Colabra makes scientific teams more efficient with real-time collaboration, integration of all data, and agile project management

Colabra

Vor kurzem erst entdeckt und für gut befunden: Googles Plattform um schnell mal Scripts zu erstellen und in Form eines Jupyter Notebooks verwalten. Alle Möglichkeiten sind unter: ersichtlich. Ausgaben über die , Nutzung von , installieren weiterer Libs über und und und. Python https://colab.research.google.com matplotlib NumPy pip Zur Nutzung benötigt man eine Einladung. Beim Aufruf des o.g. Links wird man darauf hingewiesen und kann eine Einladung anfordern. Bei mir hat das keine 24 Stunden gedauert. Einfach mal ausprobieren. Hier mal ein Beispielnotebook von mir: Webscraping mit Python und BeautifulSoup Momentan ist (leider) nur Python 2.7.x verfügbar. Für eine _Jupyter_artige Arbeit aber auf jeden Fall mehr als empfehlenswert. Besonders die nahtlose Integration in finde ich super. Man kann aus Drive heraus einfach ein neues Notebook erstellen und natürlich auch den Link dazu wie bei Google Drive für andere Dateien gewohnt, freigeben und teilen. Tolle Sache. Google Drive Google Hoffen wir mal, dass dieses Produkt weiterentwickelt und gepflegt wird. Man weiß ja, dass Google schnell mal bestehende Produkte beerdigt. Hier will Google aber die Möglichkeiten des oder auf eine einfache Plattform heben. Deshalb bin ich optimistisch, dass dieses Projekt nicht eingestampft wird. Machine Learning Deep Learning Daumen hoch dafür! Anregungen, Korrekturen und weitere Meinungen wie immer Willkommen! Stay tuned. :-) Postet on mosenturm.de

Colaboratory — Google — Research at Google — am Beispiel Webscraping mit Python

A Digestible High-Level Overview of CPU & GPU Cores

CPU এবং GPU ক্ষমতাগুলির একটি হজমযোগ্য উচ্চ-স্তরের ওভারভিউ

Ein leicht verständlicher Überblick über die CPU- und GPU-Funktionen

Una descripción general digerible de alto nivel de las capacidades de CPU y GPU

Un aperçu digeste de haut niveau des capacités du CPU et du GPU

CPU और GPU क्षमताओं का एक सुपाच्य उच्च-स्तरीय अवलोकन

CPU と GPU の機能について、わかりやすい概要を説明します。

CPU 및 GPU 기능에 대한 소화 가능한 상위 수준 개요

Uma visão geral resumida de alto nível dos recursos de CPU e GPU

Удобный общий обзор возможностей процессора и графического процессора

CPU ve GPU Yeteneklerine Sindirilebilir Üst Düzey Genel Bakış

Tổng quan cấp cao dễ hiểu về khả năng của CPU và GPU

CPU 和 GPU 功能的易懂高级概述

Large Language Models (LLMs) like GPT-4, Llama-2, and Claude have carved out a significant niche, offering unprecedented capabilities in natural language processing and generation. These models, though seemingly similar, harbor distinct characteristics that influence their interaction with our language, a phenomenon rooted in the core of their design - tokenization, embeddings, and prompting. Tokenization is the process where human language is dissected into manageable pieces, called tokens, which serve as the building blocks of comprehension for these models. Each employs a unique tokenizer, translating English or any other human language into a numerical language that the model understands. This translation is not universal; it's as diverse as the human languages we speak. A word or phrase might be tokenized differently by GPT-4 than by Llama-2, leading to varied interpretations and responses. LLM Prompting, on the other hand, is an art of communication with these models. It’s akin to having a conversation where the quality of the response is heavily influenced by the nature of the question asked. Mastering the art of prompting is essential to harness the full potential of LLMs, and it’s a skill that evolves with practice and understanding. In this video, we dive deep into the world of LLMs, offering insights into their distinct behaviors and delving into the nuances of tokenization and prompting. We explore advanced techniques and tools that enhance the efficacy of these interactions, ensuring that the conversation between humans and AI is not just coherent but also rich in context and relevance. I also share a new prompting technique that is particularly useful for summarization! Lastly, I cover the importance of parameter tuning when using those LLMs (not just when training them!), such as adapting the temperature, frequency penalty, and the other parameters influencing your outputs. Why should you watch this video? Because in the realm of AI, understanding is empowerment. This is a guided journey that equips you with the knowledge to navigate the complex terrains of LLMs effectively. It’s an exploration of the subtle, yet profound differences in how models like GPT-4, Llama-2, and Claude perceive and respond to human language. Let’s demystify the intricate dance between prompts, tokenization, embeddings, the most important model parameters, and their responses, and learn to control LLM outputs better: https://youtu.be/AOeKFlSLMOA?embedable=true&transcript=true References ►Amazing Prompting resource: learnprompting.org ►Chain of Density Prompting paper from Adams et al., https://arxiv.org/pdf/2309.04269.pdf ►Colab notebook of the tokenizer example: https://colab.research.google.com/drive/1IVQyGmj1t9R12oajT4OMjMWzwZk3jWgm?usp=sharing ►Tokenization in Machine Learning Explained: https://vaclavkosar.com/ml/Tokenization-in-Machine-Learning-Explained ►Twitter: https://twitter.com/Whats_AI ►My Newsletter (A new AI application explained weekly to your emails!): https://www.louisbouchard.ai/newsletter/ ►Support me on Patreon: https://www.patreon.com/whatsai ►Join Our AI Discord: https://discord.gg/learnaitogether

