DataGPT is a conversational AI data analytics software provider that delivers analysis at the speed of business questions. Our software empowers anyone, in any company, to talk directly to their data using everyday language, revealing expert answers to complex questions instantly. DataGPT is pioneering the modern era of data analysis. Our mission is to empower every person, in every role, in any industry, to use data to make decisions, big and small. Our flagship solution, the DataGPT AI Analyst, is more than just an analytics tool - it communicates in natural language and iterates with you to dig deeper for hidden answers to complex questions buried in large datasets. By combining the creative, comprehension-rich side of a large language model (LLM) (the “right brain”) with the logic and reasoning of advanced analytics techniques (the “left brain”), DataGPT’s conversational AI data analytics is your always-on partner, helping every user drive business outcomes. Companies like Mino Games, Plex and Wombo use DataGPT everyday to increase revenue and save time. 𝘋𝘢𝘵𝘢𝘎𝘗𝘛 𝘪𝘴 𝘢 𝘱𝘳𝘰𝘥𝘶𝘤𝘵 𝘥𝘦𝘷𝘦𝘭𝘰𝘱𝘦𝘥 𝘢𝘯𝘥 𝘰𝘸𝘯𝘦𝘥 𝘣𝘺 𝘊𝘰𝘮𝘱𝘢𝘳𝘢𝘵𝘪𝘷𝘦, 𝘐𝘯𝘤.

DataGPT

100 Complex LLM Terminology Explained in One Single & One Simple Sentence

The New York Times Company v. OpenAI Update Court Filing, retrieved on February 26, 2024 is part of . You can jump to any part in this filing . This part is 3 of 15. HackerNoon’s Legal PDF Series here B. The Key to Generalist Language Models: Scale So OpenAI’s research took the logical next step: “scaling up [] model size” by increasing both the “size” and “diversity” of the training data. See GPT-3 Paper at 6. To build its next generation of models, OpenAI’s researchers gathered a more robust set of data in part by “expand[ing]” WebText database into a new version called “WebText2,” which included material shared over a longer period of time. Id. at 8–9. They also used a filtered version of Common Crawl, a repository of data collected by a non-profit research organization representing “a copy of the Internet.” Compl. ¶ 88. OpenAI disclosed all of this no later than July 22, 2020. GPT-3 Paper at 8. At the time, it was common knowledge that WebText2 and Common Crawl included numerous articles published by the Times.[19] The result of this simple act of “scaling up” the training data was, as the Times reported at the time, “mind blowing.”[20] The new GPT-3 model was “by far the most powerful ‘language model’ ever created.” Manjoo, supra note 20. It could conduct “on-the-fly reasoning” and “unscrambl[e] words, perform[] arithmetic, and us[e] novel words in a sentence after seeing them defined only once.” GPT-3 Paper at 5. Increasing the scale of training led to a surprising jump in ability. See, e.g., id. at 22 (Figure 3.10). According to the Times’s reporting, GPT-3 showed that “[m]achines are gaining the ability to write.” Manjoo, supra note 20. Within days, developers began to use it to build unprecedented tools. Id. (“service that responds to email on your behalf”). The “key advance,” as the Times reported, was “GPT-3’s flexibility.” Id. And the reason it was flexible was OpenAI’s decision to “scal[e] up” the “size” of the training data. GPT-3 Paper at 6. The amount of data needed was staggering. Compl. ¶ 85. But it was that “unprecedented scale” that allowed the model to internalize not only a “map of human language,” but achieve a level of adaptability—and “emergent” intelligence—that “no one thought possible.”[21] Continue Reading . Here [19] Compl. ¶ 88 (citing a “2019 snapshot of Common Crawl”); GPT-2 Model Card (noting prevalence of Times content in WebText); GPT-3 Paper at 8 (noting WebText2 is an “expanded version of the WebText dataset”). [20] Farhad Manjoo, How Do You Know a Human Wrote This?, N.Y. Times (July 29, 2020), https://www.nytimes.com/2020/07/29/opinion/gpt-3-ai-automation.html. The Court may take judicial notice of this and other news articles cited in this Motion. In re UBS Auction Rate Sec. Litig., No. 08-cv-2967, 2010 WL 2541166, at *10-12 (S.D.N.Y. June 10, 2010) (judicial notice of “several news items” including Wall Street Journal articles). [21] Cade Metz, Meet GPT-3. It Has Learned to Code (and Blog and Argue), N.Y. Times (Nov. 24, 2020), https://www.nytimes.com/2020/11/24/science/artificial-intelligence-ai-gpt3.html. About HackerNoon Legal PDF Series: We bring you the most important technical and insightful public domain court case filings. This court case retrieved on February 26, 2024, from is part of the public domain. The court-created documents are works of the federal government, and under copyright law, are automatically placed in the public domain and may be shared without legal restriction. fingfx.thomsonreuters.com

