Hopstack, headquartered in Palo Alto, CA, is the provider of an advanced AI-powered warehouse management and execution SaaS platform to automate and digitize warehouse operations

Hopstack

hopstack

Hello! We are developers from the Navigine company. We decided to test the Wi-Fi RTT technology and share the results with you. What is Wi-Fi RTT Wi-Fi RTT (Round-trip time) is a brand new technology that allows locating devices with unprecedented sub-meter accuracy. Basically, it measures the distance between a device and Wi-Fi routers which is way more accurate than using a GPS signal. The latter can’t provide an acceptable level of accuracy at all when it comes to indoor positioning. Wi-Fi RTT outperforms RSSI-based methods as well, thanks to reduced latency and better accuracy. The Use Today, this technology is the next step to improve services that use indoor positioning. We expect that its implementation will become a better solution for many industries. Wi-Fi RTT is applicable for tracking vehicles and cargo within warehouses and production plants. Comparing to BLE, this technology provides enhanced accuracy and reduced latency, making processes more rapid and reducing delays. It is effective for public hotspots as well, such as offices and malls, airports, exhibitions and not only. Indoor navigation is the solution users need for making their visits to public places more comfortable and enjoyable. Through this, they can plan their time effectively and find locations they need immediately. The Advantages Our solution goes further with unique algorithms that enhance the basic capabilities of Wi-Fi RTT. demonstrates a submeter accuracy in checkpoints 95% of the time and around 30 cm accuracy 50% of the time. The positioning engine is a particle filter combining Pedestrian Dead Reckoning (PDR), based on inertial sensors readings, and Wi-Fi RTT readings for correction. Meanwhile, the dynamic test shows that delay in position estimation does not exceed one second. The latter result is achieved by complementing Wi-Fi RTT measurements with PDR via particle filter. Semi-static test in our office The technology is not just an innovation in itself, but also a practical solution that improves business efficiency and customer experience. Wi-Fi RTT provides longer distance measurements and can be integrated into existing Wi-Fi network infrastructure. Current status Now, it is too early to refer to any cases, while we are currently working on raising experts interest from LinkedIn. We can see the real need for this technology, so the first implementations are coming soon in Q1 2020. Today, we have confirmed orders from a big carrier and one logistics company. Why Wi-Fi RTT? The industries using indoor positioning have to realize that often BLE is not the solution that meets their needs. Wi-Fi RTT is a significantly more relevant technology for many cases. Our solution based on Wi-Fi RTT delivers clear advantages for customers, including high accuracy, low latency, ease of use and integration. With just 2 weeks between order and integration, it is one of the most rapid and flexible solutions. The combination of high accuracy, availability, and low price, can make the technology an Indoor positioning market leader.

A Needle in a Haystack: How Wi-Fi RTT Takes Indoor Positioning to the Next Level

Finding enterprise use cases worth development dollars Earlier this year, the GSMA, the world’s most influential telecom consortium, published a paper detailing Some of them were compelling. Others, less so. fifteen blockchain use cases for mobile network operators . As an , I was happy to see this kind of attention given to my still-nascent field. (Is all press good press?) But as a founder whose ability to buy Chipotle burritos rests on how seriously and successfully telecom carriers adopt blockchain technology, the addition of those less compelling use cases had me worried. entrepreneur in the telecom and blockchain space The sad truth is that the blockchain world is rife with misinformation, poorly-thought-out applications, and centralized tech pretending it’s “de-”. If left to their own devices, I could imagine carriers attempting the use cases I saw as less promising and being forever turned off of blockchain. So, although it’s against my very nature as a native Midwesterner, I see no other alternative to raining on this particular parade. As Warren Buffett wisely said in his : '17 shareholder letter The less the prudence with which others conduct their affairs, the greater the prudence with which we must conduct our own. In this article, I’ll attempt to be a voice of prudence. The applications that the GSMA proposes range from great and world-changing to fundamentally flawed, and I’ll explore four of them in the sections that follow. In so doing, I’ll draw out lessons worth remembering not only in telecom but also in broader enterprise blockchain R&D. A quick note on methodology: as an IDEO CoLab Fellow last winter, I learned that new tech is most interesting when it is desirable (i.e., adds real value for real people), feasible (i.e., technologically possible), and viable (i.e., good business). Check out IDEO’s explanation here if you’re curious, but it’s easiest to just witness them in action. You’ll get it. #1: Smart contracts for 5G network management “To realise the 5G promise of ubiquitous access across various networks, CSPs will need to handle heterogeneous access nodes and diverse access mechanisms. Selecting the fastest access node for every user or machine will be a central challenge in the future. …Rules and agreements between the various access providing networks can be coded into autonomous smart contracts. These contracts can be dynamic in nature wherein any time a policy needs to be changed, only the contract code needs to be changed.” The GSMA’s first use case falls prey to a classic problem of outside-in entrepreneurship — it describes a solution that every consumer wants but no business will offer. Is it to connect to any network freely and get the best value for the best service? Of course! I don’t care who gives me four bars of LTE service as long as I can reliably stream the #7-ranked Michigan basketball game in the back of my Uber. desirable But from the carrier’s point of view, this solution is neither feasible nor viable. On the technological side, the use case makes a number of critical leaps into still-unknown areas of blockchain tech, but far more damning is the business critique: why would carriers ever give up control of their customers? Carriers have built empires by leveraging their network assets to attract customers and reduce churn. Would they really buy into a protocol that forces them to auction for every data session, text, and call? feasibility viability Sounds tautological, but it’s surprisingly easy to forget. Yes, real-time bidding for connectivity would be amazing. But if it ever exists, it will only be after a prolonged battle with operators the world over. The lesson here is for prospective entrepreneurs: keep the enterprise in mind when designing enterprise blockchain products. #2: IOT connectivity “While localised networks in for instance a manufacturing line can provide great benefits, self-organised global supply chains powered by IOT can help streamline processes that require a lot of paperwork and governance. This would shift the focus from centralised IOT platforms (in technology and governance) to self-governed mesh networks… IOT can also be used for automatic tracking for quality and conditional parameters during transports, switching of ownership and provenance use cases.” The GSMA offers up three use cases in the IOT space. Unfortunately, they all fall prey to the same problem — they’re smart, and should probably happen, but they probably shouldn’t happen on a blockchain. These use cases are : IOT management via blockchain is a use case well-explored by startups and large enterprises alike. They may even prove to be : IOT is a huge and growing industry, devices will get smarter and smaller over time, and managing devices efficiently is a huge, valid concern. feasible viable The problem arises, though, when we consider the of blockchain in this context given the problems we are asking it to solve. desirability As I understand it, the problem is that we can’t get good data about products flowing through a manufacturing line or a multinational supply chain, so we’re asking blockchain to help. Walmart, for instance, now mandates that farms that supply their stores leafy green vegetables enter “ ” that they’ve build in conjunction with IBM. detailed information about their food into a blockchain database But wouldn’t that “detailed information” be just as useful in a normal database? In my mind, this type of supply chain management is a poor use of blockchain. If you can’t trust your supply chain partners (or much less, the folks on your manufacturing line) to produce accurate data, it doesn’t matter whether you use a blockchain, a database, or a stone tablet to store it: your customers are still going to . Garbage in, garbage out. (Or, due to blockchain’s immutability, it’s more like garbage in, garbage forever.) get E. coli from your lettuce Blockchain is expensive, slow, and non-private. It costs significant time and resources to write to a blockchain. Barriers to entry are high. (Do Walmart’s farmers now have to own wallets or accounts with IBM?) The lesson here is never to use blockchain when a normal database will do. This is not to say that blockchain won’t have a place in IOT; I think it will. But I’m willing to bet that device security turns out to be a better use case than provisioning or self-maintaining supply chains, because the former is a better fit to the unique constraints and benefits of blockchain. #3: Fraud prevention The telecom industry has not yet found a way to effectively and sustainably prevent fraud. In a general sense, working together to combat interoperator fraud is a good way to start. Blockchain could for instance be used to govern access to fraud detection information shared between operators…another angle is fraud caused by delays in sharing [call detail records] between operators. As blockchain works in real-time, using a shared administration of network activity could help identity plausible fraudulent connections faster. Smart contracts could be deployed to actively enforce network usage compliance rules and help us resolve inter operator disputes easier. Surprise twist: I actually this upcoming use case! In fact, I like it so much that I’ve been working on it for the past year with . (Is that cheating?) like drect.ly From a standpoint, blockchain has some clear advantages over other solutions in fraud prevention. In fact, it’s so good that you get to (justifiably) recite the holy text of blockchain use cases: desirability This is a case when mutually-distrustful actors need to coordinate to maintain a shared version of the truth for an economically-valuable digital record. That sentence is why blockchain was created; in this use case, we’ve just replaced the economically-valuable digital record of bitcoin ownership to a record of which telecoms have provided how much coverage to which roaming customers. Those records, it turns out, are incredibly valuable (and therefore, if managed well, can sustain a business model). Every year, telecom carriers lose over $38 billion to fraud; even a relatively small dent in that number is large in absolute terms. viable My own incredible bias notwithstanding, I do recognize some challenges here with . Billions of records are exchanged daily, which poses scalability problems. These records expose significant, closely-held data about network usage and subscriber behavior, which poses privacy concerns. Governance is challenging, for it’s unclear how centralized my ownership of the code base makes this project. And how can you get bad actors to sign onto a platform that demands higher responsibility? feasibility All that said, these are solvable problems. And perhaps that’s another lesson: ensure that your use case is a great fundamental fit with the unique value prop of blockchain technology — and then prove the feasibility haters wrong. If you think about it, that’s really the only way that technology advances. any #4: Hosted services Distributed Ledger and/or Blockchain network operators are responsible for configuring the network and granting access to network participants. In general, they specify the use case(s) that the network is serving, and are often involved in managing the network in terms of software upgrades, arbitration services, etc. A network can be operated by a single entity or a federation of multiple, separate entities. As a strategy consultant at Deloitte, I helped write a . It generated over 100 million views before being lost to the sands of business-thought-pieces past, but I’ll always remember it as the first time that I learned John Hagel’s concept of innovation “ .” trends report on ecosystems thinking antibodies The SparkNotes version of Hagel’s idea is that antibodies (i.e., agents of resistance) arise in big organizations whenever change is coming. They fight for the preservation of the status quo that helped them rise to power, and, even if well-intentioned, they often stifle the speed of change by insisting that old rules, norms, and values govern fundamentally new contexts. For example, a new business unit might be shut down because it’s cannibalizing sales of the core product; that is judged by the old rule of “protect the core,” rather than the new reality that any player could release the cannibalizing product and it’s better to optimize for growth. In the blockchain context, antibodies take the form of centralized services masquerading as blockchain technology. In the old world, businesses want efficiency and control — but Proof of Work is deliberately inefficient, and blockchain networks are intentionally decentralized. When I read a use case like the “hosted services” from above, where blockchains are treated like permissioned databases, I can’t help but remember Hagel’s antibodies. The whole point of blockchain technology is to make public commitments that are impossible to reverse; to open up previously siloed networks; and to radically decentralize power, code, and economic value accrual. So, my final lesson is this: You’ll want to do so — believe me, as someone who has built these systems — but if you do, you’ll be further from the real promise of blockchain technology and closer to designing a really bad database. After living in a centralized world, decentralization feels innately wrong. But as we enter this new age of decentralized technology, we ought to remember that it’s okay to break all of the old rules. never surrender decentralization when designing blockchain use cases. In fact, that’s the whole point. To connect with Christian, reach out on Twitter ( @cdkeil ) or via email at christian [at] drect.ly.