GPT-4, Llama-2, Claude: How Different Language Models React to Prompts

In a , I explored the role of efficiency in the success of a quick-commerce organization. recent piece Now, I'm going to delve further into this topic, focusing on efficiency within the commercial function, a vital component of any such organization. This topic carries significant importance as it ties directly to business metrics such as average bill, order frequency, front and back margins, and write-offs, all of which contribute to the company's PnL and are overseen by the commercial department. For this article, I'm not going to discuss the aspect of efficiency in negotiations. Instead, I'll be sharing insights into how we can uncover untapped potential in pricing, assortment management, and stock logistics with data-based instruments and processes. The description provided in the article is intended to provide a high-level overview, and I understand that each topic covered could be explored in much greater detail. If you would like more comprehensive information or detailed explanations of specific tools and approaches mentioned, I encourage you to leave a comment. This way, I can address your specific interests and provide more in-depth insights as needed. 1. Pricing By pricing here I mean the process of setting prices on SKUs, both regular prices and promo prices. Usually, within a q-commerce company with a limited assortment, prices are set manually given the expertise of commerce executives. Still, there are two approaches and tools that might greatly increase the efficiency of pricing. 1.1 Pricing Rules In today's economic climate, many countries face high inflation, necessitating frequent price revisions. This scenario presents a substantial challenge for commercial teams: Consistent price modifications may introduce the risk of inaccuracies, potentially compromising profit margins. Additionally, maintaining an accurate and up-to-date pricing level can be demanding, both in terms of time and departmental resources. The practical remedy to these challenges is pricing automation, utilizing a predefined set of rules for each SKU. Such rules could encompass maintaining a minimum profit margin for an SKU, aligning the pricing index with certain competitors, or establishing minimum/maximum prices. The frequency of price adjustments can be tailored to a company's objectives. For example, Ozon, a prominent player in the Russian marketplace, revises prices multiple times daily to ensure the most competitive rates for certain categories. However, such frequent updates may not be crucial for an average quick-commerce entity. While the solution may appear simple, it necessitates considerable collaboration from analytics, commercial, product, and development departments. The following tasks are key: : Design a data-flow encompassing competitors’ prices and purchasing costs. Analytics : Establish the pricing rules, ideally built upon a transparent pricing strategy. Commerce : Automate the entire process, including data collection, new price computation, and price updates. Product & Development 1.2. Pricing Segmentation In many developed countries, price segmentation and differentiation are often viewed as ethically questionable and, in many cases, legally restricted. However, in this piece, we won't delve into ethical debates for the following reasons: Instead of rejecting reality, we must acknowledge that price segmentation is a pervasive practice, varying in flexibility and extent across markets. Regardless of a company's stance on price segmentation, the capability to conduct pricing experiments is indispensable for understanding price elasticity and developing effective pricing strategies. Laws governing pricing modifications within the same brand of stores can range from absolute prohibition (as in Israel) to essentially full approval of pricing adjustments for different users (as seen in Russia). Stringent regulations may limit the elasticity of price segmentation but do not necessarily ban pricing experiments outright. For example, even in Israel, a retailer can form a separate brand with lower prices or conduct pricing tests using personalized discounts instead of standard prices. So, how does one implement effective price segmentation? The primary objective is to either boost orders with minimal impact on unit economics or, conversely, enhance unit economics with minimal growth disruption. In a more granular sense, the aims of this exercise are as follows: Determine the correlation between price levels and cohort retention, frequency for distinct customer segments – essentially, establish the elasticity between pricing and long-term growth. Ascertain the relationship between unit economics and price levels. There are multiple ways to design a pricing test, but here's a simplified approach: For each SKU in the assortment, calculate its importance to customers, also known as the Key Value Items (KVI) index. Create several pricing scenarios by modifying prices for top N items based on the KVI index, or proportionally, based on the KVI index: decrease, increase, substantially increase, etc. Divide users/stores (subject to legal flexibility in your jurisdiction) into multiple test groups and a control group. Ensure the number of observations is ample for identifying effects within a reasonable timeframe. Wait to identify the cohort effect (typically a few months if run by stores), then analyze variations among stores to detect patterns, such as demand elasticity depending on differing real estate prices or number of groceries, etc. Through analyzing these patterns and identified elasticities, it's possible to derive either a robust pricing strategy or an optimal price segmentation approach. 2. Assortment Management When it comes to assortment management, it's important to recognize the complexity of this process. Data analytics, while powerful, might not provide a fully automated solution, a reality that became clear to me while trying to develop an algorithm to optimally set assortment quotas: the ideal size of assortment within different categories. Despite our best efforts, this project was unsuccessful. The challenges arose from our inability to confidently estimate cannibalization curves (predicting how category sales would adjust if we added one more SKU) and integrate all pertinent market information such as agreements with suppliers, suppliers' service levels, and the quality of remaining market SKUs. Consequently, I believe that assortment can be defined properly only within a colaboration between analytics and the commercial department. Analytics in this case can assist commercial managers in 3 ways: 1. Enabling commercial with necessary tools and BI-reports to promote data-awareness and data-driven decision making Data accessibility: assisting in defining a transparrent assortment strategy Assortment Strategy: regular procedures targetted on category potential
identification Category Reviews: Let’s deep dive in each field separately. Based on this experience, I'm convinced that successful assortment management requires a collaborative approach between analytics and the commercial department. Here, analytics can play a critical role in supporting commercial managers in three areas: : Provision of necessary tools and business intelligence reports to foster data-awareness and facilitate data-driven decision-making. Data Accessibility Guidance in formulating a transparent assortment strategy. Assortment Strategy: : Routine procedures aimed at identifying category potentials. Category Reviews Let's delve into each of these areas in detail. 2.1. Data Accessibility Specific recommendations for data accessibility are challenging to prescribe, as the required tools and reports can vary significantly. At the very least, commercial managers should be able to extract various commercial metrics (Sales Quantity, GMV, Revenue, Write-Offs, OSA etc.) at a daily level, with various levels of detail (by SKUs, stores, categories etc.) readily available. In the E-Grocery companies where I've worked, this kind of dashboard was referred to as the Commercial PnL. As for other dashboards, open and clear communication between the commercial and analytics departments is critical, ensuring that analysts are responsive to the challenges and requirements faced by commercial managers. 