnified image of IBM’s 2-nanometer chip technology. It looks like an X-ray of some strange animal’s dental formula. Each transistor in the above image is roughly the size of 5 atoms and the 50 billion transistors in the chip can fit on a fingernail. All this miniaturization has been possible because we did not need a lot of energy or material to represent a bit of information, at the fundamental level. We did not know it 49 years ago when Gordon Moore postulated his law. We do now. However, we need a lot of energy and materials to represent a logical qubit of information. And that is precisely why quantum computers will not scale according to Moore’s law. As the number of qubits we need increases, the energy needed hence the cost of running these quantum computations will likely increase to match the growth of real, carbon-neutralized GDP for the entire planet. But qubits are cold. What do you mean they need a lot of energy? Theoretically, a quantum computation consumes less energy than a classical computation. Because the energy changes are reversible, we could, theoretically, run quantum computations on zero energy! Because theoretically still, any energy we input is never output. For example, the fundamental logical gate in classical computations, called a NAND gate, looks as below. You can see that 2 bits of information, A and B, go in, but only one bit of information goes out. Since that information is energy, this shows that classical computers are wasteful of energy. Landauer showed The quantum version of the NAND gate is called the Tofolli gate and is seen below. 3 inputs of energy to give 3 outputs of energy. No energy wasted. But wait, you need 3 qubit inputs while in the classical case, you only needed 2 bits. Already, we can see we need more than is usual. Simulating nature is materially and energetically very expensive Needing more energy pulses for our inputs is not even a scratch on the surface of our energy budget when using quantum computers. Thing is, we are simulating nature using artificial nature, millions of times bigger than what we are simulating. Take superconducting qubits. While natural qubits like “spinning” electrons are extremely tiny things small enough to survive their quantum states from being hit by most cosmic rays, superconducting qubits behave like artificial spinning electrons we can easily control. The downside is that millions of cosmic rays see them very clearly. And they hit them. That is just one form of external noise. It costs us and it is going to keep costing us. We need and shall keep needing a lot of energy to keep energetic noise like cosmic rays out of our quantum systems. As Mr. Guillaume explains, the trick up our sleeve is to build error-correcting codes of codes of codes of codes of the noisy system, which eventually gives us a noise-less system (in the average case). But this makes the system bigger hence a bigger budget for the cool down. Heat is the second form of external noise. Should have been the first one we talk about. Taking the analogy of the spinning electron again, sitted comfortably in its orbit, infrared signals carrying heat rarely bump the electron enough to interrupt its spin. The electron is electrostatically bound to the nucleus and if that electron is sharing the orbit with another electron, the two are essentially entangled and again, harder to bother with a spin-out of their stable configurations. Thirdly, why we suffer so much in energetic costs is because it takes a very huge amount of energy to keep track of quantum states with our measurements. Nature doesn’t do measurements, we do. We are thus adding information (a measurement) to our simulation of nature. Hence more energy (remember Landauer). Measurement also introduces unnecessary energy into a quantum system which affects the position and momentum of our particles a la Heisenberg’s uncertainty principle. It also doesn’t help that collapsing the wave function of a superposed quantum state into 1 of many states, albeit randomly, gives us exactly zero information about the hidden variables of that state. This is true randomness, not what classical computers mimic which is pseudo randomness. Our classical randomization algorithms are all pseudorandom. Einstein . complained about this Therefore we usually have to make multiple repeated measurements to figure out the probability distribution of our quantum states, be it to learn how to nudge them or mitigate errors. These repeated initializations plus measurements are called “shots” and we normally need them in the thousands. They all consume a lot of energy. Hence money. We need a bigger budget Here’s a law for you, “If we check every 2 years for the next 2 decades, qubits will keep being more expensive to run for computation than transistors” The funding is going exponential but the best we can show for it so far is . However, we cannot afford to slow down, we are close to a great use-case. 433 qubits For now, progress in quantum computers shall keep growing to match increasing funding to quantum computation but with zero profits, and it will not break even in the same way classical computing has broken even. You know, with some sort of Moore’s law for qubits. Certainly, there are learning curves that take us from poor qubits to better qubits and this curve reduces energy costs hence the budgetary constraints. But at the lowest limit, running perfect simulations of natural models is going to cost us big time and will never be cheaper than classical computation en masse. Classical computers might be approximations of nature, but they show us that we can go very far with approximating nature. However, since we need to sometimes take a critical look at reality if only to break out of approximations that are hitting their limit, then we need to be able to pay for it. Big quantum computational systems like LIGO cost $1.1 billion but it has to be paid if you want to look at the clear night sky and see gravitational waves. No classical computational systems, no matter how cleverly built, will be able to encode everything that is going on in LIGO. The E/acc movement is right, humanity needs to reach for more rungs up the ladder of the Kardashev scale. We need to be able to consume hundreds of times more usable, carbon-free energy than we use today. This energy is what truly pays for everything including quantum computations and while the technology may not be good enough for making profitable desktop-quantum-PCs, quantum computers will, once every while, give us a view of reality we did not expect. A view that re-ignites our interest in the world. Quantum Computers help nudge humanity and its classical computational approximators ever closer to truer greatness than if we had never discovered the power that is simulating nature with a quantum computational model. Even though it might be the most expensive technology on earth. *** P.S. >> Because running quantum computations is so expensive our fears of futuristic quantum computers hacking our email are unfounded. Social hacking will remain a cheaper way to hack your email, quantum computers, not so much. And if your password is PASSWORD please do not blame quantum computers.