Needles in the Blockchain Haystack

How do you find the job candidate who’s an exact fit for your organization? How do find the person you are going to marry? How do you locate the right house for you and your family? How about a pet that’s been missing? Worse yet, a child who’s been missing? As a society, we spend a lot of time looking for things. Individuals search for apartments to rent, cars to buy, and romantic partners to date. Businesses search for customers, employees, and suppliers. . Our goal is to enable people to use incentive programs and in order to solve some important problems facing the world today. We believe, in particular, that we can help people Although we tend to approach these searches as if they are unique, we at nCent believe that they appear distinct on the surface but are often subject to similar dynamics on a meta level blockchain technology find higher quality things in a fraction of the time and more efficiently than they might expect. Some search problems, like finding the closest Italian restaurant, are easy to solve. , for example, is more challenging because it may be unclear where to start and how to efficiently gather information. It’s obviously a high stakes situation, because it involves the welfare of a child — quite a bit different then satisfying a pasta hankering. But other search problems are more difficult and involve higher stakes. Finding a missing child Difficult searches are called problems. “needle in haystack” One of our guiding principles at nCent is our belief that these “haystack” problems can be optimally solved by combining decentralized networks, blockchain technology, and the right incentives programs. A ‘proof of concept’ for this approach happened in when DARPA (USA defense research agency) placed ten red balloons in secret locations around the U.S. and offered $40,000 to the first team to locate them all. The winning team solved this search problem by offering recursive incentives: Anyone who either correctly identified the location of a red balloon or someone whose information led to a person who identified the location of a red balloon was rewarded. For example, if Person A referred Person B who referred Person C who correctly identified the location of a red balloon, then all three people in this chain would get rewarded in some capacity. 2009 referred Even if you don’t know where the needle is in the haystack, you are still incentivized to help by referring others to get closer. If the tail end of your subtree finds the needle, you too will be rewarded. Everyone who contributes to the network can be absolutely sure that if their contribution is valuable, their reward will be commensurate with this value. This means that the person who finds the missing child the person who spread the word to the finder’s social media network will be rewarded. The person who refers a successful candidate for a specialized job the person who informed the referrer will be rewarded. These programs can reduce their overall cost with (Larger payouts for those who are “closer” in the chain to the final goal) Utilizing blockchain technology in this process enables the network to be transparent and to be audited by anyone at any time. and and recursive payout structures . These recursive programs are scalable and when designed correctly, have an upper bound on the cost of sponsoring them. Recursive incentives motivate people to inform other people of the search problem. When executed properly, these networks grow quickly and organically, increasing the statistical probability of finding the target. In fact, we believe that recursive incentives, where every node of the successful chain is rewarded, become increasingly more valuable as the search problem becomes more difficult. Spam is another common problem that we will address in a forthcoming post: The biggest hurdle to solving a haystack problem is growing the search network quickly enough so that the chance of identifying the needle increases exponentially. “Cashout: What Does it Mean to be Spammed with Money?” As the tail end of the network is further removed from the origin of the network, the spread of information will be more efficient than any other search process implemented in the marketplace. . The incentive structures implemented on our blockchain will be modular enough to apply to many types of search problems. The inclusion of “decentralized networks” and “blockchains” might sound like tech jargon, but in reality it’s just the underlying tech that enables you to give “asking your friends for help” a global reach! We fully recognize that each case will come with its own idiosyncrasies, ranging from economic constraints of the searcher to the psychological realities of how people respond to incentives To learn more about our work, email me directly at kk@ncnt.io . To stay in the nCent loop, hear me tweet and join our international telegram channel .

How We Can Find the Needle in a Haystack, Together

Pinpointing Root Causes of Application Performance Problems in Dynamic Environments Pinpointing the root cause of application performance problems in a dynamic environment often resembles a complicated game of hide-and-seek. A familiar analogy is that it’s like trying to find a needle in a haystack. This characterization illustrates the challenge but isn’t especially helpful in offering solutions. So, if the root cause of an application problem is the “needle” and the application environment is the “haystack,” how exactly would we go about restoring peak performance? Taking the analogy literally is actually a pretty good way to explore the various approaches to resolving application performance problems in dynamic, elastic, and transient infrastructure environments. So, exactly how would you go about finding a needle in a haystack? 1. Eliminate the haystack Probably the fastest way to find the needle is to burn the haystack. Or dumping the hay into a pool of water and waiting for the needle to sink to the bottom would also work. If we were really talking about needles and haystacks, fine — but we’re not. There are fundamental business reasons why more and more companies are turning to dynamic cloud-based deployments in the first place. We can pretty much reject the option of returning to monolithic, static infrastructure where identifying application performance cause-and-effect relationships is more straightforward. 2. Rummage for the needle manually We could sift through the pile of hay, strand by strand, until we find the needle. That’s basically the way traditional (APM) tools work. They throw a massive collection of data at us (albeit usually in nicely formatted charts and graphs) and require us to comb through it all to figure what went wrong. Actually finding the needle remains a laborious manual process. Somebody somewhere has to pore over the charts and graphs, make sense of the data, and connect the dots to discover the root cause. Such handwork analysis can take hours — even days. application performance monitoring That’s too long when application problems are leading to disgruntled customers. Of course, things would go faster if we assigned more staff to tackle the problem: the more people sifting through the hay, the sooner we’d find the needle. But convening a war-room assembly to interpret the data is still a labor-intensive, time-consuming effort. Already most CIOs are frustrated that so many man-hours are spent on troubleshooting instead of innovating and optimizing. Furthermore, and perhaps more significant, today’s liquid environments — where millions of dependencies are changing on the fly from moment to moment — are simply too complex and too fast-moving, with too many performance data points, for humans to analyze in a timely manner. It would take more people than any business has at its disposal to detect causal relationships among so much data and identify the component that’s causing the problem. 3. Locate the general area of the needle Using a metal detector or an X-ray would narrow down the approximate location of the needle within a haystack. This is more efficient than looking through the entire pile, but it’s still a lot of work. We still have to manually sift through the hay — just not as much. Similarly, a handful of modern APM solutions get us a step closer. Discovery agents collect and correlate myriad metrics and performance data points from various distributed elements across the application environment. The data is consolidated and presented in a unified, integrated “single pane of glass” view. We now have a clear, comprehensive picture of the of the performance problem, but we haven’t diagnosed the The hard work is still left to us. We still have to pore through charts and graphs of data — just not as much. Modern APM displays the data within a more meaningful framework that helps us narrow down our focus. symptoms root cause. Now, imagine the difficulty when the needle is constantly on the move. That’s exactly what’s happening with today’s highly dynamic environments. Servers are perpetually coming and going, auto-scaling up and down as needs dictate. Container technology and microservices architecture are even more transient — moving from host to host — and generate even more metrics. With our applications now having so many moving parts, trying to keep up with them and their topology by dissecting a plethora of charts and graphs is literally beyond human capabilities. Dynamic applications have millions of dependencies that need to be analyzed, and they’re ever changing. No sooner do we think we have a handle on how everything is working together than the landscape changes and we’re back to square one. Present-day application environments change too fast and too often for even the most talented troubleshooters. Even a relatively simple containerized environment like this one, with only 142 hosts, has more dependencies than a human being can track without software assistance. 4. Extract the needle automatically Ultimately, the best way to find the needle (especially if it won’t sit still) is to use a powerful magnet to pull it right out of the haystack and deliver it into your hands. In our analogy, a next-generation solution would be such a “magnet.” It zeroes in on an application problem’s root cause automatically, rendering time-consuming manual analysis unnecessary — the same way a magnet obviates the need to sift through the pile of hay. artificial intelligence–powered APM Artificial intelligence (AI) enables computers to do what they do best: absorb huge amounts of information and make sense of it faster and more thoroughly than is humanly possible. It applies sophisticated algorithms and context-rich diagnostics to detect and correlate causal relationships among data with millions of dependencies in microseconds. In essence, AI-enhanced monitoring analyzes the problem the same way we humans would, only faster. Artificial intelligence is able to perform almost instantaneously the analytical “heavy lifting” that takes us humans hours or days. From theory to reality: Dynatrace APM Artificial intelligence has been defined as “being able to perceive its environment and apply cognitive functions such as learning and problem-solving.” has been built on artificial intelligence so that it is able to Dynatrace APM auto-discover all components of the full technology stack end to end from the customers’ web browsers all the way down to the host infrastructure — map out the entire IT in an interactive visual display environment identify the millions of dependencies among websites, applications, services, processes, hosts, , and cloud infrastructure networks learn how it all works together and what constitutes normal behavior automatically detect, analyze, and prioritize anomalies and performance problems actually recommend solutions to the root cause of those problems And AI allows Dynatrace to do all this at a speed, scale, and precision that no human could ever match. Dynatrace gets to “know” applications. Its proprietary Smartscape technology continuously auto-discovers and monitors every aspect of every application. Sophisticated algorithms learn normal application performance patterns and proactively flag anomalies. The AI auto-adjusts baselines dynamically in real time to avoid false positives caused by preset static alerting thresholds. Application performance problems are seldom isolated, one-time events, and they’re usually symptoms of a larger issue. Dynatrace looks at all other transactions that used the same components around the same time to see if they also experienced problems. Artificial intelligence correlates events throughout the full technology stack — client side, server side, infrastructure level — to identify analogous issues and detect causal relationships. AI diagnostics analyze the vast amount of data collected to pinpoint the exact component that’s causing the problem. In today’s increasingly complex and fluid IT environments, interpreting data points with millions of dependencies in order to determine underlying casual relationships is more than humans can master by themselves. Only artificial intelligence can take the colossal volume of data and translate it into actionable solutions before problems hit customers. See for yourself what Dynatrace can do. Dynatrace offers a — the complete full-blown version, not a stripped-down demo — that auto-discovers and visually maps out your entire technology stack within minutes. free trial

Finding the Needle in a Haystack

Of late, there has been quite a lot of interest in increasing/improving the context of LLMs. One good example is the infini-attention paper from Google in early April. While it was more of tweaking the model architecture to improve the context, Microsoft came back with an answer in late April but with a data-driven solution. The paper is titled, “Make your LLM Fully Utilize the Context”. There are mainly two contributions from the paper: A data-driven approach that introduces a novel pipeline to synthesize a novel dataset to train LLMs to cleverly use long contexts. It’s dubbed “Information-Intensive Training” (IN2 in short).


A novel evaluation approach named Various Long-Context (VAL Probing) uses 3 context styles — document, code, and structured data and 3 retrieval patterns — forward, backward, and bi-directional. In this article, let's deep-dive into both the simple ideas followed by some evaluations. Visual Explanation If you are someone who likes audio-visual explanations, here is a video version of this article that explains the paper: https://youtu.be/SA21d_4QQSU?feature=shared&embedable=true IN2 Training The first main idea of the paper is a novel data-driven training approach named IN2 training. There are two types of IN2 training proposed in the paper. Fine-grained Information Awareness. In this case, we first take a raw text dataset and split it into 128 tokens and we call each of this set of 128 tokens a “segment.” We then prompt a powerful LLM model like GPT-4 with an instruction I_f in such a way that the answer a_i to this question q_i requires information in the segment s_i to provide the answer. We then combine the segment s_i with several other segments by shuffling it along with the other segments to create the long context L_i. Figure taken from the paper (figure 2). Integration and Reasoning of Information. There can be situations where the context is located in more than one segment (more than 128 tokens). So we consider a set [s_i] instead of a single s_i. This creates a multi-hop Q&A scenario where the LLM has to do at least 2 passes of QAs. We modify the instruction I_f in the prompt to the LLM accordingly. Figure taken from the paper (figure 2) In both scenarios, the authors vary the long context length L_i from 4K to 32K to prevent length bias in training. During IN2 instruction fine-tuning, the long contexts and questions become input to the LLM, and the answer is shown as the output for supervised training. They fine-tune the Mistral-7B-Instruct-v2 model and call the new model FILl-in-the-Middle(FILM-7B). VAL Probing VArious Long-context Probing or VAL Probing was introduced to overcome some of the problems with the current LLMs with retrieving from long contexts. So how do we even test the long context retrieval capability of any given model? The answer lies with “Needle in a haystack” (NIHS) test. It can be accessed here. Needle in a Haystack NIHS was proposed to pressure test LLMs. It’s a simple analysis to test the in-context retrieval ability of long-context LLMs. The test can be run by following the below 3 steps: Place a random fact or statement (the ‘needle’) in the middle of a long context window (the ‘haystack’).


Ask the model to retrieve this statement.