2.2. Assortment Strategy An assortment strategy is essentially a set of guiding principles that aid category managers in managing assortment. These principles are bespoke to each category and align with the broader brand image and development trajectory of the service. The structure for such a strategy may be laid out as follows: Each category should have a defined role. This role is a strategic classification within a logical framework. Each category role should be governed by specific management principles, which are applied uniformly to all categories within that role. Different perspectives exist on how to define category roles. For instance, Walmart's framework classifies categories along two dimensions: the development status of a category and whether there's an intention to develop it further. However, I find the following approach more useful: Short description of the roles: These are the SKUs that drive customers to shop. Common examples include staple items like dairy products or fresh bread. Although these categories typically see high volumes of sales, they seldom offer a high front margin. Traffic Generators: These are the SKUs that customers often buy impulsively or without premeditation, much unlike eggs or milk. Examples might be chips, chocolates, or beer. While these categories offer significant variety, the penetration of a single SKU in an order is usually low. Despite this, Money Generators significantly contribute to the overall profit margin. Money Generators: These are the SKUs that are only occasionally used, like dishes or birthday candles, within a grocery store. Their function is to draw customers into the store over competitors or to boost the average bill. Supplementary SKUs: These are the SKUs that hold importance for both customers and retailers due to their profitability. A subset of Focus SKUs is the KVI (Key Value Items), which are items that a specific retailer is renowned for ( ). Focus SKUs: "Retailer ABC is the go-to place for fresh fish" Please note, correctly assigning category roles may vary depending on customer profiles. For instance, some customers may primarily use your service for ordering snacks, while others opt for ready-to-eat meals. Therefore, more customer-centric approaches could be explored in alignment with the overarching brand strategy. Delving into the guiding principles for each category role would indeed be quite comprehensive, meriting a discussion of its own. If you're interested in this topic, I encourage you to comment below. Still, here is how I see the overall strategy and goals for each category role: be on par with direct competitors but do not differentiate; Traffic Generators: target metric - penetration into orders or sales in items differentiate from competitors and address growing demand; Money Generators: target metric - commercial profit, abs cover the basic offer not trying to differentiate; Supplementary SKUs: target metric - penetration into orders offer the broadest and highest quality assortment, differentiate from competitors; Focus SKUs: target metric - SKU / Category retention or LTV Feel free to comment on the topic, as it seems there might be a lot to discuss. 2.3 Category Reviews A category review is a process where analytics is directly deployed to ascertain potential deficiencies and investigate growth opportunities within distinct categories. However, it's vital to note that, if mishandled, this can become a contentious issue for commercial managers. Firstly, it could potentially be perceived as an unwarranted audit of their work. Secondly, if executed improperly, it can erode the trust between the commercial and analytics teams. I won't delve into the full methodology of a category review in this discussion, but it largely concentrates on identifying uncovered customer attributes by constructing what is referred to as a Customer Decision Tree (CDT). The CDT is a systematically organized map of customer needs, each of which can be met by a range of interchangeable SKUs, referred to as customer attributes. The creation of a CDT is generally an analytical process using clustering techniques based on internal sales data. However, it can be further enhanced with insights from commercial managers or customer feedback gathered from surveys. Example of a CDT built analytically: 3. Stock Logistics Efficient stock logistics is essential to ensuring high levels of availability while minimizing write-offs and store overload. The key elements of efficient stock logistics are the following: Proper availability measurement Accurate and automized stock replenishments 3.1. Proper availability measurement Availability is a critical metric within every retail operation, influencing both immediate sales and long-term customer loyalty. Therefore, in order to accurately understand the situation and prioritize efforts aimed at enhancing availability, it's imperative for retailers to develop a robust measure of availability. This measure should align with the customer's perspective of availability, that is, it should have a strong correlation with key performance indicators such as order conversion rates or average bill. Availability is usually measured with a metric called out-of-stock availability (OSA). Still, this metric is often an inaccurate indicator of availability due to the following reasons: (1) Traditional retailers often lack real-time data on stocks, thus limiting OSA estimation to once or twice a day. This snapshot approach does not accurately represent the dynamic nature of stock availability. For instance, if you measure the OSA at 7 a.m. right after restocking, sell out all the stock by 1 p.m., and cease operations for the rest of the day, the OSA wouldn't reflect the true availability scenario. (2) OSA does not account for the fact that a retailer might have multiple similar SKUs that satisfy the same customer need (as per CDT). Take eggs as an example; If a retailer has two SKUs of eggs and one of them is sold out while the other remains, it wouldn't impact sales but would decrease the OSA. (3) OSA does not consider the varying importance of different SKUs to customers. For instance, while calculating OSA, eggs and birthday candles would be given equal weight, despite the fact that eggs are generally of greater importance to customers. Here is the logic solution for the problems listed above: (1) To overcome the challenge of time-specific OSA measurement, retailers need to have real-time access to stock data. Then, OSA should be estimated as the total number of hours each SKU was present, divided by the total number of hours each SKU should have been present (the number of active SKUs multiplied by the number of operational hours of a store). (2) SKUs should be classified using the concept of the Customer Decision Tree (CDT). For instance, if you have two bottles of milk from different suppliers that are indistinguishable for customers, these two SKUs are within the same CDT property. Measure availability by CDT properties using the same approach as in point (1). This metric is called . Group OSA (GOSA) (3) Finally, to account for the varying importance of items, weight properties by # orders. Although more sophisticated methods for measuring item importance could potentially be devised, # orders seems to accurately illustrate the relative importance of a property for customers. This resulting metric is called . Adjusted Group OSA (AGOSA) Once AGOSA has been determined, it should serve as a pivotal business KPI. A recommended practice is to establish weekly meetings involving both commercial and supply chain managers. During these sessions, the team can review customer properties with low AGOSA scores (these could be properties with a high weight or low GOSA). The team can then collaborate on strategies and activities aimed at improving AGOSA for these identified properties. 3.2. Accurate and automized stock replenishments The topic of automated stock replenishment is extensive enough to warrant a dedicated textbook. For the purposes of our discussion, however, I'll briefly outline the key elements that underpin an effective stock replenishment system: Accurate and transparent demand prediction by SKUs and stores Creating a strong stock logistics system requires accurate sales predictions for each day, SKU, and store, and these predictions should cover at least the next seven days. These predictions should be clear and open, and we should be able to easily change them if something unexpected happens or if there are mistakes. A good way to do this could be to make these predictions step-by-step. We start by estimating the total sales, then estimate the share for each category, and finally look at the share for each store and SKU. We need to remember that predicting sales is usually quite hard for SKUs that aren't sold often. In these cases, the prediction is more about the chance of an item being bought within a day, rather than a normal sales forecast. So, for retailers with a lot of different products, getting a completely accurate forecast might not be possible. Autoorder algorithm that estimates the economically optimal order Beyond sales forecasting, it's essential to develop an integrated auto-ordering algorithm. This should incorporate various factors, such as sales projections, minimum supply quantities, suppliers’ delivery schedules, etc., to calculate the optimal quantity to order from suppliers. For retailers operating with a Distribution Center (DC), the system's complexity escalates. This scenario necessitates a two-step procedure: firstly, estimating the optimal order for the DC, and secondly, determining the ideal order from the DC to individual stores. It is vital that the auto-ordering algorithm does more than just crunch numbers based on sales forecasts. A sophisticated algorithm should evaluate the economically ideal level of OSA that maximizes commercial profit and calculate the necessary order volume needed to achieve this target OSA level. To achieve this, one needs to statistically identify the multiple relationships listed below: Operational analytics responsible for demand prediction corrections Regardless of the sophistication of sales forecasting, it is crucial to have a specialized team tasked with monitoring and adjusting sales predictions and automated order suggestions. This is primarily due to instances where the system may not account for variables like public holidays, internal trials, or promotional campaigns. This dedicated team should also be equipped to manage logistical intricacies, such as preventing store overload and managing stock acceptance queues. Additionally, the handling of new SKUs, which lack historical sales data, falls under their purview, necessitating an individualized approach to their management.