The Most Expensive Technology on Earth 

পৃথিবীর সবচেয়ে ব্যয়বহুল প্রযুক্তি

Die teuerste Technologie der Welt

La tecnología más cara del planeta

La technologie la plus chère au monde

पृथ्वी पर सबसे महंगी तकनीक

地球上で最も高価なテクノロジー

지구상에서 가장 비싼 기술

A tecnologia mais cara do planeta

Самая дорогая технология на Земле

Dünyanın En Pahalı Teknolojisi

Công nghệ đắt nhất trên trái đất

地球上最昂贵的技术

A few months ago, the Yelp engineering team published an article about how they use embeddings to represent and organize online content, particularly within the context of Yelp’s efforts to offer easily accessible high-quality content. https://engineeringblog.yelp.com/2023/04/yelp-content-as-embeddings.html?source=post_page-----2c1d62f74e10--------------------------------&embedable=true Yelp aims to offer easily accessible high-quality content. We need to tag, organize and rank online content to attain this goal. For this purpose, Yelp engineers have started using general embeddings on different data. It improves usability and efficiency for all kinds of model development. How refreshing is it not to have one more article about embeddings in the context of RAG (Retrieval-Augmented Generation) isn’t it!? That was the hook for me! Additionally, having worked many times on enterprise search systems (intricate implementation details but also ranking/relevancy aspects), I was also looking at use cases that were not search-oriented, or at least not directly. Although it can be interesting to explore embeddings in the context of semantic search (vs full-text search) or hybrid search, which is a valid topic by itself, that’s not something for this time. :) Without further ado, let’s dive deep into this article! I’m excited to learn about how Yelp supports its various use cases related to tagging, organizing, and ranking. Embeddings are the common backbone for many ML models and facilitate cross-team collaborations. I will use this article to introduce some machine learning concepts along the way that can help to better understand the domain. But first, this is all about embeddings! Text embedding has been researched in depth in the scientific community. First, embeddings were generated with sparse vectors representing words. Spoiler alert: the article not only explores but also delves into photo embeddings (yep, Yelp has a lot of different content…)! text embeddings At its core, one of the article’s focuses is embeddings, which are vectors utilized to represent diverse types of content. Named appropriately, these vectors are ‘embedded’ into a designated space, serving as a numerical representation for the encoded content. This space is a coordinate system where each vector has a specific position. The arrangement of vectors in this space captures the relationships and context between them. In the context of embeddings, each vector in the space has multiple dimensions. Each dimension represents a specific feature or characteristic of the content. Embeddings developed further with context-aware embeddings since the same word can have different meanings depending on how it is used in a sentence. With the use of transformers in recent years, we now have text snippet embeddings that capture more semantic meaning. Consider a simplified text embedding with two dimensions: one representing the of terms and the other for terms. A vector [4, 1] might represent a text heavily focused on technology. Embeddings can encode entire text passages, not just single words, and can have numerous dimensions tailored for specific use cases (e.g., similarity, sentiment analysis). This is a simplified representation of . The model used by Yelp is more advanced than this example. A good starting point to learn about is to explore Word2Vec. frequency technology-related health-related text embeddings word embeddings Whether vectors represent words, photos, or text snippets, proximity in space implies meaningful connections. This representation enables advanced operations and use cases. Small thing: sparse vs dense vectors. Sparse vectors offer memory efficiency but may be computationally expensive, while dense vectors provide richer representations at the cost of increased memory requirements. Both are compact representations of the content. To wrap up this refresher on embeddings (though we spent quite some time there), in the realm of multimodal models (models that can handle different types of data, such as both text and images), the direct use of raw data representations is infrequent. Instead, content is effectively encoded as vectors, providing a convenient means for manipulation. Once the content is represented numerically, basic mathematical operations become feasible (the basis for many ML algorithms). Oh, and I almost forgot: how do we create these embeddings? A transformer is a neural network architecture designed for various tasks, including generating nuanced and semantically meaningful embeddings not limited to text alone. It considers broader contexts, whether within a sentence, text