ChatGPT's Predecessor Was Trained on 'Filtered' Internet Data from an Open Source Repository

Introducing the GPT OpenData Explorer The past few days after the from OpenAI that users will be able to create custom versions of ChatGPT or GPTs, I decided to do an experiment creating one myself. Something that has always been bugging me is the difficulty of finding information of high quality when searching most topics on the web regarding Greece or, frankly, any search using the Greek language, especially compared to English search results. The aforementioned, combined with the generally always tricky task of getting information from trustworthy sources, led to the creation of the** **. This GPT is mainly powered by the Open Data published by the Greek central government, local authorities, and public bodies at . announcement OpenData Explorer GPT https://data.gov.gr OpenData Explorer GPT The big advantage that an chatbot like this GPT can provide is natural language understanding and how that can . LLM simplify the process of people searching for information Journalists, students, or anyone else can access (and soon properly visualize) data regarding education, the environment, health, transport, crime, etc. By making open data more accessible, we can increase their impact on society. A few examples: • Increase transparency and accountability: Users can check the trend of how many people take public transport and compare it to the available fleet; they can monitor the quality of broadband internet • • OpenData Explorer currently supports only data that derive from , but if positive results come out of this attempt, the goal would be to integrate it with open data from other countries as well. This would enable intriguing combinational prompts and provide access to a much greater audience. It's still worth mentioning, like any other piece of technology, but even more so with such a new one, it is still very crucial to check the validity of any AI-generated response. If you've read this far, the fun visual part comes now. Below, I will share some of the example prompts you can give the OpenData Explorer GPT and compare it to plain ChatGPT. I would love people to try it and test out its potential and, of course, receive feedback about it. Raise awareness on environmental issues: Users can check the number of fires, the electricity consumption, and the renewable energy sources used all over the country Social impact: Users can observe unemployment trends, crime rates, drunk driving accidents https://data.gov.gr Let me know what you think! As a disclaimer for this GPT: for a lot of dates, there is no data GPT is limited to the available datasets from data.gov.gr it's not good yet at doing basic operations with the results (e.g., adding them) if you don't guide it properly graph creation is not the best, either some of the responses are too big for it to handle Open Data If you are interested in Open Data, below are a few observations and interesting points that I discovered during my investigation of Open Data in Greece. According to https://data.europa.eu Open (Government) Data refers to the information collected, produced or paid for by the public bodies (also referred to as Public Sector Information) and made freely available for re-use for any purpose. It is worth mentioning that the quality of this data is very important, and a key part of that is providing a machine-readable friendly format (e.g., via API technologies). For : https://data.gov.gr • the API access option to the data is very useful • documentation and instructions on how to use the API are lacking • there are only 52 datasets available • a lot of the datasets are not updated (e.g., data for forest wildfires stops on 2018-12-31 or electricity consumption data is provided only for a few days of 2020) • on the site, you can find the number of people who have used the API, and I was surprised to see that OpenData Explorer was, in a lot of cases, the first to make requests to it Greek Open Data: • there is more available open data from Greek organizations, but it's not in , and in a lot of cases, it is in non-friendly formats (such as PDFs) • Greece ranked 28th (out of 35) in Open Data Maturity in 2022, according to • according to the "Greece 2.0" plan, a total budget of 5.227.840,00 euros will be invested, having started from 2023 and until 2025, to the project of "Open data and enhancing participation" see If you want to learn more about the state of open data in Greece, I highly recommend reading* *, an article by . data.gov.gr https://data.europa.eu/en/publications/open-data-maturity/2022#country-overview https://greece20.gov.gr/wp-content/uploads/654.SUB5_ANOIKTA_DEDOMENA_EDYTE_5190853.pdf The slow death of open data in Greece Eliza Triantafyllou I'm also following the initiatives of , which contributes to the development of Open Technologies in Greece. ΕΕΛ/ΛΑΚ GFOSS