Commercial Analytics

Large language models. You must’ve heard these words before. They represent a specific type of machine learning-based algorithm that understand and can generate language, a field often called or NLP. natural language processing You’ve certainly heard of the most known and powerful language model: . GPT-3 GPT-3, as I’ve described in the video covering it is able to take language, understand it and generate language in return. But be careful here; it doesn’t really understand it. In fact, it’s far from understanding. GPT-3 and other language-based models merely use what we call dictionaries of words to represent them as numbers, remember their positions in the sentence, and that’s it. Let's dive into those powerful machine learning models and try to understand what they see instead of words, called word embeddings, and how to produce them with an example provided by Cohere. Learn more in the video... References ►Read the full article: ►BERT Word Embeddings Tutorial: ►Cohere's Notebook from the code example: ►Cohere Repos focused on embeddings: ►My Newsletter (A new AI application explained weekly to your emails!): https://www.louisbouchard.ai/text-embedding/ https://mccormickml.com/2019/05/14/BERT-word-embeddings-tutorial/#why-bert-embeddings https://colab.research.google.com/github/cohere-ai/notebooks/blob/main/notebooks/Basic_Semantic_Search.ipynb https://github.com/cohere-ai/notebooks https://www.louisbouchard.ai/newsletter/ Video Transcript 0:07 language models you must have heard 0:10 these words before they represent a 0:13 specific type of machine learning 0:14 algorithms that understand and can 0:16 generate language a field often called 0:19 natural language processing or NLP 0:22 you've certainly heard of the most known 0:24 and Powerful language models like gpt3 0:26 gpt3 as I've described in the video 0:28 covering it is able to take language 0:30 understand it and generate language in 0:33 return but be careful here it doesn't 0:35 really understand it in fact it's far 0:38 from understanding gbd3 and other 0:41 language-based models merely use what we 0:44 call dictionaries of words to represent 0:46 them as numbers remember their positions 0:49 in the sentence and that's it using a 0:52 few numbers and positional numbers 0:53 called embeddings they are able to 0:55 regroup similar sentences which also 0:58 means that they are able to kind of 1:00 understand sentences by comparing them 1:02 to known sentences like our data set 1:05 it's the same process for image sentence 1:07 models that take your sentence to 1:10 generate an image they do not really 1:11 understand it but they can compare it to 1:13 similar images producing some sort of 1:16 understanding of the concepts in your 1:18 sentence in this video we will have a 1:20 look at what those powerful machine 1:22 learning models see instead of words 1:24 called word embeddings and how to 1:27 produce them with an example provided by 1:29 the sponsor of this video a great 1:31 company in the NLP field cohere which I 1:35 will talk about at the end of the video 1:36 as they have a fantastic platform for 1:39 NLP we've talked about embeddings and 1:42 gpt3 but what's the link between the two 1:44 emittings are what is seen by the models 1:47 and how they process the words we know 1:50 and why use embeddings well because as 1:53 of now machines cannot process words and 1:56 we need numbers in order to train those 1:59 large models thanks to our carefully 2:01 built data set we can use mathematics to 2:04 measure the distance between embeddings 2:06 and correct our Network based on this 2:08 distance iteratively getting our 2:10 predictions closer to the real meaning 2:12 and improving the results and meetings 2:15 are also what the models like clip 2:17 stable diffusion or Dali used to 2:19 understand sentences and generate images 2:21 this is done by comparing both images 2:24 and text in the same embedding space 2:26 meaning that the model does not 2:28 understand either text or images but it 2:31 can understand if an image is similar to 2:33 a specific text or not so if we find 2:36 enough image caption pairs we can train 2:38 a huge and Powerful model like Dali to 2:41 take a sentence embed it find its 2:43 nearest image clone and generate it in 2:46 return so machine learning with text is 2:48 all about comparing embeddings but how 2:51 do we get those embeddings we get them 2:53 using another model trained to find the 2:56 best way to generate similar embeddings 2:58 for similar sentences while keeping the 3:01 differences in meaning for similar words 3:03 compared to using a straight one for one 3:06 dictionary the sentences are usually 3:08 represented with special tokens marking 3:10 the beginning and end of our text then 3:13 as I said we have our poses from all 3:15 embeddings which indicate the position 3:17 of each word relative to each other 3:19 often using sinusoidal functions I 3:22 linked a great article about this in the 3:25 description if you'd like to learn more 3:26 finally we have our word embeddings we 3:29 start with all our words being split 3:31 into an array just like a table of words 3:34 starting now there are no longer words 3:36 they are just tokens or numbers from the 3:40 whole English dictionary you can see 3:42 here that all the words now are 3:44 represented by a number indicating