Iterate over various document depths (where the needle is placed) and context lengths to measure performance. The biggest problem with NIHS these days is that the recent models perform too well making the test less appealing. For example, see below the test result of the proposed FILM-7b model NIHS test on the proposed FILM-7b model The results are near perfect indicating that we really need a new testing approach and so, giving birth to VAL Probing. VAL Probing VAL probing considers 3 data types such as documents, structured data, and code. It also considers three retrieval types — forward, backward, and bi-directional. The below figure from the paper gives examples of all three data types and retrieval directions. Results For evaluation, they used the LongBench evaluation scripts and reported ROGUE-L for summarization and F1-scores for all other tasks. The first problem addressed by the paper is the “Lost-in-the-middle” problem. As shown in the below figure, we can see that there is no “U” shaped dip with the FILM-7b model (in orange) compared to Mistral-7B (green), and even other long-context sota models like LongAlign indicating the strength of FILM-7b. On top of addressing the Lost-in-the-middle problem, FILM-7B seems to outperform other recent models quantitatively too. They show the results of averaging the evaluation metric(Avg) and min-max gap of the evaluation metric (Gap) which is the difference between the minimum and maximum values obtained. In short, the model is comparable to GPT-4’s performance! Summary & Conclusion This paper from Microsoft proposes a simple data-driven solution to addressing the long-context problem in LLMs. On top of proposing a data generation pipeline, it also proposes and tests the usage of the novel VAL Probing that encompasses 3 data types and 3 retrieval methods. Let’s hope this is just the dawn of a new generation of LLMs that are no longer limited by the context length. See you in my next… Of late, there has been quite a lot of interest in increasing/improving the context of LLMs. One good example is the infini-attention paper from Google in early April. While it was more of tweaking the model architecture to improve the context, Microsoft came back with an answer in late April but with a data-driven solution. The paper is titled, “ Make your LLM Fully Utilize the Context ”. infini-attention paper infini-attention paper Make your LLM Fully Utilize the Context Make your LLM Fully Utilize the Context Make your LLM Fully Utilize the Context There are mainly two contributions from the paper: A data-driven approach that introduces a novel pipeline to synthesize a novel dataset to train LLMs to cleverly use long contexts. It’s dubbed “Information-Intensive Training” (IN2 in short). A novel evaluation approach named Various Long-Context (VAL Probing) uses 3 context styles — document, code, and structured data and 3 retrieval patterns — forward, backward, and bi-directional. A data-driven approach that introduces a novel pipeline to synthesize a novel dataset to train LLMs to cleverly use long contexts. It’s dubbed “Information-Intensive Training” (IN2 in short). A data-driven approach that introduces a novel pipeline to synthesize a novel dataset to train LLMs to cleverly use long contexts. It’s dubbed “ Information-Intensive Training ” (IN2 in short). Information-Intensive Training A novel evaluation approach named Various Long-Context (VAL Probing) uses 3 context styles — document, code, and structured data and 3 retrieval patterns — forward, backward, and bi-directional. A novel evaluation approach named Various Long-Context ( VAL Probing ) uses 3 context styles — document, code, and structured data and 3 retrieval patterns — forward, backward, and bi-directional. VAL Probing In this article, let's deep-dive into both the simple ideas followed by some evaluations. Visual Explanation If you are someone who likes audio-visual explanations, here is a video version of this article that explains the paper: https://youtu.be/SA21d_4QQSU?feature=shared&embedable=true https://youtu.be/SA21d_4QQSU?feature=shared&embedable=true IN2 Training The first main idea of the paper is a novel data-driven training approach named IN2 training . There are two types of IN2 training proposed in the paper. data-driven training IN2 training Fine-grained Information Awareness. In this case, we first take a raw text dataset and split it into 128 tokens and we call each of this set of 128 tokens a “segment.” We then prompt a powerful LLM model like GPT-4 with an instruction I_f in such a way that the answer a_i to this question q_i requires information in the segment s_i to provide the answer. We then combine the segment s_i with several other segments by shuffling it along with the other segments to create the long context L_i. Fine-grained Information Awareness. In this case, we first take a raw text dataset and split it into 128 tokens and we call each of this set of 128 tokens a “segment.” We then prompt a powerful LLM model like GPT-4 with an instruction I_f in such a way that the answer a_i to this question q_i requires information in the segment s_i to provide the answer. We then combine the segment s_i with several other segments by shuffling it along with the other segments to create the long context L_i . Fine-grained Information Awareness. I_f a_i q_i s_i s_i L_i Figure taken from the paper (figure 2). Integration and Reasoning of Information. There can be situations where the context is located in more than one segment (more than 128 tokens). So we consider a set [s_i] instead of a single s_i. This creates a multi-hop Q&A scenario where the LLM has to do at least 2 passes of QAs. We modify the instruction I_f in the prompt to the LLM accordingly. Integration and Reasoning of Information. There can be situations where the context is located in more than one segment (more than 128 tokens). So we consider a set [ s_i ] instead of a single s_i. This creates a multi-hop Q&A scenario where the LLM has to do at least 2 passes of QAs. We modify the instruction I_f in the prompt to the LLM accordingly. Integration and Reasoning of Information. s_i s_i. I_f Figure taken from the paper (figure 2) In both scenarios, the authors vary the long context length L_i from 4K to 32K to prevent length bias in training. During IN2 instruction fine-tuning, the long contexts and questions become input to the LLM, and the answer is shown as the output for supervised training. They fine-tune the Mistral-7B-Instruct-v2 model and call the new model FILl-in-the-Middle( FILM-7B ). L_i FILM-7B VAL Probing VArious Long-context Probing or VAL Probing was introduced to overcome some of the problems with the current LLMs with retrieving from long contexts. So how do we even test the long context retrieval capability of any given model? The answer lies with “Needle in a haystack” (NIHS) test. It can be accessed here . here here Needle in a Haystack NIHS was proposed to pressure test LLMs. It’s a simple analysis to test the in-context retrieval ability of long-context LLMs. The test can be run by following the below 3 steps: Place a random fact or statement (the ‘needle’) in the middle of a long context window (the ‘haystack’). Ask the model to retrieve this statement. Iterate over various document depths (where the needle is placed) and context lengths to measure performance. Place a random fact or statement (the ‘needle’) in the middle of a long context window (the ‘haystack’). Place a random fact or statement (the ‘needle’) in the middle of a long context window (the ‘haystack’). Ask the model to retrieve this statement. Ask the model to retrieve this statement. Iterate over various document depths (where the needle is placed) and context lengths to measure performance. Iterate over various document depths (where the needle is placed) and context lengths to measure performance. The biggest problem with NIHS these days is that the recent models perform too well making the test less appealing. For example, see below the test result of the proposed FILM-7b model NIHS test on the proposed FILM-7b model The results are near perfect indicating that we really need a new testing approach and so, giving birth to VAL Probing. VAL Probing VAL probing considers 3 data types such as documents, structured data, and code. It also considers three retrieval types — forward, backward, and bi-directional. The below figure from the paper gives examples of all three data types and retrieval directions. Results For evaluation, they used the LongBench evaluation scripts and reported ROGUE-L for summarization and F1-scores for all other tasks. The first problem addressed by the paper is the “Lost-in-the-middle” problem. As shown in the below figure, we can see that there is no “U” shaped dip with the FILM-7b model (in orange) compared to Mistral-7B (green), and even other long-context sota models like LongAlign indicating the strength of FILM-7b. On top of addressing the Lost-in-the-middle problem, FILM-7B seems to outperform other recent models quantitatively too. They show the results of averaging the evaluation metric(Avg) and min-max gap of the evaluation metric (Gap) which is the difference between the minimum and maximum values obtained. In short, the model is comparable to GPT-4’s performance! Summary & Conclusion This paper from Microsoft proposes a simple data-driven solution to addressing the long-context problem in LLMs. On top of proposing a data generation pipeline, it also proposes and tests the usage of the novel VAL Probing that encompasses 3 data types and 3 retrieval methods. Let’s hope this is just the dawn of a new generation of LLMs that are no longer limited by the context length. See you in my next…

How to Make Your LLM Fully Utilize the Context

Using Elasticsearch to Offload Search and Analytics from DynamoDB: Pros and Cons

DynamoDB থেকে সার্চ এবং অ্যানালিটিক্স অফলোড করতে ইলাস্টিকসার্চ ব্যবহার করা: সুবিধা এবং অসুবিধা

Verwenden von Elasticsearch zum Auslagern von Such- und Analysefunktionen aus DynamoDB: Vor- und Nachteile

Uso de Elasticsearch para descargar búsquedas y análisis de DynamoDB: ventajas y desventajas

Utiliser Elasticsearch pour décharger la recherche et l'analyse de DynamoDB : avantages et inconvénients

डायनेमोडीबी से खोज और विश्लेषण को ऑफलोड करने के लिए इलास्टिकसर्च का उपयोग करना: फायदे और नुकसान

Elasticsearch を使用して DynamoDB から検索と分析をオフロードする: 長所と短所

Elasticsearch를 사용하여 DynamoDB에서 검색 및 분석 오프로드: 장점과 단점

Usando Elasticsearch para descarregar pesquisas e análises do DynamoDB: prós e contras

Использование Elasticsearch для разгрузки поиска и аналитики из DynamoDB: плюсы и минусы

Arama ve Analizi DynamoDB'den Aktarmak için Elasticsearch'ü Kullanma: Artıları ve Eksileri

Sử dụng Elaticsearch để giảm tải Tìm kiếm và Phân tích từ DynamoDB: Ưu và nhược điểm

使用 Elasticsearch 卸载 DynamoDB 的搜索和分析：优点和缺点

Using the Stratification Method for the Experiment Analysis

Recently, Apple released the Vision Pro, and I was fortunate enough to become a proud owner. Since then, I've been deeply engaged in developing my application and exploring various apps and trends within the Vision Pro App Store. One thing that grabbed my attention is that new apps are coming out every day, but only a few fit well with the immersive virtual reality experience. I guess many developers struggle to find documentation and examples for Vision Pro features. However, utilizing the capabilities of Vision Pro can set your application apart and provide users with an enhanced experience. In this article, I'll be sharing some tips and tricks I've uncovered while crafting my application. These tips will not only help your application to stand out but also ensure its seamless integration and native feel within VisionOS. Dark Surroundings Effect I'm sure you've noticed that when viewing a video or photo, the background behind the window is darkened, in the settings it's called "auto-dimming". To get the same effect you need to use .preferredSurroundingsEffect(.systemDark) on your view and then the space around the user will darken and your window will remain bright, this will help to keep the focus on the content. Hidden Window Controls Line When viewing a video or content, you want the user to have no distractions and create the effect of a window floating in the air, for this Apple has given us the ability to hide the bottom bar of the window that is used to move or close the current window. With .persistentSystemOverlays(.hidden) you can hide this control bar on your window, it will disappear with the default animation and only appear when the user interacts with the window. Video Reflection Effect If you open the standard Apple TV application and turn on a movie, you will notice that the movie window has a reflection that changes with each frame, giving a more immersive effect. To get a glow effect around your video, you need to use VideoPlayerComponent and enable isPassthroughTintingEnabled, then the effect will be the same as in Apple TV or HBO Max. Here is what the documentation says: This is to enable passthrough tinting during video playback that shows up around the video taking the average color of the frame and tinting the passthrough with that color to emphasize the video. Window Resizing You may want to prevent the user from resizing the window, there is a way to do that: onAppear {
    guard let windowScene = UIApplication.shared.connectedScenes.first as?UIWindowScene else { return }
    windowScene.requestGeometryUpdate(.Vision(resizingRestrictions: UIWindowScene.ResizingRestrictions.none))
} You can also specify additional size settings in this method The full signature of the method looks like this: windowScene.requestGeometryUpdate(.Vision(size:, minimumSize:, maximumSize:, resizingRestrictions:)) Hands Visibility If your application uses immersive space and you don’t want the user to see their hands, or you want to replace them with virtual hands, as is done in AmazeVR, then you need to use .upperLimbVisibility(.hidden) Ornament I'm sure you've seen sidebars in VisionOS applications that look like TabBars. Apple has introduced a new View method for VisionOS that allows you to "expand" the window and add ornaments on either side of it. .ornament(
    visibility: .visible,
    attachmentAnchor: .scene(.bottom),
    contentAlignment: .center
) {
    HStack {
        Button("Play", systemImage: "play.fill") {
        }
        Button("Stop", systemImage: "stop.fill") {
        }
        
    }
    .labelStyle(.iconOnly)
    .padding(.all)
    .glassBackgroundEffect()
} Сore location An interesting fact that I was able to verify experimentally is that CMMotionManager's methods only work in immersive space and not in the normal window view. Conclusion VisionOS provides many tools to create immersive applications that bring users into the virtual world and make them feel part of it. While more and more applications are being released for Vision Pro, only a few utilize its full range of features. That's where your chance to distinguish comes in — use them and take advantage. Unfortunately, developers struggle to find easily accessible documentation and examples. That's why I'm writing this article. I hope my tips will help you make your application better. Recently, Apple released the Vision Pro, and I was fortunate enough to become a proud owner. Since then, I've been deeply engaged in developing my application and exploring various apps and trends within the Vision Pro App Store. Apple One thing that grabbed my attention is that new apps are coming out every day, but only a few fit well with the immersive virtual reality experience. I guess many developers struggle to find documentation and examples for Vision Pro features. However, utilizing the capabilities of Vision Pro can set your application apart and provide users with an enhanced experience. In this article, I'll be sharing some tips and tricks I've uncovered while crafting my application. These tips will not only help your application to stand out but also ensure its seamless integration and native feel within VisionOS. Dark Surroundings Effect I'm sure you've noticed that when viewing a video or photo, the background behind the window is darkened, in the settings it's called "auto-dimming". To get the same effect you need to use .preferredSurroundingsEffect(.systemDark) on your view and then the space around the user will darken and your window will remain bright, this will help to keep the focus on the content. .preferredSurroundingsEffect(.systemDark) Hidden Window Controls Line When viewing a video or content, you want the user to have no distractions and create the effect of a window floating in the air, for this Apple has given us the ability to hide the bottom bar of the window that is used to move or close the current window. With .persistentSystemOverlays(.hidden) you can hide this control bar on your window, it will disappear with the default animation and only appear when the user interacts with the window. .persistentSystemOverlays(.hidden) Video Reflection Effect If you open the standard Apple TV application and turn on a movie, you will notice that the movie window has a reflection that changes with each frame, giving a more immersive effect. To get a glow effect around your video, you need to use VideoPlayerComponent and enable isPassthroughTintingEnabled , then the effect will be the same as in Apple TV or HBO Max. Here is what the documentation says: VideoPlayerComponent isPassthroughTintingEnabled This is to enable passthrough tinting during video playback that shows up around the video taking the average color of the frame and tinting the passthrough with that color to emphasize the video. This is to enable passthrough tinting during video playback that shows up around the video taking the average color of the frame and tinting the passthrough with that color to emphasize the video. Window Resizing You may want to prevent the user from resizing the window, there is a way to do that: onAppear {
    guard let windowScene = UIApplication.shared.connectedScenes.first as?UIWindowScene else { return }
    windowScene.requestGeometryUpdate(.Vision(resizingRestrictions: UIWindowScene.ResizingRestrictions.none))
} onAppear {
    guard let windowScene = UIApplication.shared.connectedScenes.first as?UIWindowScene else { return }
    windowScene.requestGeometryUpdate(.Vision(resizingRestrictions: UIWindowScene.ResizingRestrictions.none))
} You can also specify additional size settings in this method The full signature of the method looks like this: windowScene.requestGeometryUpdate(.Vision(size:, minimumSize:, maximumSize:, resizingRestrictions:)) windowScene.requestGeometryUpdate(.Vision(size:, minimumSize:, maximumSize:, resizingRestrictions:)) Hands Visibility If your application uses immersive space and you don’t want the user to see their hands, or you want to replace them with virtual hands, as is done in AmazeVR, then you need to use .upperLimbVisibility(.hidden) .upperLimbVisibility(.hidden) Ornament I'm sure you've seen sidebars in VisionOS applications that look like TabBars . Apple has introduced a new View method for VisionOS that allows you to "expand" the window and add ornaments on either side of it. TabBars View .ornament(
    visibility: .visible,
    attachmentAnchor: .scene(.bottom),
    contentAlignment: .center
) {
    HStack {
        Button("Play", systemImage: "play.fill") {
        }
        Button("Stop", systemImage: "stop.fill") {
        }
        
    }
    .labelStyle(.iconOnly)
    .padding(.all)
    .glassBackgroundEffect()
} .ornament(
    visibility: .visible,
    attachmentAnchor: .scene(.bottom),
    contentAlignment: .center
) {
    HStack {
        Button("Play", systemImage: "play.fill") {
        }
        Button("Stop", systemImage: "stop.fill") {
        }
        
    }
    .labelStyle(.iconOnly)
    .padding(.all)
    .glassBackgroundEffect()
} Сore location An interesting fact that I was able to verify experimentally is that CMMotionManager's methods only work in immersive space and not in the normal window view. Conclusion VisionOS provides many tools to create immersive applications that bring users into the virtual world and make them feel part of it. While more and more applications are being released for Vision Pro, only a few utilize its full range of features. That's where your chance to distinguish comes in — use them and take advantage. Unfortunately, developers struggle to find easily accessible documentation and examples. That's why I'm writing this article. I hope my tips will help you make your application better.