snippets, or other data types like images in multimodal settings. Here’s if you want to learn more: the introductory paper So what does Yelp want to do? Semantic comprehension of the text is essential for Yelp. Yelp reviews are our most valuable asset since they contain a lot of business context and sentiment. We want to capture the essence of each review text to serve their information to our users better. We looked for versatility in our embedding as we try to use the same embedding in various tasks: tagging, information extraction, sentiment analysis and ranking. I briefly mentioned it earlier, but here’s the core idea! The goal is to have versatile embeddings that can be applied in different contexts to accomplish various tasks. Once these embeddings are obtained, they can be used as input for different machine learning models, each tailored to specific tasks. To achieve this versatility, a transformer embedding model is employed, labeled input data relevant to . In the end, the space where the embeddings are encoded captures the knowledge necessary for these diverse tasks. trained on the desired tasks Embeddings based on reviews are currently generated by the Universal Sentence Encoder off-the-shelf model offered by Tensorflow. This blog section will present the USE model, any modifications tested to improve it and its advantages for the Yelp dataset. The Universal Sentence Encoder (USE) serves as the base transformer model they’ve chosen. If you’re interested in more details, you can refer to their article or read the paper introducing the Universal Sentence Encoder. Alright, I’ll skip this part. I’m not particularly interested in the inner workings of the USE transformer; it’s a bit too detailed for me. I’m sure it’s impressive, though! ;) You can find the base model on Hugging Face, and it also includes a comprehensive description of its internals. https://huggingface.co/Dimitre/universal-sentence-encoder?source=post_page-----2c1d62f74e10--------------------------------&embedable=true Here’s something helpful for better visualizing these embeddings: earlier, I provided an example of a 2-dimensional vector used to encode whether a text was more focused on technology or health. USE generates, by default, 512-dimensional embeddings. It helps to relativize how they capture enough context to be used for all these various tasks. Then, the article goes on to describe how they used this model to encode reviews and the results. By nature, most NLP models will perform better when trained on domain-specific text. With this hypothesis, we developed and compared a Yelp fine-tuned encoder with the pre-trained USE model available on TensorFlow-Hub. We aimed to create a better model adapted to the Yelp domain than the pre-trained model. After fine-tuning the model, we wanted to use it to generate an embedding for reviews specifically. Some vocabulary clarifications first: the terms ‘off-the-shelf model,’ ‘foundation model,’ and ‘base model’ are often used interchangeably. They typically refer to pre-trained models available for immediate use without the need for extensive customization or additional training. The development of a base model is a resource-intensive process, with a focus on data quantity over quality. The costs associated with this first stage can amount to millions of dollars in resource processing and can span several weeks. The exact details of large model development, including training time and dataset sizes, are proprietary information held by the developers and researchers involved in the project. Access to this information is crucial to identify biases and understand the ethical impact on applications that use them. Open-source models, like USE, have an advantage here. Given the significant costs associated with creating base models, not every company can afford this investment. Consequently, having access to these pre-trained models off the shelf for free proves to be highly convenient. In the second stage, these pre-trained base models can be fine-tuned to better capture business domain knowledge. While the first stage focuses on learning the structure, features, and recurrent patterns from extensive datasets (task and content-specific), the second stage involves training the base model on a smaller, domain, task-specific dataset, prioritizing . quality over quantity However, in Yelp’s case, the fine-tuned model did not outperform the base model during inference. The model evaluation made on the Yelp domain showed that the ready-to-use model performed better than or as well as the Yelp pre-trained encoder for all tasks. These results happened because either the Yelp domain touches many generic subjects in the USE model or our experiments lacked task diversity to gain an edge. Based on these results, we decided to keep the off-the-shelf USE pre-trained model. Definitely, the quality of the fine-tuning dataset could be a significant reason. Factors like diversity and biases might impede the fine-tuning process. This raises questions