where 3:46 they are in the dictionary thus having 3:49 the same number for the word Bank even 3:51 though their meaning are different in 3:53 the sentence we have now we need to add 3:56 a little bit of intelligence to that but 3:58 not too much this is done thanks to a 4:00 model trained to take this new list of 4:03 numbers and further encode it into 4:05 another list of numbers that better 4:08 represent the sentence for example it 4:10 will no longer have the same embedding 4:13 for the two words bank here this is 4:15 possible because the model used to do 4:17 that has been trained on a lot of 4:19 annotated Text data and learned to 4:21 encode similar meaning sentences next to 4:24 each other and opposite sentences far 4:27 from each other thus allowing our 4:29 embeddings to be less biased by our 4:31 choice of words then the initial simple 4:34 one for one word embedding we initially 4:37 had here's what using imagings looks 4:39 like in a very short NLP example there 4:42 are more links below to learn more about 4:44 embeddings and how to code it yourself 4:46 here we will take some Hacker News posts 4:49 and build a model label to retrieve the 4:51 most similar post of a new input 4:53 sentence to start we need a data set in 4:56 this case it is a pre-embedded set of 4:58 3000 Hacker News posts that have already 5:01 been emitted into numbers then we build 5:04 a memory saving all those embeddings for 5:07 future comparison we basically just 5:09 saved these embeddings in an efficient 5:11 way when a new query is done for example 5:13 here asking what is your most profound 5:16 life inside you can generate its 5:18 embedding using the same embedding 5:20 Network usually it is bird or a version 5:23 of it and we compare the distance 5:25 between the embedding space to all other 5:27 Hacker News posts in our memory note 5:30 that it's really important here to 5:32 always use the same network whether for 5:34 generating your data set or for querying 5:36 it as I said there is no real 5:38 intelligence here nor that it actually 5:40 understands the words it just has been 5:42 trained to embed similar sentences 5:45 nearby in the unmanning space nothing 5:47 more if you send your sentence to a 5:50 different network to generate an 5:51 embedding and compare the embedding to 5:53 the ones you had from another Network 5:55 nothing will work it will just be like 5:58 the nice people that try to talk to me 5:59 in Hebrew at eccv last week it just 6:02 wasn't in an embedding space my brain 6:04 could understand fortunately for us our 6:06 brain can learn to transfer from one 6:08 embedding space to another as I can with 6:11 French and English but it requires a lot 6:13 of work and practice and it's the same 6:16 for machines anyways coming back to our 6:18 problem we could find the most similar 6:21 posts that's pretty cool but how could 6:23 we achieve this as I mentioned it's 6:25 because of the network birth in this 6:28 case it learns to create similar 6:30 embeddings from similar sentences we can 6:32 even visualize it in two Dimensions like 6:35 this where you can see how two similar 6:37 points represent similar subjects you 6:39 can do many other things once you have 6:41 those embeddings like extracting 6:43 keywords performing a semantic search 6:45 doing sentiment analysis or even 6:47 generating images as we said and 6:49 demonstrated in previous videos I have a 6:52 lot of videos covering those and listed 6:55 a few interesting notebooks to learn to 6:57 play with encodings thanks to the cohere 6:59 team now let me talk a little bit about 7:02 kohilu as they are highly relevant to 7:05 this video cook here provides a 7:07 everything you need if you are working 7:09 in the NLP field including a super 7:11 simple way to use embedding models in 7:14 your application literally with just an 7:16 API call you can embed the text without 7:18 knowing anything about how the embedding 7:21 models work the API does it for you in 7:23 the background here you can see the 7:25 semantic search notebook that uses 7:27 cohere API to create embeddings of an 7:30 archive of questions and question 7:32 queries to later perform search of 7:34 similar questions using cook here you 7:37 can easily do anything related to text 7:39 generate categorize and organize at 7:42 pretty much any scale you can integrate 7:44 large language models trained on 7:46 billions of words with a few lines of 7:48 code and it works in any Library you 7:51 don't even need machine learning skills 7:53 to get started they even have learning 7:55 resources like the recent cohere for 7:57 ai's colors program that I really like 8:00 this program is an incredible 8:01 opportunity for emerging talent in NLP 8:04 research around the world if selected 8:06 you will will work alongside their team 8:08 and have access to a large-scale 8:10 experimental framework and cohere 8:12 experts which is pretty cool I also 8:15 invite you to join their great Discord 8:17 Community ingeniously called Co Unity I 8:21 hope you've enjoyed this video and will 8:23 try out cohere for yourself with the 8:25 first link below I am sure you will 8:27 benefit from it thank you very much for 8:29 watching the whole video and thanks to 8:31 anyone supporting my work by leaving a 8:33 like comment or trying out our sponsors 8:36 that I carefully select for these videos