OpenData Explorer GPT: Unlocking Information through AI

This is the second blog post in a 3-part series where I explain how my team and I created GPT Pilot – the AI coding agent that’s designed to work at scale and build production-ready apps with a developer’s help. of this series, I discussed the high-level overview of GPT Pilot. The idea is that AI can now do 95% of all coding that we, developers, are doing. See in 2 hours, which would usually take 20-30 developer hours. However, an app is of no use if it doesn’t fully work or solve the user’s problem. So, until real AGI arrives, you need a developer. In part #1 how I used ChatGPT to code out an entire Redis proxy So, this is how GPT Pilot came to life. It is designed to do 95% of the required coding and asks developers for reviews, such as when it becomes stuck and cannot move forward or needs something outside the app, like an API key. In this post, I walk you through the entire process GPT Pilot goes through when coding an app. I share diagrams to provide a visual representation of everything that’s going on behind the scenes in GPT Pilot. I’m a visual person, so I always create diagrams. To understand how GPT Pilot’s coding works, there are three concepts – context rewinding, recursive conversations, and TDD. See my introduction, which I described in them in part #1 of this series. The GPT Pilot coding workflow contains 5 steps: Take the next development task in line Break down the task into development steps Take the next development step Fetching of currently implemented code Write code for the current step Run the code or a command Test the new code changes Debug the development step or go to the next step Coding workflow is my favorite part of GPT Pilot, so let’s dive in. Here is a diagram of how it looks visually: #1 Task breakdown Two important concepts will be mentioned throughout this blog post – and . development tasks development steps GPT Pilot works in a way that, after breaking down the specifications for developing an app, it creates development tasks that will lead to a fully working app. are basically high-level descriptions of what needs to be done that a developer will take and start implementing. Think of them as tasks in Jira (btw, I hate Jira…not sure if anyone relates, but I just wanted to let it out of my system). Here is an example of a development task: Development tasks In the diagram above, you see 3 task properties: : what needs to be implemented to fulfill this task description : how can the lead developer determine if the current task is finished – a crucial pillar of GPT Pilot is a developer must be involved throughout the coding process so that you can ensure the development process is going as planned and understand the codebase along the way user_review_goal : the kind of automated test GPT Pilot should write to test if this entire development task works as expected. After a development , GPT Pilot writes unit tests, and after a development , it writes integration or E2E tests. programmatic_goal step task Now, when you start developing a task from Jira ( ), you will split it into smaller chunks (we call them ) that are actionable items you would set out to implement into the codebase. Each development step can be one of the following: development task development steps – a command that needs to be run on the machine, such as a command to install a dependency, start the app to check if the previously implemented steps work, or create a folder. Command run – the most important development step that explains what exactly needs to be implemented in the actual code to fulfill the current step. It can contain new code that needs to be written or code that needs to be changed. The way it works is the code change is a detailed, human-readable description of what needs to be implemented. It contains both the code that needs to be implemented and the description of what it is being used for. This is very similar to when you ask to code something. It will give you the code as well as the explanations of why it wrote that code. Code change ChatGPT The reason for this is that code implementation is not so simple. Sometimes, we need to add a snippet to existing code or change the existing implementation. That is why we separated the outline of the coding change (this development task) and the actual implementation of this change that the CodeMonkey agent is dedicated to. I will go deeper into that in the #3 Coding section. Here is an example of a code change: – a development step that AI cannot do by itself and needs human help in fulfilling the step. Then, GPT Pilot asks the developer to do something, and when he/she is done, they write “continue,” and GPT Pilot will continue with the