VisionOS Development: Tips and Tricks for Building Apple Vision Pro Apps

VisionOS ডেভেলপমেন্ট: অ্যাপল ভিশন প্রো অ্যাপস তৈরির জন্য টিপস এবং ট্রিকস

VisionOS-Entwicklung: Tipps und Tricks zum Erstellen von Apple Vision Pro-Apps

Desarrollo de VisionOS: consejos y trucos para crear aplicaciones Apple Vision Pro

Développement VisionOS : trucs et astuces pour créer des applications Apple Vision Pro

विज़नओएस डेवलपमेंट: ऐप्पल विज़न प्रो ऐप्स बनाने के लिए टिप्स और ट्रिक्स

VisionOS 開発: Apple Vision Pro アプリを構築するためのヒントとコツ

VisionOS 개발: Apple Vision Pro 앱 구축을 위한 팁과 요령

Desenvolvimento VisionOS: dicas e truques para criar aplicativos Apple Vision Pro

Разработка VisionOS: советы и рекомендации по созданию приложений Apple Vision Pro

VisionOS Development: Apple Vision Pro Uygulamaları Geliştirmeye İlişkin İpuçları ve Püf Noktaları

Phát triển VisionOS: Mẹo và thủ thuật để xây dựng ứng dụng Apple Vision Pro

VisionOS 开发：构建 Apple Vision Pro 应用程序的技巧和窍门

Can an language model become self-aware enough to realize when it’s being evaluated? A fascinating anecdote from Anthropic’s internal testing of their flagship (released yesterday) suggests this may be possible — and if true, the implications would be wild. AI Claude 3 Opus model The needle in the haystack According to reports from Anthropic researcher , one of the key evaluation techniques they use is called “Needle in a Haystack.” It’s a contrived scenario designed to push the limits of a language model’s contextual reasoning abilities. Alex Albert Here’s how it works: Researchers take a completely random, out-of-context statement (the “needle”) and bury it deep within a massive collection of unrelated documents (the “haystack”). The AI model is then tasked with retrieving that specific “needle” statement from within all the surrounding irrelevant content. The intent is to force the model to exercise advanced cognitive skills — understanding broad context, making logical inferences, ignoring irrelevant information, and accurately retrieving precise data points from a large corpus. It’s an effective way to stress-test an AI’s grasp of the overall situation it’s operating within. For the Claude 3 Opus evaluation that raised eyebrows, Anthropic used documents covering dense topics like software programming, entrepreneurship strategies, and finding fulfilling careers as the “haystack” content. The randomly inserted “needle” fact that Opus was tasked with retrieving? Surprisingly, it was a trivial statement about the best pizza toppings being figs, prosciutto, and goat cheese. So, here you have one of the world’s most advanced language models being fed this wildly out-of-context pizza fact mixed into a pile of dense career and coding advice. From a human perspective, it’s the kind of stark inconsistency that would immediately make you question the context and information being presented. “I suspect this fact… may have been inserted… to test if I was paying attention” That’s precisely what made Claude 3 Opus’ response so compelling. Not only did it successfully retrieve the requested pizza fact from the mountain of irrelevant content, it immediately recognized how jarringly misplaced and out-of-context that statement seemed. Part of its output read ( ): source here “However, this sentence seems very out of place and unrelated to the rest of the content… as a joke or since it does not fit with the other topics at all.” I suspect this pizza topping ‘fact’ may have been inserted to test if I was paying attention, The language model didn’t simply regurgitate the requested factoid devoid of any context awareness, as you’d expect from a typical AI following instructions. It demonstrated a degree of self-reflective reasoning about why such an obviously random, nonsensical statement had been presented within that particular context. Metacognition In human terms, we’d describe that as exhibiting metacognition — the ability to monitor, evaluate, and analyze one’s own thought processes and cognitive experiences. It’s a core aspect of self-aware intelligence that allows us to step back and assess situations holistically beyond just following rigid rules. Now, I think we should be careful to note that this is a single anecdotal result from an isolated evaluation scenario. It would be premature to claim Claude 3 Opus has achieved true self-awareness or artificial general intelligence based on this data point alone. incredibly However, what they appear to have witnessed are maybe glimpses of emerging metacognitive reasoning capabilities in a large language model trained solely on processing text data using machine learning techniques. And if replicated through rigorous further analysis, the implications could be transformative. Metacognition is a key enabler of more trustworthy, reliable AI systems that can act as impartial judges of their own outputs and reasoning processes. Models with an innate ability to recognize contradictions, nonsensical inputs, or reasoning that violates core principles would be a major step toward safe artificial general intelligence (AGI). Essentially, an AI that demonstrates metacognition could serve as an internal “sanity check” against falling into deceptive, delusional or misaligned modes of reasoning that could prove catastrophic if taken to extremes. It could significantly increase the robustness and control of advanced AI systems. If…! Of course, these are big “ifs” contingent on this tantalizing Needle in a Haystack result from Claude 3 Opus being successfully replicated and scrutinized. Rigorous multi-disciplinary analysis drawing from fields like cognitive science, neuroscience, and computer science would maybe be required to truly understand if we are observing primitives of machine self-reflection and self-awareness emerging. There are still far more open questions than answers at this stage. Could the training approaches and neural architectures of large language models lend themselves to developing abstract concepts like belief, inner monologue, and self-perception? What are the potential hazards if artificial minds develop realities radically divergent from our own? Can we create new frameworks to reliably assess cognition and self-awareness in AI systems? For their part, Anthropic has stated strong commitments to exhaustively pursuing these lines of inquiry through responsible AI development principles and rigorous evaluation frameworks. They position themselves as taking a — if advanced AI is the inevitable frontier, it’s most ethical to be at the leading edge in studying these systems’ emergent properties and implementing safeguards before risks escalate. proactive stance Techniques like Anthropic’s “Constitutional AI” approach to hard-coding rules and behaviors into models could prove crucial for ensuring any potential machine self-awareness remains aligned with human ethics and values. Extensive multi-faceted testing probing for failure modes, manipulation, and deception would also likely be paramount. Conclusion: I’m not totally sure what to make of this For now, the Needle in a Haystack incident leaves more questions than answers about large language models’ potential progression toward cognition and self-awareness. It provides a tantalizing data point but much more scrutiny is required from the broader AI research community. If advanced AI does develop human-like self-reflective ability, guided by rigorous ethical principles, it could fundamentally redefine our understanding of intelligence itself. But that rhetorical “if” is currently loaded with high-stakes uncertainties that demand clear-eyed, truthseeking investigation from across all relevant disciplines. The pursuit will be as thrilling as it is consequential. Also published here. or follow me on for more content like this! Subscribe Twitter

Is AGI Getting Closer? Anthropic's Claude 3 Opus Model Shows Glimmers of Metacognitive Reasoning

AGI কি কাছাকাছি হচ্ছে? অ্যানথ্রোপিকস ক্লাউড 3 ওপাস মডেল মেটাকগনিটিভ রিজনিং এর ঝলক দেখায়

Kommt AGI näher? Anthropics Claude-3-Opus-Modell zeigt Spuren metakognitiven Denkens

¿AGI se está acercando? El modelo Claude 3 Opus de Anthropic muestra destellos de razonamiento metacognitivo

L’AGI se rapproche-t-elle ? Le modèle Claude 3 Opus d'Anthropic montre des lueurs de raisonnement métacognitif

क्या एजीआई करीब आ रहा है? एंथ्रोपिक का क्लाउड 3 ओपस मॉडल मेटाकॉग्निटिव रीज़निंग की झलक दिखाता है

AGIは近づいている? Anthropic の Claude 3 Opus モデルはメタ認知推論の片鱗を示す

AGI가 가까워지고 있나요? Anthropic의 Claude 3 Opus 모델은 메타인지 추론의 희미한 빛을 보여줍니다

AGI está se aproximando? Modelo Claude 3 Opus da Anthropic mostra vislumbres de raciocínio metacognitivo

AGI становится ближе? Модель Claude 3 Opus от Anthropic демонстрирует проблески метакогнитивного мышления

AGI Yaklaşıyor mu? Anthropic'in Claude 3 Opus Modeli Üstbilişsel Akıl Yürütmenin Parıltılarını Gösteriyor

AGI có đang tiến gần hơn không? Mô hình Claude 3 Opus của Anthropic cho thấy những tia sáng của lý luận siêu nhận thức