Text Embedding Explained: How AI Understands Words

Incrustación de texto explicada: cómo la IA entiende las palabras

Explication de l'intégration de texte : comment l'IA comprend les mots

पाठ एंबेडिंग समझाया: एआई शब्दों को कैसे समझता है

テキスト埋め込みの説明: AI が単語を理解する方法

Explicação da incorporação de texto: como a IA entende as palavras

Giải thích về nhúng văn bản: Cách AI hiểu các từ

文本嵌入解释：AI 如何理解单词

Have you ever dreamed of a good transcription tool that would accurately understand what you say and write it down? Not like the automatic YouTube translation tools… I mean, they are good but far from perfect. Just try it out and turn the feature on for the video, and you’ll see what I’m talking about. Luckily, OpenAI just released and open-sourced a pretty powerful AI model just for that: Whisper. It understands stuff I can’t even comprehend, not being a native English speaker (listen in the video) and it works for language translation too! Learn more in the video below... References ►Read the full article: ► Radford, A., Kim, J.W., Xu, T., Brockman, G., McLeavey, C. and Sutskever, I., Robust Speech Recognition via Large-Scale Weak Supervision. ►Project link: ►Code: ►Google Colab notebook: ►YouTube Whisperer app: ►My Newsletter (A new AI application explained weekly to your emails!): https://www.louisbouchard.ai/whisper/ https://openai.com/blog/whisper/ https://github.com/openai/whisper https://colab.research.google.com/github/openai/whisper/blob/master/notebooks/LibriSpeech.ipynb https://huggingface.co/spaces/jeffistyping/Youtube-Whisperer https://www.louisbouchard.ai/newsletter/ Video Transcript 0:00 have you ever dreamed of a good 0:01 transcription tool that will accurately 0:03 understand what you see and write it 0:05 down not like the automatic YouTube 0:07 translation tools I mean they are good 0:09 but far from perfect just try it out and 0:12 turn the feature on for this video and 0:14 you'll see what I'm talking about well 0:16 open AI just released an open sourced 0:18 and a pretty powerful AI model just for 0:21 that whisper it even understands stuff I 0:24 can't even comprehend not being a native 0:26 English speaker 0:28 this is the micro machine man presenting 0:29 the most miniature motorcade of 0:30 micro machine each one has dramatic 0:32 details terrific trim Precision page Art 0:33 Plus incredible micro machine pocket 0:34 place that says a PlayStation fire it 0:36 works for language translation too 0:38 whisper a consistent the liquid it 0:41 sounds automatically alternator 0:48 the results and precision are incredible 0:51 but what's even cooler is how it works 0:53 let's dive into it but first let me 0:56 introduce this episode's sponsor that is 0:58 highly related to This research assembly 1:00 AI assembly AI is the API platform for 1:03 state-of-the-art AI models from startups 1:06 to Fortune 500 companies developers and 1:08 product teams around the world leverage 1:10 assembly AI to build better ai-based 1:13 products and features if you are 1:15 building a meeting summarizer podcast 1:17 analyzer or really anything related to 1:19 audio or video and want to leverage AI 1:22 to power transcription or insights at 1:24 scale definitely check out their API 1:26 platform more specifically I wanted to 1:29 share their summarization model which I 1:31 find really cool as the name says with 1:34 this model you can build tools that 1:36 automatically summarize your audio and 1:38 video files the model is flexible to fit 1:41 your use case and can be customized to 1:44 different summary types bullets 1:46 paragraphs headlines or adjust it all 1:48 Works through simple API calls and you 1:51 can find all the information you need 1:53 for the summarization model and assembly 1:55 AI with the first link below 1:59 when it comes to the model itself 2:01 whisper is pretty classic it is built on 2:04 the Transformer architecture stacking 2:06 encoder blocks and decoder blocks with 2:08 the attention mechanism propagating 2:10 information between both it will take 2:13 the audio recording split it into 30 2:16 second chunks and process them one by 2:18 one for each 30 second recordings it 2:21 will encode the audio using the encoder 2:23 section and save the position of each 2:25 word said and leverage this encoded 2:28 information to find what was said using 2:30 the decoder the decoder will predict 2:33 what we call tokens from all this 2:34 information which are basically each 2:37 words being said then it will repeat 2:39 this process for the next word using all 2:41 the same information as well as the 2:43 predicted previous word helping it to 2:46 guess the next one that will make more 2:48 sense as I said the overall architecture 2:50 is a classic encoder and decoder and I 2:53 covered it in multiple videos similar to 2:55 gpt3 and other language models which I 2:58 invite you to check for more 3:00 architectural details this works as it 3:02 was trained on more than 600 000 hours 3:05 of multilingual and multitask supervised 3:08 data collected from the web meaning that 3:11 they trained their Audio model in a 3:12 similar way as gpt3 with data available 3:15 on the internet making it a large and 3:18 general Audio model it also makes the 3:20 model way more robust than others in 3:23 fact they mentioned that whisper 3:24 approaches human level robustness due to 3:27 being trained on such a diverse set of 3:29 data ranging from Clips TED Talks 3:32 podcasts interviews and more which all 3:34 represent real world-like data with some 3:36 of them transcribed using machine 3:38 learning based models and not humans 3:40 using such imperfect data certainly 3:43 reduces the possible Precision but I 3:45 will argue it helps for robustness when 3:47 used as sparsely compared to Pure human 3:49 curated audio data sets with perfect 3:52 transcriptions having such a general 3:54 model isn't very powerful in itself as 3:57 it will be beaten at most tasks by 3:58 smaller and more specific models adapted 4:01 to the task at hand but it has other 4:03 benefits you can use this kind of 4:05 pre-trained models and fine-tune them on 4:08 your task meaning that you will take 4:10 this powerful model and retrain a part 4:13 of it or the entire thing with your own 4:15 data this technique has been shown to 4:17 produce much better models than starting 4:19 training from scratch with your data and 4:21 what's even cooler is that openai open 4:24 sourced their code and everything 4:25 instead of an API so you can use whisper 4:28 as a pre-trained foundation architecture 4:30 to build upon and create more powerful 4:33 models for yourself some people have 4:35 already released the tools like the 4:37 YouTube Whisperer unhugging face by Jeff 4:39 is typing taking a YouTube link and 4:42 generating transcriptions which I found 4:44 thanks to Yannick kilter they also 4:46 released a Google collab notebook to 4:48 play with right away while something 4:50 competition is key I'm glad openai is 4:53 releasing some of its work to the public 4:54 I'm convinced such collaborations are 4:57 the best way to advance in our field let 5:00 me know what you think if you'd like to 5:01 see more public releases of openai or if 5:04 you like the final products they build 5:06 like dally as always you can find more 5:08 information about whisper in the paper 5:11 and code linked below and I hope you've 5:13 enjoyed this video I will see you next 5:15 week with another amazing paper