implementation. Here are some reasons why human intervention might be needed: Human intervention There is a needed API key (e.g., Twitter API key to fetch data from Twitter) GPT Pilot became stuck in a debugging process, and it either filled the entire context length or the recursion conversation was too deep and became unproductive to continue down the recursion depth. GPT Pilot needs a verification if something works as expected – e.g., GPT Pilot is not sure if Mongo is installed properly on the machine and might ask the developer to run some sudo commands and see if it works as expected. #2 Fetching of currently implemented code It’s easy for AI to write a new file that contains code, but in reality, that is rarely the case. For the most part, we write into the existing files and either change the existing code or add new code. Now, AI can do this easily if you give it all of the existing code and instructions for what needs to be implemented. The problem arises when an app scales and the codebase becomes so large that it cannot fit into the LLM context. And this is actually a very common case – at least until we have LLMs with 1M tokens, which doesn’t seem to be coming soon. When you work on a task in a big codebase, you usually look at a smaller part of the codebase (maybe 1,000 lines) and work only with that subset of code to implement the task. So, to address this issue and make GPT Pilot truly scalable so that it can create and upgrade large production-ready codebases, we must create a way for the AI to select the smaller part of the codebase (e.g., those 1,000 lines) on which it will implement the current task. Once it’s finished, we can simply add the finished lines back into the original codebase. Let me start explaining this by telling you what happens when GPT Pilot writes code and creates new files and folders. For each file and folder it must create, it needs to write a description of what the idea is behind the file or folder it wants to create. For example, it might want to create a folder utils for which it will write: Contains utility modules that provide generic, reusable solutions to common problems encountered throughout the application.
These utilities are not specific to the app's core domain but offer auxiliary functionality to support and streamline the primary codebase.
They encapsulate best practices, reduce code repetition, and make the overall code cleaner and easier to maintain. Examples include functions for data formatting, error handling, debugging tools, string manipulation, data validation, and other shared operations that don't fit within specific modules or components of the app. Now, for each function, GPT Pilot creates, it writes a description of what the function is supposed to do – that is a pseudocode for the entire codebase. Now that you know what happens when GPT Pilot writes code, you can understand how it fetches the relevant code for each development step. Before GPT Pilot codes each step, it first fetches the relevant part of the codebase in a completely separate LLM conversation. That conversation is set up in 3 steps. AI is given the development step description along with the entire project file/folder structure and descriptions for each file and folder. From this, LLM tells us which files are relevant for the mentioned step. After narrowing down the necessary files, we give the LLM pseudocode for each file it listed and ask it to tell us which functions are relevant for the current development step. Once we know the pseudocode it selected, we can fetch the actual code and put it into the original conversation, where the LLM will write the description of what needs to be implemented. If the app becomes extremely huge, we can improve this by first giving the LLM the folders, from which it will select folders, and then we give it relevant files. Before each of these steps, we can also rewind the conversation to the beginning to leave more room in the context. Here is a diagram of what this looks like: #3 Coding Now that we can that contains all the code necessary for someone to implement a specific task, we can start with the actual coding process. This happens in a 2-part process: create an LLM message First, the LLM writes the description of what needs to be implemented along with the code. If the entire file needs to be coded, LLM’s response will contain all the code, but if only a part of the code inside a file needs to be changed, LLM will tell us things, like after the Mongo setup, add the following lines of code… As you can imagine, by this being stochastic rather than deterministic, we need to ensure that the written code is inserted into appropriate places or changed correctly. Here is where the agent steps