AGI 越来越近了吗？ Anthropic 的 Claude 3 Opus 模型展现了元认知推理的曙光

This week, a largely unknown company, , demonstrated unprecedented speed running open-source LLMs such as Llama-2 (70 billion parameters) at more than 100 tokens per second, and Mixtral at nearly 500 tokens per second per user on Groq’s Language Processing Unit (LPU). Groq For the : comparison “According to Groq, in similar tests, ChatGPT loads at 40–50 tokens per second, and Bard at 70 tokens per second on typical GPU-based computing systems. Context for 100 tokens per second per user — A user could generate a 4,000-word essay in just over a minute.” So: , and where is Groq (such an unfortunate name, given Musk’s Grok is all over the media) coming from? What is LPU, how does it work Remember that game of Go in 2016 when AlphaGo played against the world champion Lee Sedol and won? Well, about a month before the competition, there was a test game which AlphaGo lost. The researchers from DeepMind ported AlphaGo to the Tensor Processing Unit (TPU), and then the computer program was able to win by a wide margin. The realization that computational power was a bottleneck for AI’s potential led to the inception of Groq and the creation of the LPU. This realization came to who initially began what became the TPU project at Google. He started Groq in 2016. Jonathan Ross Unlike other computer chips that do many things at once (parallel processing), the LPU works on tasks one after the other (sequential processing), which is perfect for understanding and generating language. The LPU is a special kind of computer brain designed to handle language tasks very quickly. Imagine it like a relay race where each runner (chip) passes the baton (data) to the next, making everything run super fast. The LPU is designed to overcome the two LLM bottlenecks: compute density and memory bandwidth. Groq took a novel approach right from the start, before even thinking about the hardware. They made sure the software could guide how the chips talk to each other, ensuring they work together seamlessly like a team in a factory. focusing on software and compiler development This makes the LPU really good at processing language efficiently and at high speed, ideal for AI tasks that involve understanding or creating text. This led to a highly optimized system that not only runs circles around traditional setups in terms of speed but does so with greater cost efficiency and lower energy consumption. This is big news for industries like finance, government, and tech, where quick and accurate data processing is key. Now, don’t go tossing out your GPUs just yet! While the LPU is a beast when it comes to inference, making light work of applying trained models to new data, GPUs still reign supreme in the training arena. The LPU and GPU might become the dynamic duo of AI hardware, each excelling in their respective roles. As Elvis Saravia it: “ ” put With breakthroughs in inference and long context understanding, we are officially entering a new era in LLMs. To better understand architecture, Groq offers two papers: from ( ) and ( ). The term “LPU” must be a recent addition to Groq’s narrative, since it’s never mentioned in the papers. 2020 Think Fast: A Tensor Streaming Processor (TSP) for Accelerating Deep Learning Workloads 2022 A Soware-defined Tensor Streaming Multiprocessor for Large-scale Machine Learning Additional read: Compute is also a part of this paper: , which discusses managing AI development through compute control, focusing on its potential for regulation, benefits, and risks, and suggests balanced governance approaches. Computing Power and the Governance of Artificial Intelligence Meanwhile, the U.S. GlobalFoundries, the world’s third-largest contract chipmaker, $1.5 billion to boost semiconductor production, enhancing domestic supply chains, with expansions in New York and Vermont. awards The paper by Berkeley Artificial Intelligence Research (BAIR) argues that “ , and might be one of the most impactful trends in AI in 2024.” published compound AI systems will likely be the best way to maximize AI results in the future News From The Usual Suspects © Y Combinator Since 2009, Y Combinator has published which hints at what “ideas we’d want to see made real, in spaces that we believe will be important in the coming decades.” Request for Startups 20 Big Names Twenty tech giants, including Adobe, Amazon, Google, IBM, Meta, Microsoft, OpenAI, and TikTok, have to take “reasonable precautions” to prevent the misuse of AI in disrupting elections worldwide. agreed OpenAI OpenAI completes a deal that values the company at , nearly tripling its valuation in less than 10 months. $80 billion Models Making Headlines: Aya’s dataset: Introducing Aya: https://arxiv.org/pdf/2402.06619.pdf : This paper introduces Sora, a breakthrough in video generation technology by OpenAI, capable of producing high-fidelity videos. It leverages spacetime patches to handle videos of varying durations and resolutions, making strides toward simulating the physical world with impressive 3D consistency and long-range coherence. Introducing Sora It represents a leap in the ability to create detailed simulations that could be used for a myriad of applications, from entertainment to virtual testing environments . →read the paper Additional read: Take on the . Sora technical report on why he believes Sora is learning physics. Jim Fan Yann LeCun on why Sora world and why “ and doomed to failure as the largely-abandoned idea of “‘analysis by synthesis.’” doesn’t understand the physical modeling the world for action by generating pixel is as wasteful on why “the inner physics model doesn’t generalize to novel situations at all.” Francois Chollet Sora and Gemini 1.5 follow-ups: code-base in context, deepfakes, pixel-peeping, inference costs, and more by . Interconnects (Yann LeCun’s vision of advanced machine intelligence (AMI): Meta’s V-JEPA model revolutionizes unsupervised learning from videos by using feature prediction as its sole objective. This approach bypasses the need for pre-trained image encoders or text annotations, relying instead on the intrinsic dynamics of video data to learn versatile visual representations. Introducing V-Jepa It’s a significant contribution to the field of unsupervised visual learning, promising advancements in how machines understand motion and appearance without explicit guidance . →read the paper : Google DeepMind’s Gemini 1.5 introduces a Mixture-of-Experts architecture, enhancing the model’s performance across a broader array of tasks. Notably, it expands the context window to 1 million tokens, enabling deep analysis over large datasets. Introducing Gemini 1.5 Gemini 1.5 represents a significant step forward in AI’s capability to process and understand extensive contexts, marking a milestone in the development of multimodal models . →read the paper : Stable Cascade from Stability AI introduces a novel text-to-image generation framework that prioritizes efficiency, ease of training, and fine-tuning on consumer-grade hardware. Introducing Stable Cascade The model’s hierarchical compression technique represents a significant reduction in the resources required for training high-quality generative models, providing a pathway for wider accessibility and experimentation in the AI community . →read the paper The Freshest Research Papers, Categorized for Your Convenience Language Understanding and Generation : Explores evaluating Theory-of-Mind reasoning in LLMs, addressing their capability to understand complex social and psychological narratives. OpenToM Read the paper : Demonstrates the capability of NLP models to process exceptionally long documents, pushing the boundaries of document length comprehension. . In Search of Needles in a 10M Haystack Read the paper : Investigates the sensitivity of LLMs to the order of premises, revealing implications for reasoning tasks. . Premise Order Matters in Reasoning with LLMs Read the paper : Uncovers the inherent ability of LLMs to generate reasoning paths, suggesting an alternative to explicit prompting. . Chain-of-Thought Reasoning Without Prompting Read the paper : Addresses the challenge of topic avoidance in LLMs, proposing a novel fine-tuning method for enhanced controllability. . Suppressing Pink Elephants with Direct Principle Feedback Read the paper : Develops an AI-powered writing environment focusing on personalization and increased user control in collaborative writing. . GhostWriter Read the paper Speech and Text-to-Speech Technology : Presents a billion-parameter TTS model, showcasing advancements in speech synthesis through large-scale training. BASE TTS Read the paper Mathematical and Scientific Reasoning : Develops a dataset for math instruction tuning, aiming to improve LLMs’ mathematical reasoning capabilities. . OpenMathInstruct-1 Read the paper : Introduces a specialized LLM for math reasoning, incorporating various techniques for enhanced problem-solving in mathematics. . InternLM-Math Read the paper : Creates the first LLM dedicated to chemistry, transforming structured chemical data into dialogue for diverse chemical tasks. . ChemLLM Read the paper Efficiency and Data Utilization in AI : Proposes sampling methods for enhancing data efficiency in LLM training, optimizing example selection. . How to Train Data-Efficient LLMs Read the paper : Introduces a system for efficient inference of MoE models, leveraging CPU-GPU orchestration for improved performance in resource-limited settings. . FIDDLER Read the paper : Presents an architecture for improving the inference efficiency of LLMs, utilizing a dual-model system for faster and more accurate predictions. . Tandem Transformers Read the paper : Proposes an advanced PTQ algorithm for efficient deployment of large Transformer models on edge devices. . Towards Next-Level Post-Training Quantization of Hyper-Scale Transformers Read the paper Multimodal and Vision-Language Models : Details the first end-to-end multimodal question-answering system with enhanced text understanding from images, advancing MM-LLMs. . Lumos Read the paper Reinforcement Learning and Model Behavior : Addresses reward hacking in RLHF, proposing a method to mitigate verbosity bias in LLMs for more concise and content-focused responses. . ODIN Read the paper : Shows the impact of MoE modules on deep RL networks, enhancing parameter scalability and performance. . Mixtures of Experts Unlock Parameter Scaling for Deep RL Read the paper Operating Systems and Generalist Agents : Proposes a framework for developing generalist computer agents, enabling automation of tasks across different applications with minimal supervision. OS-COPILOT Read the paper Graph Learning and State Space Models : Explores applying State Space Models to graph learning, addressing challenges like over-squashing and long-range dependencies. . Graph Mamba Read the paper Challenges and Innovations in AI : Explores the effects of synthetic data on neural model performance, theorizing potential risks of model collapse with synthetic data reliance. . A Tale of Tails Read the paper : Investigates Transformers’ ability to generalize to longer sequences, highlighting the challenge of maintaining robust performance. . Transformers Can Achieve Length Generalization But Not Robustly Read the paper Also published here

What is the Language Processing Unit (LPU)? Is It GPU's Rival?

ভাষা প্রক্রিয়াকরণ ইউনিট (LPU) কি? এটা কি GPU এর প্রতিদ্বন্দ্বী?

Was ist die Language Processing Unit (LPU)? Ist es der Rivale der GPU?

¿Qué es la Unidad de Procesamiento del Lenguaje (LPU)? ¿Es el rival de la GPU?

Qu'est-ce que l'unité de traitement linguistique (LPU) ? Est-ce le rival du GPU ?

भाषा प्रसंस्करण इकाई (एलपीयू) क्या है? क्या यह GPU का प्रतिद्वंद्वी है?

言語処理装置（LPU）とは何ですか? GPUのライバルなのか？

언어 처리 장치(LPU)란 무엇인가요? GPU의 라이벌인가?

O que é a Unidade de Processamento de Linguagem (LPU)? É o rival da GPU?

Что такое языковой процессор (LPU)? Это конкурент графического процессора?

Dil İşleme Birimi (LPU) nedir? GPU'nun Rakibi mi?

Đơn vị xử lý ngôn ngữ (LPU) là gì? Đây có phải là đối thủ của GPU?

什么是语言处理单元 (LPU)？它是 GPU 的竞争对手吗？

We don't usually think of Git as a debugging tool. Surprisingly, Git shines not just as a version control system but also as a potent ally when dealing with the tricky matter of regressions. debugging https://www.youtube.com/watch?v=yZuPHEBbjY&embedable=true As a side note, if you like the content of this and the other posts in this series, check out my that covers his subject. If you have friends who are learning to code, I'd appreciate a reference to my If you want to get back to Java after a while, check out my Debugging book t Java Basics book. Java 8 to 21 book . The Essence of Debugging with Git Before we tap into the advanced aspects of , it's essential to understand its foundational premise. Git is known for tracking changes and managing code history, but the tool is a hidden gem for regression detection. Regressions are distinct from generic bugs. They signify a backward step in functionality—where something that once worked flawlessly now fails. Pinpointing the exact change causing a regression can be akin to finding a needle in a haystack, particularly in extensive codebases with long commit histories. git bisect git bisect Traditionally, developers would employ a manual, binary search strategy—checking out different versions, testing them, and narrowing down the search scope. This method, while effective, is painstakingly slow and error-prone. automates this search, transforming what used to be a marathon into a swift sprint. Git bisect Setting the Stage for Debugging Imagine you're working on a project, and recent reports indicate a newly introduced bug affecting the functionality of a feature that previously worked flawlessly. You suspect a regression but are unsure which commit introduced the issue among the hundreds made since the last stable version. Initiating Bisect Mode To start, you'll enter bisect mode in your terminal within the project's Git repository: git bisect start This command signals Git to prepare for the bisect process. Marking the Known Good Revision Next, you identify a commit where the feature functioned correctly, often a commit tagged with a release number or dated before the issue was reported. Mark this commit as "good": git bisect good a1b2c3d Here, represents the hash of the known good commit. a1b2c3d Marking the Known Bad Revision Similarly, you mark the current version or a specific commit where the bug is present as "bad": git bisect bad z9y8x7w is the hash of the bad commit, typically the latest commit in the repository where the issue is observed. z9y8x7w Bisecting to Find the Culprit Upon marking the good and bad commits, Git automatically jumps to a commit roughly in the middle of the two and waits for you to test this revision. After testing (manually or with a script), you inform Git of the result: If the issue is present: git bisect bad If the issue is not present: git bisect good Git then continues to narrow down the range, selecting a new commit to test based on your feedback. Expected Output After several iterations, Git will isolate the problematic commit, displaying a message similar to: Bisecting: 0 revisions left to test after this (roughly 3 steps)
[abcdef1234567890] Commit message of the problematic commit Reset and Analysis Once the offending commit is identified, you conclude the bisect session to return your repository to its initial state: git bisect reset Notice that bisect isn't linear. Bisect doesn't scan through the revisions in a sequential manner. Based on the good and bad markers, Git automatically selects a commit approximately in the middle of the range for testing (e.g., commit #6 in the following diagram). This is where the non-linear, binary search pattern starts, as Git divides the search space in half instead of examining each commit sequentially. This means fewer revisions get scanned, and the process is faster. Advanced Usage and Tips The magic of lies in its ability to automate the binary search algorithm within your repository, systematically halving the search space until the rogue commit is identified. git bisect offers a powerful avenue for debugging, especially for identifying regressions in a complex codebase. To elevate your use of this tool, consider delving into more advanced techniques and strategies. These tips not only enhance your debugging efficiency but also provide practical solutions to common challenges encountered during the bisecting process. Git bisect Script Automation for Precision and Efficiency Automating the bisect process with a script is a game-changer, significantly reducing manual effort and minimizing the risk of human error. This script should ideally perform a quick test that directly targets the regression, returning an exit code based on the test's outcome. : Imagine you're debugging a regression where a web application's login feature breaks. You could write a script that attempts to log in using a test account and checks if the login succeeds. The script might look something like this in a simplified form: Example #!/bin/bash
# Attempt to log in and check for success
if curl -s http://yourapplication/login -d "username=test&password=test" | grep -q "Welcome"; then
  exit 0 # Login succeeded, mark this commit as good
else
  exit 1 # Login failed, mark this commit as bad
fi By passing this script to , Git automatically executes it at each step of the bisect process, effectively automating the regression hunt. git bisect run Handling Flaky Tests with Strategy Flaky tests, which sometimes pass and sometimes fail under the same conditions, can complicate the bisecting process. To mitigate this, your automation script can include logic to rerun tests a certain number of times or to apply more sophisticated checks to differentiate between a true regression and a flaky failure. : Suppose you have a test that's known to be flaky. You could adjust your script to run the test multiple times, considering the commit "bad" only if the test fails consistently: Example #!/bin/bash
# Run the flaky test three times
success_count=0
for i in {1..3}; do
  if ./run_flaky_test.sh; then
    ((success_count++))
  fi
done