What is OpenAI's Whisper Model?

¿Qué es el modelo de susurro de OpenAI?

Qu'est-ce que le modèle Whisper d'OpenAI ?

OpenAI का व्हिस्पर मॉडल क्या है?

OpenAI のウィスパー モデルとは?

O que é o modelo Whisper da OpenAI?

Mô hình thì thầm của OpenAI là gì?

OpenAI 的耳语模型是什么？

Smart contract oracles protect DeFi protocols from price manipulation attacks: make the protocols safe, and are their most significant dependency simultaneously. Thus, choosing an oracle is a critical decision in the design of a protocol. StarkNet   is a ZK-rollup on the Ethereum chain that solves the “scaling” problem of Ethereum in a highly-efficient manner. The network was released in November 2021, and in a short life of around 10 months, it has given birth to multiple projects from the DeFi space, NFTs, and even gaming (a whole list of projects building on StarkNet can be found ). Though the scaling problem of Ethereum can probably be solved with L2 solutions built on top of the secure layer of the Ethereum mainnet, the problem of permissionlessly getting off-chain data on chain persists. here The two largest oracles building on StarkNet are and the . I had a conversation with Jonas Nelle, the co-founder of Empiric, and Vlad, founder and CEO of Stork, on their view of how these oracles are going to provide accurate data on-chain and the immense possibilities that have now been unlocked with StarkNet’s cheap computation (and a bit expensive storage cost). Empiric Network Stork Oracle Network Currently available data Right now, both Empiric and Stork provide price data for some of the most prominent cryptocurrencies (BTC, ETH, DAI, etc). Along with that, Empiric Network has plans to make data like weather data in a specific place or the outcome of a sports match available. Meanwhile, the Stork team is also interested in equity prices (i.e. stocks), NFTs, and other asset classes. Once these are added, non-price-related feeds might also get added by Stork as well. The wide diversity of information made available by these oracles is very helpful for the ecosystem. Since only price data is available as of this writing, I set out to do an exploratory data analysis on the data sent by these oracles on the Starknet testnet and compare the quality (and quantity) for both. Stork Some basic facts to note are: Stork is an oracle with on-chain price computation. This means that median prices are calculated based on price quotes provided by individual publishers rather than being aggregated off-chain on StarkNet Allowlisted publishers can publish and sign data on-chain Right now 2 publishers are live publishing price updates on the testnet (Dexterity - a financial firm in crypto - and the other is Stork) Has 5 price tickers available (a more updated list can be found ) here The median time between price updates was 180 seconds Now, let’s get to the exciting part - the numbers . I analyzed the price published for BTC and ETH on the Stork contract between July 31, 2022, and August 12, 2022. 🤓 As we can see from the graph above, the price range for BTC in the given timeframe was $2403 (i.e. there are no discrepancies in the range of data). However, in some cases, there are large intervals between updates (up to 4 hours) according to the dataset. The dataset was an offline record of transactions that were sent as transactions to the network. Since this record is not supposed to be saved by Stork (there is no use case for this), the data isn’t supposed to be reliable. Hence, I analyzed the on-chain frequency of transactions on the contract that was shared by Vlad. The on-chain transactions revealed a more optimistic observation. The longest time difference between two updates was 3000 seconds, and these updates probably coincided with the downtimes of the StarkNet network itself. Most of the updates were in the range of 2-3 minutes. A distribution plot of the differences can be found below: The Stork team is actively adding new publishers in the coming weeks as of this writing. I believe that once more publishers are added, the updates will easily become more frequent, and the downtimes will not be that long. Empiric Network The Empiric oracle seems more mature and provides 20+ price feeds, including both spot and options (full list here: ). Their codebase has been open source since day 1, and I was told they had just successfully completed their first audit. Though let us take a closer look at the data. A few facts first: https://docs.empiric.network/using-empiric/supported-assets Allowlisted publishers can publish data on-chain (Empiric Network, CMT, Gemini, FTX, CEX, BitStamp, Coinbase, Kraken, Binance, Bitfinex, Coingecko) Has 17 price tickers available (list ) here The median time between two price updates for the empiric publisher was 93 seconds The median time between two price updates when considering all publishers was less than 5 seconds We got the following results upon analyzing Empiric for the price of BTC between July 31, 2022, and August 12, 2022. The price range for BTC was 2468, and the maximum downtime was for 7 minutes, with most of the updates taking less than one minute (see stats below) It might be surprising that some updates had less than 5 seconds of gap between them. This is probably due to the number of publishers sending the transactions with the updated prices. Empiric’s cofounder, Jonas, mentioned during our interaction that specific mechanisms could be implemented to explicitly favor timing coordination among publishers, and the updates may remain at small intervals. Stork can indeed implement this mechanism in the future once they have a more extensive network of publishers. Since the price updates for Empiric also mentioned the source of the data, I looked at the actual closeness of the data to real-time prices (from the CoinGecko API) at 5-minute intervals. An example of the data retrieved from Empiric is given below for a clearer understanding. The is the actual source of price (market-maker, AMM, etc.) of the asset, while is the entity that sends the transaction on-chain. source the publisher By taking only the updates from set as publisher and as the source, we saw that the price was within -0.002 % of the price returned by CoinGecko API at 5-minute intervals over around 24 hours. I could not run a similar analysis on the Stork data since the data we received from Stork did not have the sources of data segregated and instead had an aggregated price feed. the empiric coingecko It is very important to understand that a larger number of publishers on the Empiric network gives it the upside of having more frequent