in. It is called code monkey because it doesn’t make any decisions but rather simply implements the code that the Developer agent writes. It is given the code relevant for the current task (that is previously selected by LLM in the code-fetching phase) and the description that the Developer agent created in development step #1. Then, the only thing it needs to return is the completely coded sections/files that we can just insert/replace in the codebase. CodeMonkey #4 Testing There are two places where testing is done – (1) after each development task when GPT Pilot creates integration tests that test if the high-level features work as intended and (2) after each development step when it creates smaller unit tests that ensure all functions work as expected. GPT Pilot has three different types of tests it can do: are the preferred way of testing a step or task because they will be used in a regression test suite so that GPT Pilot can be sure that new code changes don’t break old features. However, automated tests are not always the most optimal way to test new code. Automated tests is a test where we run a specific command and give the output to the LLM, which then tells us if the implementation was successful. For example, we don’t need to create an automated test that will check if we can run an app with npm run start – for that, a simple command run is enough to check if we successfully set up our environment. Command run is the final way to test the app, and it is needed whenever AI cannot test the implementation itself. This is needed, for example, when there are some visual aspects (e.g., CSS animations) that must be checked to see if they work correctly. Human intervention After running each test, if successful, GPT Pilot takes on the next task or step and continues with coding, but when the test fails, GPT Pilot needs to debug the error. #5 Debugging The debugging process needs to be robust so that it can be started on any bug that arises, regardless of the error. It also needs to be able to debug any issue that happens during the debugging process. This is where recursive conversations come in, which are conversations with the LLM that are set up in a way that they can be used “recursively.” Let’s look at the example in the image below. It represents a flow that the GPT Pilot goes through when working on a development task that has five development steps. In this example, during the development of step #3, an error occurs – let’s say it implements a specific code change, but after running a test, it fails. Then, it goes into the recursion level #1 to debug this issue. It breaks down what needs to be done to fix this issue into two steps, but during the implementation of the first step, another error happens. For example, a needed dependency for fixing error #1 doesn’t exist. GPT Pilot then goes into the recursion level #2, which it breaks down into three steps. In the third step, another error occurs. Then, it goes to the third recursion level, which has only 1 step. Once that step is successfully executed, GPT Pilot goes back to the recursion level #2 and finishes debugging error #2. After that, it goes back to debugging error #1, and finally, after error #1 is fixed, it goes back to development step #3, after which it continues the app implementation. When the recursions go five levels deep, GPT Pilot will stop the debugging process and ask the developer to fix the initial issue it started with. Once the developer resolves this issue, they write the results to GPT Pilot. Then, it can continue the development process as if it debugged the issue itself. Conclusion In the first post of this series, I discussed the high-level overview of how GPT Pilot works. In this post, I described the GPT Pilot Coding Workflow, including: How Developer and CodeMonkey agents work together to implement code (write new files or update existing ones), How recursive conversations and context rewinding work in practice and Rewinding the app development process and restoring it from any development step. In the final post, I will dive deep into how all the agents are structured. We built the agents modularly because we know they will evolve over time. Please head over to GitHub, clone the , experiment with it, and send me your feedback. I want GPT Pilot to be as helpful to developers as possible, so let me know what you think, how it can be improved, or what works well. Add comments at the bottom of this post, or email me at . GPT Pilot repository zvonimir@pythagora.ai 🌟🌟🌟 Finally, we're trying to raise funds to continue developing GPT Pilot, so . Thank you 🙏 🌟🌟🌟 it would mean A LOT if you could and/or share it with your friends star GPT Pilot Github repository Also published . here