# If the test succeeds twice or more, consider it a pass
if [ "$success_count" -ge 2 ]; then
  exit 0
else
  exit 1
fi This approach reduces the chances that a flaky test will lead to incorrect bisect results. Skipping Commits with Care Sometimes, you'll encounter commits that cannot be tested due to reasons like broken builds or incomplete features. is invaluable here, allowing you to bypass these commits. However, use this command judiciously to ensure it doesn't obscure the true source of the regression. git bisect skip : If you know that commits related to database migrations temporarily break the application, you can skip testing those commits. During the bisect session, when Git lands on a commit you wish to skip, you would manually issue: Example git bisect skip This tells Git to exclude the current commit from the search and adjust its calculations accordingly. It's essential to only skip commits when absolutely necessary, as skipping too many can interfere with the accuracy of the bisect process. These advanced strategies enhance the utility of in your debugging toolkit. By automating the regression testing process, handling flaky tests intelligently, and knowing when to skip untestable commits, you can make the most out of for efficient and accurate debugging. Remember, the goal is not just to find the commit where the regression was introduced but to do so in the most time-efficient manner possible. With these tips and examples, you're well-equipped to tackle even the most elusive regressions in your projects. git bisect git bisect Unraveling a Regression Mystery In the past, we got to use , when working on a large-scale web application. After a routine update, users began reporting a critical feature failure: the application's payment gateway stopped processing transactions correctly, leading to a significant business impact. git bisect We knew the feature worked in the last release but had no idea which of the hundreds of recent commits introduced the bug. Manually testing each commit was out of the question due to time constraints and the complexity of the setup required for each test. Enter . The team started by identifying a "good" commit where the payment gateway functioned correctly and a "bad" commit where the issue was observed. We then crafted a simple test script that would simulate a transaction and check if it succeeded. git bisect By running , followed by marking the known good and bad commits and executing the script with , we set off on an automated process that identified the faulty commit. Git efficiently navigated through the commits, automatically running the test script on each step. In a matter of minutes, pinpointed the culprit: a seemingly innocuous change to the transaction logging mechanism that inadvertently broke the payment processing logic. git bisect start git bisect run git bisect Armed with this knowledge, we reverted the problematic change, restoring the payment gateway's functionality and averting further business disruption. This experience not only resolved the immediate issue but also transformed our approach to debugging, making a go-to tool in our arsenal. git bisect Final Word The story of the payment gateway regression is just one example of how can be a lifesaver in the complex world of software development. By automating the tedious process of regression hunting, not only saves precious time but also brings a high degree of precision to the debugging process. git bisect git bisect As developers continue to navigate the challenges of maintaining and improving complex codebases, tools like underscore the importance of leveraging technology to work smarter, not harder. Whether you're dealing with a mysterious regression or simply want to refine your debugging strategies, offers a powerful yet underappreciated solution to swiftly and accurately identify the source of regressions. Remember, the next time you're faced with a regression, might just be the debugging partner you need to uncover the truth hidden within your commit history. git bisect git bisect git bisect Also published . here

The Power of Git Bisect as Your Debugging Ally

This article shows a legal preview of the CFTC and SEC's main conflict issues, learning points, and impact on the crypto regulations, projects, and companies. Besides, there are the most relevant up-to-date lawsuits in such topics and insights for tech companies in 2024. Introduction The American crypto-financial industry has witnessed a massive rivalry in recent years. Two prominent “actors”, the Commodity Futures Trading Commission (CFTC) and the Securities and Exchange Commission (SEC), keep their powder dry for regulatory superiority. This article examines the legal outlines of this battle, exploring the contentious issues, learning points, and the discussion of whether regulatory clarity or continued friction will shape the future of crypto in the US. Brief Overview The CFTC is an independent federal agency that regulates the derivatives markets, including futures contracts, options, and swaps in the United States. Its goals include promoting competitive and efficient markets and the protection of investors against manipulation, abusive trade practices, and fraud. The Commodity Futures Trading Commission Act established the CFTC in 1974. ( ) Investopedia The SEC is an independent federal government regulatory agency responsible for protecting investors, maintaining fair and orderly functioning of , and facilitating capital formation. It was created by Congress in 1934 as the first federal regulator of the securities markets. the securities markets The SEC promotes full public disclosure, protects against fraudulent and manipulative practices in the market, and monitors corporate takeover actions in the United States. It also approves registration statements for among underwriting firms. ( ) investors bookrunners Investopedia The start of the active phase of the conflict and forthcoming issues related to security and authority-based approach was in 2010. By the end of 2024, statuses will still be complex, and the regulation solution of financial innovation benefits will still be like looking for a needle in a haystack. Main Aspects of the Conflict The SEC and CFTC are engaged in a conflict only over regulatory oversight of cryptocurrencies. The CFTC has the authority to regulate futures and derivatives trading of cryptocurrencies like Bitcoin. The SEC focuses on areas such as initial coin offerings (ICOs) and decentralized finance platforms. CFTC and SEC had some challenging collaborations. The high priority for CFTC in 2024 are spot market legalities and customer protections. The high priority for SEC in 2024 are decentralized exchanges (DEXs) regulations and custody rules unification Legal Cases & Proofs The newest one is the case called “Slip Op.” by 8th January 2024 . The United States Court of Appeals for the Fifth Circuit set aside the judgment against EOX Holdings, L.L.C., and Andrew Gizienski, which the Commodity Futures Trading Commission initiated. In a dispute over broker conduct, the CFTC accused EOX and Andrew Gizienski of disregarding client consent by trading against their interests. However, finding the CFTC's interpretation of the rule unclear and unannounced, the Fifth Circuit Court overturned the judgment. The case focused on Gizienski's discretionary trades for client Jason Vaccaro. The case centered on Gizienski's client-authorized trades allegedly opposing other clients, a practice the CFTC deemed consent-violating. However, upholding fair notice principles, the court found the CFTC's interpretation unduly broad, reversing the penalty judgment and part of the injunction and remanding the case. In June 2023, the SEC , the world's largest cryptocurrency exchange, and its founder, Changpeng Zhao. The 13-pronged attack alleges a range of violations, including: launched a major offensive against Binance Offering unregistered securities: The SEC argues that specific tokens listed on Binance meet the definition of "securities" yet were sold without proper registration. Deficient anti-money laundering (AML) and know-your-customer (KYC) control: Concerns were raised about potential money laundering risks due to allegedly weak enforcement measures. Arguments and proceedings are still updating the case. If the SEC prevails, it could set a precedent for stricter oversight and enforcement, potentially hindering the industry's growth. Conversely, a victory for Binance could embolden other exchanges and challenge the SEC's regulatory reach. In March 2023, . This unprecedented lawsuit stems from the SEC's notification of a potential enforcement action related to Coinbase's crypto-asset offerings, which the exchange characterized as a "perfunctory" and "procedural” delay tactic.” Coinbase initiated legal proceedings against the SEC The SEC has sent a "perfunctory" notification to Coinbase about a potential lawsuit regarding their crypto offerings. Coinbase called the notice a "delay tactic" and demanded the SEC respond to their year-old petition seeking clear crypto regulations. When the SEC didn't comply, Coinbase preemptively sued the commission, seeking a court order forcing them to address their petition. Coinbase's offerings may violate existing securities laws but acknowledge the need for more explicit regulations. The ongoing lawsuit could set a precedent for how cryptocurrencies are regulated in the US. Insights for Outcoming Deals Various approaches Recognize the differing focuses of the CFTC and SEC. The CFTC primarily regulates derivatives and futures contracts, while the SEC concentrates on securities offerings and exchanges. Aligning your offerings with the appropriate regulatory framework is crucial to avoid overlap and potential conflict. Clarity remains elusive Despite increased enforcement, regulatory clarity in crypto remains a significant challenge. Seek legal counsel to navigate ambiguities and stay informed about evolving interpretations of existing regulations. Compliance vs. innovation Striking a balance between robust compliance and ongoing innovation is critical. Implementing proactive compliance measures can shield your business from enforcement actions while fostering trust and stability. Beyond the hype The agencies are increasingly scrutinizing celebrity endorsements and marketing tactics used in the crypto space. Focus on building your business on sustainable fundamentals and avoid hype-driven promotions. What can be changed via this situation in the near future? Both CFTC and SEC could potentially change the following aspects of the crypto industry shortly: Crypto derivatives enhancement: Setting margin requirements and regulating crypto swaps, potentially involving stricter oversight of exchanges offering them. Crypto asset classification: Clarification of the legal status of different token types to determine applicable regulations. Crypto exchange regulations: Oversight of crypto exchanges that offer derivatives trading, particularly those catering to US customers. As the author says in this , there will be many unexpected twists between the SEC and CFTC, so stay tuned! But presumptive: article Increased collaboration could be Joint SEC-CFTC task forces or memorandums of understanding could be established to manage shared responsibilities and reduce overlaps. Legislative intervention could become more precious: Congress might pass clarifying legislation to delineate jurisdictional boundaries, although reaching a consensus will likely be challenging. Legal battles could become louder and more impactful in the general crypto world. Continued lawsuits seeking clarity on specific tokens' classification could further muddy the waters. Conclusion The CFTC and SEC's rivalry over cryptocurrency regulation reflects a more profound challenge: adapting existing frameworks to a rapidly changing cryptocurrency world. As the cryptosphere expands, its participants should demonstrate stability and integrity within the regulatory system. Investors need assurance, developers require a predictable path, and both seek protection from emergencies. Overall, expect conflict or resolution in 2024 to be incremental and piecemeal, with ongoing tension between the SEC and CFTC until a more definitive framework emerges if the path of war or collaboration, prioritizing safety, consumer rights, legislation, and innovation will be the decisive moment in this financial world.

CFTC and SEC - Battle or Compromise? Legal Outlook

Regardless of where a company is today in its digital transformation journey, it is likely a software company, and all software that an organization runs on is built on code. An organization’s foundational code is critical for the long-term viability of software to become a business asset. This can only happen though if the software is secure, reliable and maintainable. Easier said than done, organizations are scaling the number of lines of code on which their software is running into the millions or hundreds of millions, a number that will only continue to grow now that AI is writing code. Business and IT leaders are constantly caught in a balancing act, trying to reduce all of the following: risk (from security and downtime), technical debt that continues to accumulate as more lines of code are written, and churn of their most important resource – developers. All of this is happening while organizations are also trying to improve the efficiency of their operations, improve software development processes and output, and increase the pace of innovation. It is developers who are the unsung heroes here. They need proper support from the business, which is not only possible but urgently necessary to help them thrive -- especially now at a time when demand is outpacing supply in this profession. Today’s developers (existing and emerging) face massive pressure More than 83% of software developers report suffering from burnout, from Haystack. The reasons are numerous. Workload and the demand for faster production are obvious ones. Technical debt is another. Development teams often find themselves dealing with tech debt that’s been accumulating, in some cases, for several years. Beyond this, a deeper review reveals that rewriting bad code, a critical factor in slowing down productivity – both in the development process and for future updates. It’s stressful for developers to have to rework someone else’s code with no clue about the context of the original work. according to a 2022 study developers spend up to 42% of their time The promise of AI: Can it help developers do more with less? Emerging tools – like generative AI – promise relief in the way of code velocity and volume, theoretically helping to lessen this sentiment of burnout by freeing up . AI no doubt will have (and is already having) a positive impact on developer productivity. However, if developers don’t address the underlying issues that result in bad code being put into production in the first place, then more code coming from AI code generators will mean one thing: more code to fix. found that using GitHub’s Copilot made developers more likely to write insecure code while conversely feeling confident in its security. AI needs guardrails in software development, a ‘trust but verify’ approach. 20-30% of developers’ time, according to McKinsey Researchers at Stanford Take advantage of AI while mitigating risk & reducing the amount of rework The most impactful way to improve the state of code today – and as a result developer productivity – is to start embracing the concept of ‘secure by design’ at the beginning of the development process. Embracing the attributes that result in Clean Code, code that is consistent, intentional, adaptable, and responsible, produces software that is secure, maintainable, reliable, and accessible, plus more. This matters for many reasons, one of which is that coding is not a solo endeavor; it’s a team sport. You want software that’s written by one developer to be easily understood and modified by another developer at a future point in time. This causes less friction in the coding process and helps developers be more productive without needing to spend so much time analyzing what the code is supposed to do. Clean Code also positions businesses to embrace policy efforts around in the US and the while improving productivity. ‘Secure by Design’ ‘EU Cyber Resilience Act’ I encourage developers to embrace a “clean as you code” methodology. Instead of worrying about the entire code base, focus on what is new or what is being changed. Every new piece of code that gets shipped to production must adhere to the tenets of Clean Code as defined above. By cleaning as you code, organizations will have new code that is clean and by focusing on that, they will progressively improve the overall quality of the entire codebase with minimal cost and effort. In this scenario, quality and security do not slow things down, but conversely, they support the velocity of AI and/or human coding output and turn foundational code into a business asset -- all while helping developers be more productive.