about how they assess the quality of their input data, specifically reviews, and whether they undergo preprocessing. Hyperparameters can also play a role in impacting the process. Now, the choice of evaluation metrics is crucial. If the metrics used do not align with the goals of the fine-tuning tasks, they might not accurately reflect the model’s performance. We computed the cosine similarity between different embedding representations of reviews Cosine similarity is commonly used to compare text embeddings. It does not take into account the magnitude of the vectors (i.e., the length of the text). Instead, it focuses on the orientation of the vectors rather than their absolute values. It’s easy to understand that the semantic meaning of a sentence should not be impacted by the length, but if the vectors point in opposite directions, they probably express opposite meanings. Note also that, as mentioned earlier, USE generates fixed-length vectors. In this case, it would not have been an issue with cosine similarity, but fixed-length vectors are easier to integrate with different ML algorithms (and this is what they want, remember: tagging, sentiment analysis, etc.). … It transforms varying sentence lengths into a fixed-length vector representation. But ultimately, what I’m really missing in this testing approach, beyond the metric that was used, is the process. I would have liked to know more about their testing approach, including details about the data they used, whether it is automated, some insights into the infrastructure, etc. Now that they have explained how they generate text embeddings from user reviews of businesses. They mention that they use these initial embeddings to generate another type of embeddings: . business embeddings We chose to base our business embedding on user content. We select the 50 most recent reviews and average their vector embeddings to create our first business embedding representation. It’s a great way to start since reviews contain quality content describing the businesses. The next step will be to add the photo embeddings as well. So basically, they average the top 50 reviews’ embeddings of a given business to generate a single embedding per business. Something like that? business_embedding = sum(top_50_embeddings) / len(top_50_embeddings) After all, that’s the beauty of having vectors… However, they don’t explain how they pick out those top 50 embeddings. Perhaps they include the review rating as part of the text embedding and use it down the road for ranking. A clearer explanation of the approach regarding additional metadata handling would have been nice. And then there’s this assumption that “reviews contain quality content.” Really? I mean, as a user, when I drop a note on a business page, it’s not always a masterpiece… So, I’m still hung up on my earlier question about sizing up the quality of their input data. And why do they do that…? Business embeddings help generate a top-k similarity list to relate businesses to other businesses, users to businesses and users to users based on their matching business interaction history. This correlation matrix of similarity helps show signification recommendations like “Users like you also liked…” or “Since you like business A, you might like business B”. You can learn more about this use case in . this blog post Ah, that’s cool — recommendations based on vectors. That’s a use case that makes sense (comparing related vectors) and adds a lot of end-user value. Well done! The approach they mention is based on a technique named , either item-based or user-based. In their case, the correlation matrix mentioned in the context of generating recommendations indicates that a similarity score, possibly Pearson correlation, is computed for each pair of businesses or users. collaborative filtering They refer to another post that provides a detailed explanation of their approach. It reminds me of an amazing book, old but incredibly well-written, that actually sparked an obsession with recommendations for me at that time — by the late . But there’s so much more than recommendations in this book! Toby Segaran https://www.oreilly.com/library/view/programming-collective-intelligence/9780596529321/?source=post_page-----2c1d62f74e10--------------------------------&embedable=true Ok, we are almost at the end, 2/3 done… They now introduce another type of embedding: . Well, this one is not for the faint of heart, and honestly, I will go through it quickly because if you want nitty-gritty details, head to the article. I think I have more newbie questions than comments. photo embeddings This time, they chose to use a different model than USE to generate these embeddings: CLIP. CLIP (Contrastive Language-Image Pre-training) and USE are tailored for different purposes, with CLIP being more adept at handling image-text associations and abstract visual concepts. Here are some of the advantages highlighted: The CLIP model is a zero-shot model, meaning it can infer successfully from an unseen dataset. A