How GPT Pilot Codes 95% of Your App [Part 2]

Yesterday, the non-profit organization Center for AI Safety about the existential risks of AI: released a one-sentence statement “ ” Mitigating the risk of extinction from A.I. should be a global priority alongside other societal-scale risks, such as pandemics and nuclear war The statement was signed by 350 prominent AI researchers and industry leaders such as OpenAI CEO, , Google CEO, Demis Hassabis, and Anthropic CEO, Dario Amodei. Earlier this month, the same three tech leaders to discuss future AI regulations.  AI is one of the few industries where major companies ask the regulators to impose boundaries on what they can and cannot do, usually, it works the other way around. Sam Altman DeepMind met with the Biden administration The mid-to-long-term AI risks are serious, although difficult to understand and predict. In this light, new copyright infringement issues posed by gigantic generative AI models, may not seem like a big deal. There is certainly not any life-or-death urgency involved in clarifying the legal landscape. On the other hand, the tension between generative AI models and copyright law is imminent and the dangers are easy to understand and explain. The Copyright Issues Before Open AI’s commercial breakthrough with , and before the competitive race among BigTech companies to develop larger and more impressive AI models, legal scholars were discussing who would own the right to a piece of original work created by an AI. As it turns out, the question of authorship has had very little practical relevance.  Neither providers nor users of generative AI models have shown much interest in claiming rights over content generated by AI. ChatGPT More relevant and pressing issues are: If using large volumes of copyright-protected material in the training process of large AI models is infringing the rights of copyright holders, and 2) What if outputs generated by AI models look suspiciously similar to copyrighted works used in the training process? In my next post, I will attempt to answer both questions by looking at the AI image model which was recently targeted with lawsuits from two different fronts. In this post, I will provide some relevant background information on the EU’s upcoming AI Act and the debate about transparency requirements. Stable Diffusion The EU AI Act and Foundation Models The European Union agreed on a on the 11th of May. Among the changes from previous drafts comes additional regulation of so-called “foundation models”. Compromise Text for its EU AI Act A  foundation model is defined as . Examples of AI model that falls into this category is ChatGPT and Stable Diffusion. "an AI system model that is trained on broad data at scale, is designed for the generality of output, and can be adapted to a wide range of distinctive tasks" The Compromise Text sets out requirements for developers of foundational models to register the model in an EU database. In this context, they have to disclose certain information and technical documentation of a range of factors related to the development of the model. According to Section C in ANNEX VIII, foundation model providers such as OpenAI and would be required to disclose data sources and which training resources are used in the development of their models. Google In addition to the requirements in ANNEX VIII, providers of foundational models are among other things required to “ (Article 28b (4) (c)). document and make publicly available a sufficiently detailed summary of the use of training data protected under copyright law” This provision is interesting and I am curious to learn how it will be interpreted in practice. OpenAI has famously not revealed anything about how or with what data GPT-4 was trained. Once the EU's AI Act enters into force, this has to change if they wish to continue operating in the EU. OpenAI, Upcoming Regulations & Copyright Issues It seems like OpenAI is speaking in two tongues. On one hand, Sam Altman and his team support AI regulation and worry about the future implications of AI and its dangers to humanity. On the other hand, they have so far refused to open up GPT-4 for public scrutiny due to competitive concerns. OpenAI’s split-personality is foreseeable since the company has a stated mission , while they are also riding on a ten-billion-dollar investment from Microsoft. The strange mixture of altruistic concerns for humanity and colossal funding from BigTech makes me wonder if OpenAI is willing to play ball with regulators, or if it will be more profitable for them to not do so and remain closed. “to ensure that artificial general intelligence benefits all of humanity” While the training process behind GPT-4 is closed for public scrutiny, we can be sure of one thing: it includes the use of copyright-protected material on an unfathomably large scale. which consists of more than 15 million domains and it still only accounts for a relatively small part of the size of GPT-3’s complete training data set. Presumably, the large majority of text GPT-3 used to analyze and learn from was protected by copyright. Washington Post analyzed the Google C4 dataset I wonder if and how OpenAI and other providers of foundation models can prepare a “sufficiently detailed summary" of all this data in accordance with EU's AI Act. The interpretation is important. Last Wednesday, if OpenAI could not comply with the upcoming EU regulations. He has since reversed the statement in a Tweet from last Friday, where he clarified: " . " Sam Altman floated the idea of leaving EU We are excited to continue to operate here and of course have no plans to leave At a recent event in London during , the : Sam Altman’s “Euro tour” OpenAI CEO said " " The current draft of the EU AI Act would be over-regulating, but we have heard it's going to get pulled back Dragos Tudorache, an EU parliament member who is leading the drafting of EU proposals disagrees: "I don't see any dilution happening anytime soon (..) These provisions relate mainly to transparency, which ensures the AI and the company building it are trustworthy. I don't see a reason why any company would shy away from transparency."