updates published. The bottleneck, however, is the block frequency of the StarkNet network itself. The data we analyzed above is the history of transactions sent and the corresponding timestamp of the transaction being sent. However, they get added to a block whenever a new block is created and published (which takes around 3 mins today). By taking only the updates from set as publisher and as the source, we saw that the price was within -0.002 % of the price returned by CoinGecko API at 5-minute intervals over around 24 hours. I could not run a similar analysis on the Stork data since the data we received from Stork did not have the sources of data segregated and instead had an aggregated price feed. the empiric coingecko It is very important to understand that a larger number of publishers on the Empiric network gives it the upside of having more frequent updates published. The bottleneck, however, is the block frequency of the StarkNet network itself. The data we analyzed above is the history of transactions sent and the corresponding timestamp of the transaction being sent. However, they get added to a block whenever a new block is created and published (which takes around 3 mins today). How “permissionless” are current oracles? As mentioned in this post, Stork and Empiric only enable allowlisted publishers (mostly high-volume market makers and exchanges) to publish the latest price to the respective smart contracts. I am sure both teams are working on a truly permissionless way of proving the price of an asset. However, right now, the centralized nature of the data being published is to be expected from projects at a nascent stage. Empiric recently the in collaboration with Coinbase and OKX. This allows any user to publish the price data received from their APIs with the exchanges’ signatures. As a result, even though the data source remains permissioned (and multiple publishers might use the same source), these are positive steps toward a completely decentralized oracle system. ported open oracle standard The most prominent example available today is Chainlink. Anyone is allowed to stake some assets, spin up a node, and start sharing their knowledge about the off-chain (IRL) world. However, only certain entities can publish data to the official Chainlink price feeds ( ) - which are used mainly by Defi protocols in the space. The approach that Chainlink has developed makes sense because entities with large market caps have a reputation to uphold and minimal upside to gain from (if any) by publishing dishonest data on the chain. On the other side, since anyone can spin up a node, it is the responsibility of the node maintainer to increase their outreach and build a historical reputation for being a high-quality data provider. here Moreover, since both Empiric and Stork oracles are decentralized in the way that the originator of the price update transaction always is the publisher (who has the private key to their public wallet), any malicious attempts to skew the price reflected by the oracle will only result in reputation loss. Any oracles on StarkNet can implement simple aggregation modes like the median of most recent updates to prevent a single source from skewing the data. Both Stork and Empiric employ these and similar techniques. Will the updates remain as frequent when moving to mainnet? It is a big “if“ to assume that the updates remain just as frequent once these oracles move to the mainnet. On StarkNet, the computation cost is infinitesimal, yet the storage cost is relatively high. Hence any updates to the storage of the network may cost higher than expected once moved to the mainnet. However, another interesting fact is that overwriting the same storage slot multiple times in a single block will not result in high storage costs. Hence you only pay storage fees while updating the value - if it stays the same, you pay nothing. The cost would still be of a single write since the final value of the slot is sent to Ethereum Mainnet as calldata. Due to these reasons, StarkNet may issue batching rebates in the future for multiple overwrites of the same storage slot. Computational feeds and other possibilities for future research As the computational effort required to verify STARK proofs is compared to the computation proven, StarkNet has the potential to scale Ethereum by orders of magnitude; this gives rise to new possibilities like running live computational feeds that can aggregate data provided by oracles, calculating yield curves, and possibly even modeling them on-chain. o, a fixed rate loan protocol, is already working on an integration with Empiric's data feeds to price their fixed rate loans on StarkNet. Right now, CurveZero is live on testnet, and you can take out a test fixed yield loan on their . Similarly. Projects like can use a wide source of information from oracles on Starknet and aggregate the feeds to make decisions for depositing their users' liquidity at the highest yields possible. Yagi is integrating with Empiric and will be utilizing their volatility feeds. exponentially small CurveZer website Yagi Finance Yet another approach for further improvement is the possibility of moving to . Since L3s build on top of L2s, similar to how L2s build on top of L1s, publishing data will become even cheaper and more accessible, and multiple transactions will be provably posted on the L2 (unless there is a verifier contract on it). Depending on the use case, data availability may be tuned, and the recursive proofs may be employed to power the progress of this ecosystem. L3s Resources The link to Colab notebooks containing all the results of DA mentioned in the blog post can be found below: Stork evaluation: https://colab.research.google.com/drive/1-K1t4BCAXdVdliI3Dby2jyT84ulLdX44?usp=sharing Empiric evaluation: https://colab.research.google.com/drive/1Bkct5khBkWAp-BvOZnu6VFDvMtRI1Tqz?usp=sharing Empiric Evaluation against CoinGecko: https://colab.research.google.com/drive/17-ilkX0mgWwo9tHUCVmSypb6DrWILmnp?usp=sharing

Oracles on StarkNet

70-Page Report on the COCO Dataset and Object Detection [Part 2]

COLABRA

colabra.app

-- employees

Since n.d.

COMPANY RANKING

COLABRA

EVERGREEN INDEX #2685

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

Colabra WIKI

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

COLABRA

colabra.app

-- employees

Since n.d.

COMPANY RANKING

COLABRA

EVERGREEN INDEX #2685

RELATED COMPANIESRANK

STORIES ABOUT COLABRA

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

Colabra WIKI

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