zero-shot model is an opportunity for Yelp to better identify and tag unseen photo categories to improve photo search. Our classifier won’t need a thousand examples for each new tag or label added. The next section compares the CLIP model with their existing classifiers, ResNet50, for the classification of Food, Drinks, Menu, Interior, Exterior, and Home Services Contractor. Without any fine-tuning, the pre-trained CLIP model outperforms ResNet50, as assessed by precision/recall measures. The main advantages of CLIP include its ability: To recognize categories it has never seen during training (zero-shot learning), while ResNet50 requires examples for each category. To capture semantic relationships and concepts within images and text, unlike ResNet50, which lacks an inherent understanding of textual information associated with the images. This is a really interesting finding: just like USE, the base CLIP model outperforms a fine-tuned model (note: ResNet50 is not a transformer architecture) . for Yelp It would have been interesting to know how a fine-tuned CLIP using Yelp data would have fared. Additionally, I would have liked more information about the testing methodology. Finally, they look at some of CLIP’s vulnerabilities. Before using CLIP and publishing its results, it’s better to know its vulnerabilities and how to optimize the model performance. The key vulnerability highlighted in the text is related to what the authors refer to as ‘typographic attacks’ — the model’s susceptibility to being misled by the presence of specific words or text in images. https://gizmodo.com/if-skynet-takes-over-try-writing-robot-on-your-shirt-1846433086?source=post_page-----2c1d62f74e10--------------------------------&embedable=true In summary, the vulnerability stems from CLIP’s sensitivity to textual elements in images, posing a risk of misclassification when certain words are strategically placed. That’s something to know when deploying CLIP in real-world scenarios, particularly in contexts where typographic attacks could be a concern. I understand that USE could not handle the joint processing of images and text, leading them to explore CLIP. However, the embeddings generated by the two models exist in different spaces, preventing cross-content operations by mixing these two types of embeddings. I wonder if such a requirement could arise. Additionally, I read about Hugging Face providing an encoder-transformer template that can be used to build your own transformer architecture by combining several base models if I understand correctly. https://huggingface.co/blog/warm-starting-encoder-decoder?source=post_page-----2c1d62f74e10--------------------------------&embedable=true Could they use this if needed? This is clearly beyond my expertise. In conclusion, that was a very interesting article! It showcased: the usage of embeddings to power different use cases, the usage of embeddings as an interface between products and teams, the efficacy of off-the-shelf models compared to models that have been finely tuned (for Yelp), and the advantage of having these base models available for free and, even better, open source (either known training set and/or parameters/weights). Still, some open questions persist, such as: validation of the quality of the input data (e.g., reviews), details about the testing methodologies, The storage infrastructure is missing: where did they decide to store and serve these embeddings? It could be a basic data lake (blob storage) or a more advanced vector database; however, the required search capabilities are not really advanced at this point. Unclear also about the approach to managing metadata, which is also related to the previous point. I would have loved to learn more about the sentiment analysis use case that was mentioned at the beginning. Investigate and delve deeper into the usage of embeddings in the context of graph representation, as embeddings effectively capture relationships between concepts. I  hope this post gives you some insights into what you could do with embeddings, How you could generate them, and what parts of your products could benefit from them. These off-the-shelf models can help you easily incorporate this technology to make it work for you and deliver more end-user value. The sky is the limit, and you don’t actually need to know all the intricacies. Found this article useful? Follow me on Linkedin and ! Please 👏 this article to share it! Hackernoon

BITS TECHNOLOGY

bits.bi

1-10 employees

Since 2022

COMPANY RANKING

BITS TECHNOLOGY

EVERGREEN INDEX #2099

Representing this company? Click here to claim and customize this page!

Bits Technology WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!

BITS TECHNOLOGY

bits.bi

1-10 employees

Since 2022

COMPANY RANKING

BITS TECHNOLOGY

EVERGREEN INDEX #2099

RELATED COMPANIESRANK

STORIES ABOUT BITS TECHNOLOGY

Representing this company? Click here to claim and customize this page!

Bits Technology WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!