The Legal Copyright Battle Against AI: An Introduction to the EU’s Requirements

How to Fine-tune and Optimize GPT Assistants with OpenAI

Cómo ajustar y optimizar los asistentes de GPT con OpenAI

Comment affiner et optimiser les assistants GPT avec OpenAI

OpenAI के साथ GPT असिस्टेंट को कैसे फाइन-ट्यून और ऑप्टिमाइज़ करें

OpenAI を使用して GPT アシスタントを微調整および最適化する方法

Como ajustar e otimizar assistentes GPT com OpenAI

Cách tinh chỉnh và tối ưu hóa GPT Assistants với OpenAI

如何使用 OpenAI 微调和优化 GPT 助手

Modern applications often incorporate search solutions to enable users to quickly access existing content on demand. It is difficult to envision any other functionality that could efficiently retrieve this information, making search an essential feature in most applications. Custom-made search solution options Simultaneously, even if the need to query and search is very common, different applications take drastically different approaches. In most cases, companies remain at a very basic level of search by simply querying OLTP databases directly. Requests may look like this: . SELECT id, title FROM, entities WHERE, description LIKE ‘%bow% However, more frequently, they are represented in complex, multiple-level table join structures which are impossible to read, slow, and primitive. They are unable to comprehend context, require numerous customizations, and are very challenging to implement properly. While it is possible to improve query execution time through materialized views, query caching, and other technics, the added complexity results in a considerable lag between primary updates and consistent search results. More efficient alternatives to primitive DB-based search solutions may constitute open-source search engines such as Apache Lucene, Apache Solr, Elasticsearch, Sphinx, MeiliSearch, Typesense, etc. These tend to be comparatively faster and much better at dealing with complex queries and working with filters. But once these search engines are compared to counterparts like Search or DuckDuckGo, it becomes clear that open-source solutions fail at building proper search context and query modalities — they are incapable of understanding a query if the user provides a vague search request. Google Extracting the meaning from the query Imagine you simply cannot remember what that yellow citrus fruit with a sour taste are called! But you want to search the app for an article on how to grow this mysterious fruit. How do you go about that search? Your query may be, “How to grow yellow sour citrus indoors”. Any of the aforementioned open-source search engines may struggle significantly to return relevant results for this query, even if the database does contain articles about growing “lemons”. That is because the extraction of meaning from a query is a natural language task and is unlikely to be solved without AI components. is good at this task. GPT-3 OpenAI offers based on GPT-3 that converts natural language text into a vector of floating point numbers. An embedding is essentially a low-dimensional space into which high-dimensional vectors can be translated. In this case, the high-dimensional vector is text and the low-dimensional is an output vector of fixed size. The distance between vectors represents how strongly they correlate. The smaller the distance, the higher the correlation. By redefining text as a vector-based form, the task is reduced to a classical ML search problem. an embeddings API Choosing the right model The conversion of document text into a vector representative can happen in the background, while the vectorization of the search query should occur during runtime. There are several GPT-3 model families that OpenAI offers: text-search-ada-doc-001: 1024
text-search-babbage-doc-001: 2048
text-search-curie-doc-001: 4096
text-search-davinci-doc-001: 12288 Higher vector dimensions lead to more embedded information and, thus, also higher costs and slower searches. Documents are usually long and queries are generally short and incomplete. Therefore, the vectorization of any document differs significantly from any query’s vectorization considering the content’s density and size. OpenAI know that, and so they offer two paired models, and : -doc -query text-search-ada-query-001: 1024
text-search-babbage-query-001: 2048
text-search-curie-queryc-001: 4096
text-search-davinci-query-001: 12288 It is important to note that the query and document must both utilize the same model family and have the same length of the output vector. Example dataset It may be easiest to observe and understand the power of this search solution through examples. For this example, let us draw on the which contains metadata about roughly 5,000 movies from TMDb. We will build a search solution relying only on movie documents, not reviews. TMDB 5000 Movie Dataset The dataset contains plenty of columns, but our vectorization process will be built around the title and overview columns only. Example of the record: Title: Harry Potter and the Half-Blood Prince