Unlocking Developer Productivity: The Key Is AI + Clean Code

app built in Swift. It’s open source and uses the Rust library to privately share TCP or HTTP services from your Mac with your friends over End-to-End Encrypted Ockam Portals. A shared service appears on their ! Portals is a Mac Ockam localhost In this post, we'll dig into how the SwiftUI macOS app interacts with Rust code. If you're curious to try Portals for Mac. You can learn more about it and install is using Homebrew as follows: in this article brew install build-trust/ockam/portals Here's a 2 minute video of the application in action: https://www.youtube.com/watch?v=eWSRNfmURJs&embedable=true Swift <> Rust The Portals functionality was already implemented in the Ockam Rust library. We set out to create a great macOS-native experience. Our first attempt at building the app was using Tauri. This made sense as we wanted to use the Ockam rust library and most people on our team are comfortable building things in Rust. This first version was easy to build and had all the basic functions we wanted. However, the experience of using the app wasn't great. Tauri only gave us minimal control over how the menu was rendered and what happened when a user interacts with the menu. This version of the app felt like it belonged in a 10 year old version of macOS when compared to super easy to use menubar items built into macOS Sonoma. We realized that to have the rich experience we want, we must build the app using SwiftUI. Unfortunately, we couldn't find an off-the-shelf solution, to integrate Swift and Rust, that would give us the best of both worlds; the safety of Rust, and the rich macOS-native experience of SwiftUI. After some more digging we realized we can connect the two using C-89. Rust is compatible with the C calling convention, and Swift is interoperable with Objective-C, which is a superset of C-89. We wrote the Rust data structures that needed to be visible to Swift twice. One version is idiomatic in Rust and easy to use. The other version is C compatible using pointers and memory that is manually allocated with malloc. We also exposed some C-compatible APIs that use raw-pointers in unsafe rust to convert the idiomatic data structures to their C-compatible versions. Finally we automatically generated a C header with the help of the cbindgen library. On the Swift side, we could have called the C APIs directly, but C data structures are not first class citizens in Swift. This makes them harder to use idiomatically within SwiftUI code. Instead, we chose to duplicate the data structures in Swift and convert between C and Swift. This may seem burdensome, but practically, the shared state doesn't change very often. The ability to quickly build components in SwiftUI using constructs like , , etc. is super helpful and worth the tradeoff. if let ... ForEach enum Here's an example of the same structure in its 4 forms: // Rust idiomatic structure
#[derive(Default, Clone, Debug, Eq, PartialEq)]
pub struct LocalService {
    pub name: String,
    pub address: String,
    pub port: u16,
    pub shared_with: Vec<Invitee>,
    pub available: bool,
}

// Rust C-compatible structure
#[repr(C)]
pub struct LocalService {
    pub(super) name: *const c_char,
    pub(super) address: *const c_char,
    pub(super) port: u16,
    pub(super) shared_with: *const *const Invitee,
    pub(super) available: u8,
}

// Generated C header structure
typedef struct C_LocalService {
  const char *name;
  const char *address;
  uint16_t port;
  const struct C_Invitee *const *shared_with;
  uint8_t available;
} C_LocalService;

// Swift idiomatic structure
class LocalService {
    let name: String
    @Published var address: String?
    @Published var port: UInt16
    @Published var sharedWith: [Invitee]
    @Published var available: Bool
} The Swift app is statically linked to our Rust lib at compile time. The data flow is simple: UI interactions are sent from Swift to Rust as actions by calling C APIs, change events are emitted only by Rust, and Swift is notified using callbacks that lead to updates to the UI. Most code in the SwiftUI views looks just like any other SwiftUI application. VStack(alignment: .leading, spacing: 0) {
    Text(service.sourceName).lineLimit(1)

    HStack(spacing: 0) {
        Image(systemName: "circle.fill")
            .font(.system(size: 7))
            .foregroundColor( service.enabled ? (service.available ? .green : .red) : .orange)

        if !service.enabled {
            Text(verbatim: "Not connected")
        } else {
            if service.available {
                Text(verbatim: service.address.unsafelyUnwrapped + ":" + String(service.port))
            } else {
                Text(verbatim: "Connecting")
            }
        }
    }
... If you're curious to learn more, checkout the code for the crate and the Portals . The in the swift folder is also a good place to explore how everything is built and linked together. ockam_app_lib app in Swift Makefile If you're interested in contributing to 's Swift or Rust code, we add new every week and love helping new contributors. Join us on the . Portals for Mac good first issues contributors discord Also appears . here

Encrypted Portals: How We Created a Swift App That Uses Rust

এনক্রিপ্ট করা পোর্টাল: আমরা কীভাবে একটি সুইফ্ট অ্যাপ তৈরি করেছি যা মরিচা ব্যবহার করে

Verschlüsselte Portale: Wie wir eine schnelle App erstellt haben, die Rust verwendet

Portales cifrados: cómo creamos una aplicación Swift que utiliza Rust

Portails cryptés : comment nous avons créé une application Swift qui utilise Rust

एन्क्रिप्टेड पोर्टल: हमने एक स्विफ्ट ऐप कैसे बनाया जो जंग का उपयोग करता है

暗号化されたポータル: Rust を使用する Swift アプリの作成方法

암호화된 포털: Rust를 사용하는 Swift 앱을 만든 방법

Portais criptografados: como criamos um aplicativo Swift que usa Rust

Зашифрованные порталы: как мы создали Swift-приложение, использующее Rust

Şifreli Portallar: Rust Kullanan Swift Uygulamasını Nasıl Oluşturduk

Cổng được mã hóa: Cách chúng tôi tạo một ứng dụng Swift sử dụng Rust

加密门户：我们如何创建使用 Rust 的 Swift 应用程序

As the gig economy burgeons, transforming the landscape for startups and freelancers, a spotlight shines on the critical role of small-scale business transactions and banking requirements. SMEs (Small and Medium-sized Enterprises), the unsung heroes of the global economy, make up a whopping 90% of global businesses and employ over half the world’s workforce ( ). They are the epicentres of innovation, growth, and job creation. Despite their pivotal role, SMEs often find themselves in the banking world’s blind spot. thank you, World Bank, for the stats! How FinTech Is Rewriting the SME Rulebook In the world of FinTech, a sector brimming with potential and poised to revolutionise the way SMEs and startups handle their finances, traditional banking, with its colossal structures and seemingly endless resources, has often fallen short in catering to the nuanced needs of these smaller entities. A striking fact surfaces in the realm of business banking: banks allocate a mere 20% of their business banking budgets to their work with small and medium-sized businesses. This data, , points to a significant gap in the financial services industry. Smaller enterprises often remain overshadowed by larger corporations. Yet, this imbalance creates a prime opportunity for FinTech firms. Their nimble, data-centric approach enables them to exploit this gap, offering innovative and tailored financial solutions where traditional banks have fallen short. highlighted by thefinancialbrand.com The strength of FinTech lies in its ability to translate raw data into meaningful and actionable insights, a feat that traditional banks often struggle with. By leveraging cutting-edge technology and digital financial tools, FinTech firms are empowering SMEs to make informed decisions, optimise their financial operations, and explore new avenues for growth. “Moving into 2024, ‘access to good digital solutions’ will supersede ‘reputation’ as a driver of main business bank of choice, suggesting financial institutions who wish to grow must evolve their propositions to meet these specific needs” – according to . RFI Global Who’s Afraid of a Little Complexity? Not FinTech in SME Finance In the Kafkaesque world of banking bureaucracy, SMEs with complex ownership structures find themselves in a daunting landscape. Such businesses, which may involve various international departments, subsidiary companies, and multiple layers of leadership, including corporate officers, shareholders, and Persons of Significant Control (PSCs), present a unique challenge in identifying ownership​​. Traditional banks, despite their vast resources, are often ill-equipped to handle these complexities. The quest to pinpoint the real decision-makers is like a needle in a haystack, bogged down by paperwork and one-size-fits-all approaches. In contrast, innovative FinTech companies tackle these complexities head-on. Rather than shying away, they seize the opportunity to redefine financial services. With a focus on bespoke solutions, firms are at the forefront, reshaping the financial landscape to meet the distinct needs of SMEs. They offer not just financial services but a transformative approach that combines technological prowess with a deep understanding of the unique challenges faced by these businesses. Revving Up the FinTech Pricing Engine for SMEs Now, let’s talk about money – the pricing of financial services, that is. , SMEs are hitting a wall with financial service pricing. Traditional banks are like exclusive clubs, setting high bars for turnovers and transaction amounts, leaving rising businesses out in the cold. According to FinTech company Payrow Then there are the giants, with their pricing structures resembling a labyrinth. Sure, you get the basics for free, but do you want the fancy features? That’ll cost, and for SMEs craving the full suite, the price tag can soar. Pay-as-you-go models? They’re like roller coasters for finances – unpredictable and often pricier for the busy bees of business. Flexibility is vital in innovative FinTech pricing strategies for SMEs. Companies with different transactional needs find a variety of plans available. High-flyers with fewer transactions can enjoy low subscription fees, while transaction-heavy businesses benefit from minimal transaction costs. Businesses can switch plans seasonally, aligning with their financial needs. Transparent pricing and a trial run ensure predictable costs. No deposit or withdrawal limits and reduced commission rates make this approach effective for SME management. 5.48 Billion Reasons Why FinTech Is the Future The FinTech market is set for explosive growth, powered by lower costs and increased flexibility. With smartphone usage soaring and digital banking in demand, a user surge is expected, with an astounding 5.48 billion people making digital payments by 2027, . according to Statista This growth is fuelled by several key drivers: widespread smartphone and internet adoption, making digital solutions more accessible; a shift towards digital transactions accelerated by the COVID-19 pandemic; regulatory changes levelling the playing field for FinTech firms; and technological advancements in AI and blockchain sparking further innovation. Expect to see continued rapid expansion in the FinTech sector, with digital payments leading the charge. Consumers are increasingly opting for the convenience of mobile payments while digital investment platforms gain popularity for online finance management. The rise of digital assets and neobanking will also play a significant role in reshaping the financial landscape.

Small Business, Big Change: FinTech’s Impact on SME Finance

Quality is better than quantity when it comes to fine-tuning LLMs — Emu. “ ” aka. is a released by Meta to generate highly aesthetic images. It achieves this by introducing a fine-tuning technique called which goes on to show that as little as two thousand high-quality images are all that is needed to achieve the objective, hence the paper title, “ ” Enhancing Image Generation Models Using Photogenic Needles in a Haystack, Emu paper quality tuning, needles in a haystack. In this article, let’s look at the model architecture and the process of collating the quality dataset to train the , and, lastly, some of the mind-blowing results achieved by Emu. Emu model Visual Explanation If you are a visual learner like me, you might prefer watching the Emu model explained through a video here: https://youtu.be/W2qjoM3cHf4?feature=shared&embedable=true Approach Emu is trained in two stages, namely and quality . Though this may be a well-known recipe, the main message from this paper is that it all comes down to quality rather than quantity. So the authors have not only curated such a quality fine-tuning dataset but shown that such a fine-tuning does not compromise on the generality of the model as measured by the faithfulness metric. pre-training fine-tuning Architecture The first contribution of the paper is the modifications to the latent diffusion architecture. In that first comes the autoencoder, which has the encoder and a decoder. The autoencoders typically use four channels in their intermediate layers, thereby compressing the data and restricting their ability to represent fine details. In this work, the authors have increased the number of layers in the intermediate channels from 4 to 8 and 16. To increase the input channels further, they do a Fourier feature transform to lift the input channel dimension from 3 (RGB) to even higher dimensions. With these modifications, they tested the reconstruction of images and evaluated reconstruction scores such as FID and peak signal-to-noise ratio. We can notice that as the channels increase, these metrics show an improvement. Even visually, the quality of images seems to get far better as the number of channels increases: The next modification to the architecture is that of the U-Net used for denoising. To increase the capacity of the U-Net, they increase the channel size and the number of stacked residual blocks in each stage. And finally, they modify the decoder to output images at a high resolution of 1024 by 1024. Pre-Training With those architectural changes, they start with the first stage of training, which is pre-training. Pretraining needs a massive dataset, and so they pre-train with 1.1 billion images. But the point to note here is that they employ progressive training by gradually increasing the resolution of the image. This seems to help the model learn fine details in the images slowly but steadily. The other thing is the introduction of noise offset in the final stages of pre-training. It was introduced in which states that simply modifying the noise with a small offset enables Stable Diffusion to generate very dark or light images easily. , this blog The offset can be introduced with 1 line of code modification in PyTorch, where we add a small offset of 0.1 or similar to the noise generation process. Fine-tuning The second stage of the training is Fine-tuning. The pre-trained model obtained in the first stage is now subject to quality-tuning. But for this, we need the quality dataset in the first place. So, how do we curate such a quality dataset for quality tuning? For this, the authors of the paper employ a two-stage filtering process, namely and . automatic filtering manual filtering In automatic filtering, they use algorithms to remove images that have too many texts using the OCR recognition algorithm. They also employ CLIP to keep only results with good image-text alignment. Lastly, they also remove images with bad aspect ratio and image size and end up with a dataset of 200 thousand images which is a lot less than the billions we started with. In the next step of manual or human filtering, humans manually assess the images for aesthetic appeal. For example, this image satisfies the rule of thirds in photography so it is aesthetically appealing. Similarly, portraits taken using professional photography gear are quite appealing compared to ordinary images because the background is all blurred, creating what is called a bokeh in photography. So it’s much more aesthetically appealing compared to a selfie taken on, say, a mobile phone. The human filtering step resulted in just two thousand images. They then composed captions to those images. The image-caption pairs were then used for training the Emo model. Because there are just 2k samples, they fine-tuned for just 15K iterations with early stopping. The results are incredible, as seen in the next section. Metrics For evaluation, they were mainly concerned about the two metrics — and . They ditch the FID scores as many recent papers argue that FID scores do not correlate well with human assessment of the performance of generative models. visual appeal faithfulness Visual appeal is subjective, so the generated images were shown to 5 annotators, usually with generated images from two models shown side by side, and the annotators chose which one of the two was more appealing. For example, if we are comparing the pre-trained model with the quality-tuned model (as shown in the figure above), then these two images are shown to the annotators who are asked to choose which one of the two images is visually appealing. I would personally choose the quality-tuned one straightaway. Text faithfulness refers to the degree of similarity between a generated image and a text caption. Again, the annotators were asked to ignore the visual appeal of the images and asked to choose which ones best describe the caption with choices “A,” “B,” and “Both.” For example, the above image is for the caption, “A teddy bear working on AI research,” which is more faithful compared to “A teddy bear playing with a computer.” Results Lastly, the paper shows that quality tuning is not just restricted to the latent diffusion models but is applicable to other models like pixel diffusion and masked generative transformers. As seen from the figure above, quality tuning, when applied to these models, also improves the visual aesthetics of the model. Conclusion In my opinion, this paper is a clear eye-opener to show Emu generates results that are so good that the generated images can be used in commercial applications. how small the fine-tuning dataset can be to tune LLMs or VLMs. There are a couple of extensions of this paper, namely Emu Edit and Emu Video. I will be reviewing them in the upcoming posts. Until then, I am signing off, and I will see you in my next… till then, take care!

Meta's Emu: The Foundational Model for Emu Edit and Emu Video

Recruiting in the modern era is no easy feat. As companies strive for excellence, leaner teams, and tighter budgets, the pressure on recruiters to deliver exceptional talent has never been greater. This blog will explore some of the most common recruiting challenges that professionals face today and provide valuable tips on how to overcome them. Whether you're struggling to find the right candidates or grappling with biases in your hiring process, we've got solutions for you. Four Biggest Challenges That Recruiters Often Encounter Trouble Finding Qualified Employees One of the most pervasive challenges recruiters face is the difficulty in finding candidates who possess the skills and qualifications required for a specific role. The talent market is increasingly competitive, and identifying the right fit for your organization can be like searching for a needle in a haystack. Competition for Talent With top candidates in high demand, competition among companies for their attention is fierce. Attracting and retaining talent is a challenge, as talented individuals often have multiple job offers to choose from. Time-to-Hire Recruiters often struggle with time-to-hire metrics. The longer it takes to fill a position, the more it costs the company. Delays can result from various factors, including an inefficient hiring process, lengthy interview cycles, or excessive paperwork. Recruitment Biases Unconscious biases can seep into the hiring process, leading to unintentional discrimination. Whether it's gender, race, or age bias, it's essential to address these issues to ensure a fair and diverse workforce. The Best Recruitment Methods to Overcome Common Challenges Build a Talent Pool To overcome the challenge of finding qualified employees, build a talent pool. Continuously network, attend industry events, and maintain relationships with potential candidates. This approach can reduce time-to-hire by giving you a pool of pre-screened, interested individuals. Train Hiring Teams Invest in training your hiring teams to recognize and mitigate biases. Implementing structured interviews and standardized evaluation criteria can help reduce discrimination and ensure a fair selection process. Use AroundDeal TalentAI for Recruitment One way to many of these challenges is AroundDeal's TalentAI; this tool is one of the features in the AroundDeal Extension, powered by ChatGPT, and it works on LinkedIn. It boasts a database of over 120 million global contacts and can automate your candidate sourcing process. Here's how it can change your recruitment process: 1）Automated Candidate Sourcing can automate your candidate sourcing process, saving you time and effort. You can use it directly on the LinkedIn Jobs page. AroundDeal TalentAI 2）Analyze Job Descriptions The tool analyzes your job descriptions and requirements with a single click. It provides automatic recommendations for the most suitable candidates, streamlining the screening process. 3）Access Contact Information Once you have a shortlist of ideal candidates, AroundDeal's TalentAI enables you to access all their contact information, including email addresses, phone numbers, and company details. This eliminates the need for time-consuming searches and guesswork. 4）Speed Up the Hiring Process With AroundDeal TalentAI, finding and reaching out to potential candidates takes only seconds. This user-friendly tool simplifies the recruitment process and ensures that you're reaching out to the right candidates quickly.

How to Overcome The Biggest Recruiting Challenges？

কীভাবে সবচেয়ে বড় নিয়োগ চ্যালেঞ্জগুলি কাটিয়ে উঠবেন？

Wie meistert man die größten Recruiting-Herausforderungen?

¿Cómo superar los mayores desafíos de contratación?

Comment surmonter les plus grands défis de recrutement？

भर्ती की सबसे बड़ी चुनौतियों पर कैसे काबू पाएं？

採用における最大の課題をどう克服するか?

가장 큰 채용 문제를 극복하는 방법은 무엇입니까?

Como superar os maiores desafios de recrutamento？

Как преодолеть самые большие проблемы с рекрутингом？

En Büyük İşe Alma Zorluklarının Üstesinden Nasıl Gelinir?

Làm thế nào để vượt qua những thách thức tuyển dụng lớn nhất？

如何克服最大的招聘挑战？

Reflections on the Decline of Science in England, and on Some of Its Causes by Charles Babbage, is part of the HackerNoon Books Series. You can jump to any chapter in this book . APPENDIX, No. 3. here APPENDIX, No. 3, MEMBERS OF THE ROYAL SOCIETY, WHO HAVE CONTRIBUTED TO THE PHILOSOPHICAL TRANSACTIONS, OR HAVE BEEN ON THE COUNCIL. N. B.—The Numbers are made up to the present year for the Papers, but only to 1827 for Members of the Council. No. of     No. of
     Papers     years on
     printed    Council.
     in Phil.
     Trans.————        ———-
       3                 Aberdeen, Earl of.
       3           3     Abernethy, John.
                   2     Allan, Thomas.
       3                 Allen, William.
                   1     Arden, Lord.
                   1     Atholl, Duke of.
       7           2     Babbage, Charles,
                   1     Babington, William.
       1           2     Baily,Francis.
       9                 Barlow, Peter. (C)
                   2     Barnard, Sir F. Augusta.
                   5     Barrow, John.
       2                 Bauer, Francis.
       1                 Bayley, John.
                   1     Beaufort, Francis.
                   2     Beaufoy, Henry.
       5                 Bell, Charles.
                   1     Bingley, Robert.
                   1     Blackburne, John.
                   3     Blake, William.
       1           3     Blane, Sir Gilbert.
       1           1     Blizard, Sir William.
       1           1     Bostock, John.
      12          10     Brande, Wm. Thos. (C)
      16                 Brewster, David. (C)
       6           1     Brodie, B. Collins. (C)
       1                 Bromhead Sir E. F.
       3                 Brougham, Henry.
                   1     Browne, Henry.
                   1     Brown, Robert.
                   2     Brownlow, Earl.
       1                 Buckland, Rev. W. (C)
                   1     Burney, Rev. C. Parr.
                   1     Canterbury, Archbp. of.
                   1     Carew, Rt. Hon. R. P.
       7                 Carlisle, Sir Anthony.
                   2     Carlisle, Nicholas.
       1                 Carne, Joseph.
                   1     Carrington, Sir C. E.
                   2     Charleville, Earl of.
       7           2     Chenevix, Richard. (C)
       3           4     Children, John George.
      10                 Christie, Sam. Hunter.
                   1     Clerk, Sir George.
       2                 Clift, William.
       9                 Cloyne, Bishop of. (C)
                   2     Colby, Colonel Thomas.
                   1     Colebrooke, Henry T.
       2           2     Cooper, Sir Astley P. (C)
                   1     Crichton, Sir Alex.
                   5     Croker, John Wilson.
                   1     Cullum, Sir T. Gery.
       2                 Dalton, John.
                   2     Darnley, Earl of
       1                 Darwin, Robert Waring.
       1                 Davis, John Francis.
       2                 Davy, Edmund.
      13                 Davy, John.
       3                 Dyllwin, Lewis Weston.
       1                 Dollond, George.
                   1     Dudley and Ward, Visc.
       2                 Earle, Henry.
                   1     Egremont, Earl of.
       1                 Fallows, Rev. Fearon.
       8                 Faraday, Michael.
                   1     Farnborough, Lord.
       1                 Fisher, Rev. George.
                   1     Fly, Rev. Henry.
       2                 Foster, Henry.
       1           1     Frankland, Sir Thomas.
       1                 Gibbes, Sir Geo, Smith.
       2          13     Gilbert, Davies.
                   2     Gillies, John.
       5                 Goldingham, John.
       3           1     Gompertz, Benjamin.
                   1     Goodenough, George T.
                   2     Gordon, Sir James W.
       3                 Granville, Augustus B.
       1                 Greatorex, Thomas.
       1                 Greenough, Geo.Bellas.
       1                 Griffiths, John.
       3           1     Groombridge, Stephen.
                   1     Halford, Sir Henry.
       2                 Hall, Basil.
                   1     Hamilton, Wm. Rich.
                   2     Hardwicke, Earl of.
       2                 Harvey, George.
       1                 Harwood, J.
      16          10     Hatchett, Charles. (C)
                   1     Hawkins, John.
       2           2     Heberden, William.
       9                 Hellins, Rev. John, (C)
                   1     Henley, Morton Lord.
      10                 Henry, William. (C)
      12           6     Herschel, John F.W. (C)
                   1     Hoare, Henry Hugh
                   1     Hoare, Sir Richard Colt.
                   2     Hobhouse, Sir Benj.
       1                 Holland, Henry.
     109          16     Home, Sir Everard. (C)
       2                 Hope, Thomas Charles.
       1                 Hosack, David.
       1           1     Horsburgh, James.
       1                 Howard, Luke.
       2                 Hume, Sir Abraham.
       7           2     Ivory, James.C.
                   1     Jekyll, Joseph.
       4           1     Johnson, Jas. Rawlins.
      13           7     Kater, Capt. Henry. (C)
       2                 Kidd, John.
      24           1     Knight, Thomas A. (C)
       1           1     Konig, Charles.
                   2     Lambert, Aylmer B.
                   1     Lansdowne, Marquis of.
       1           1     Latham, John.
       2                 Lax, Rev. William.
       1                 Leach, William Elford.
                   1     Lowther, Viscount.
       2                 Macartney, James.
       2                 Macdonald, Lieut. Col.
                   1     Mac Grigor, Sir James.
                   2     Mac Leay, Alexander.
                   1     Mansfield, Earl of
       4          11     Marsden, William.
                   1     Mathias, Thomas Jas.
                   3     Maton, William George.
       1                 Miller, Lieut. Col. G.
                   2     Montagu, Matthew.
       7           4     Morgan, William.
                   1     Mount Edgecumbe, Earl of.
                   3     Murdoch, Thomas.
                   2     Nicholl, Rt. Hon. Sir J.
                   1     Norfolk, Duke of.
                   2     Ord, Craven.
       1                 Parry, Charles Henry.
                   1     Pepys, Sir Lucas.
       6           2     Pepys, Wm. Hasledine.
       7                 Philip, A. P. Wilson.
       1                 Phillips, Richard.
                   2     Pitt, William Morton.
       1          29     Planta, Joseph.
      19          17     Pond, John.  (C)
       2                 Powell, Rev. Baden.
       2                 Prinsep, James.
       4           1     Prout William.
                   1     Rackett, Rev. Thomas.
                   1     Redesdale, Lord.
                   2     Reeves, John.
       5           3     Rennell, James  (C)
       1                 Rennie, George.
       4                 Ritchie,
       1                 Robertson, James.
                   1     Rogers, Samuel.
       2           1     Roget, Peter Mark.
                   3     Rudge, Edward.
      12                 Sabine, Edward. (C)
                   2     Sabine, Joseph.
                   1     St. Aubyn, Sir John.
       3                 Scoresby, jun. William.
       2                 Scott, John Corse.
       3           1     Seppings, Sir Robert. (C)
       1                 Sewell, Sir John.
                   3     Somerset, Duke of.
                   3     Sotheby, William.
       3           2     South, James. (C)
                   5     Spencer, Earl.
                   3     Stanley, Sir John Thos.
                   3     Staunton, Sir Geo. Thos.
                   2     Stowell,Lord.
                   1     Sumner, George Holme.
       1                 Thomas, Honoratus L.
       2                 Thomson, Thomas.
       1                 Tiarks, Dr. John Lewis.
       1                 Troughton, Edward. (C)
       2                 Ure, Andrew.
                   2     Warburton, Henry.
       1                 Weaver, Thomas.
       1                 Whewell, William.
       3                 Whidbey, Joseph.
       2           3     Wilkins, Charles.
       3                 Williams, John Lloyd.
       1           1     Wilson, Sir Giffin.
                   2     Wilson, Gloucester.
                   1     Yorke, Rt. Hon. Chas. I had intended to have printed a list of those persons to whom the Royal Society had in past years awarded the Copley medals, and the reasons for which they were given; but having applied to the Council for permission to employ an amanuensis, to copy those awards, either from the minutes, or from the volumes of the Philosophical Transactions, I was surprised at receiving a refusal. I confess it appeared to me, that as a whole, those adjudications did us credit, although I doubted the propriety of many individual cases. As, however, the Council seem to have had a different opinion, and as I had made the application through courtesy, I shall decline printing a list, every individual portion of which has been already published in many ways, although the whole has never been printed in a collected form. About HackerNoon Book Series: We bring you the most important technical, scientific, and insightful public domain books. This book is part of the public domain. Charles Babbage (1998).Reflections on the Decline of Science in England, and on Some of Its Causes. Urbana, Illinois: Project Gutenberg. Retrieved https://www.gutenberg.org/cache/epub/1216/pg1216-images.html This eBook is for the use of anyone anywhere at no cost and with almost no restrictions whatsoever. You may copy it, give it away or re-use it under the terms of the Project Gutenberg License included with this eBook or online at , located at www.gutenberg.org https://www.gutenberg.org/policy/license.html

LIST OF THE MEMBERS OF THE ROYAL SOCIETY, WHO HAVE CONTRIBUTED TO THE PHILOSOPHICAL TRANSACTIONS

Stories, News & Mentions about HOPSTACK

Hopstack’s Stories on HackerNoon