Overview: As Harry begins his sixth year at Hogwarts, he discovers an old book marked as 
‘Property of the Half-Blood Prince’, and begins to learn more about Lord Voldemort’s dark past. Let’s map the dataset into ready-to-indexing text: datafile_path = "./tmdb_5000_movies.csv"
df = pd.read_csv(datafile_path)

def combined_info(row):
  columns = ['title', 'overview']
  columns_to_join = [f"{column.capitalize()}: {row[column]}" for column in columns]
  return '\n'.join(columns_to_join)

df['combined_info'] = df.apply(lambda row: combined_info(row), axis=1) The embedding process is straightforward: def get_embedding(text, model="text-search-babbage-doc-001"):
   text = text.replace("\n", " ")
   return openai.Embedding.create(input = [text], model=model)['data'][0]['embedding']

get_embedding(df['combined_info'][0]) This block of code outputs a list the size of which is equal to the parameters that the model is operating with, which in the case of is 2048. text-search-babbage-doc-001 A similar embedding process should be applied to all documents we would like to search on: df['combined_info_search'] = df['combined_info'].apply(lambda x: get_embedding(x, model='text-search-babbage-doc-001'))
df.to_csv('./tmdb_5000_movies_search.csv', index=False) Column will hold a vector representation of the combined_text. combined_info_search And, surprisingly, that’s already it! Finally, we are ready to perform a sample search query: from openai.embeddings_utils import get_embedding, cosine_similarity

def search_movies(df, query, n=3, pprint=True):
    embedding = get_embedding(
        query,
        engine="text-search-babbage-query-001"
    )
    df["similarities"] = df.combined_info_search.apply(lambda x: cosine_similarity(x, embedding))

    res = (
        df.sort_values("similarities", ascending=False)
        .head(n)
        .combined_info
    )
    if pprint:
        for r in res:
            print(r[:200])
            print()
    return res


res = search_movies(df, "movie about the wizardry school", n=3) These are the results we get: Title: Harry Potter and the Philosopher’s StoneOverview: Harry Potter has lived under the stairs at his aunt and uncle’s house his whole life. But on his 11th birthday, he learns he’s a powerful wizard — with a place waiting for him at the Hogwarts School of Witchcraft and Wizardry. As he learns to harness his newfound powers with the help of the school’s kindly headmaster, Harry uncovers the truth about his parents’ deaths — and about the villain who’s to blame.

Title: Harry Potter and the Goblet of FireOverview: Harry starts his fourth year at Hogwarts, competes in the treacherous Triwizard Tournament and faces the evil Lord Voldemort. Ron and Hermione help Harry manage the pressure — but Voldemort lurks, awaiting his chance to destroy Harry and all that he stands for.

Title: Harry Potter and the Prisoner of AzkabanOverview: Harry, Ron and Hermione return to Hogwarts for another magic-filled year. Harry comes face to face with danger yet again, this time in the form of an escaped convict, Sirius Black — and turns to sympathetic Professor Lupin for help. The overview for ‘ ’ contains the words ‘wizardry’ and ‘school’ and comes first in the search output. The second result no longer contains the word ‘school’, but still retains words close to ‘wizardry’, ‘Triwizard’. The third result contains only a synonym of ‘wizardry’ — magic. Harry Potter and the Philosopher’s Stone There are, of course, a multitude of other movies within this database that feature schools or wizards (or both), but the above were the only ones returned to us. This is clear proof that the search solution works and actually understood the context of our query. We used the Babbage model with only 2048 parameters. Davinci has six times more (12,288) parameters and can, thus, perform significantly better with regard to highly complex queries. The search solution may occasionally fail to produce output relevant to some queries. For instance, the query ‘movies about wizards in school’ produces: Title: Harry Potter and the Philosopher’s StoneOverview: Harry Potter has lived under the stairs at his aunt and uncle’s house his whole life. But on his 11th birthday, he learns he’s a powerful wizard — with a place waiting for him at the Hogwarts School of Witchcraft and Wizardry. As he learns to harness his newfound powers with the help of the school’s kindly headmaster, Harry uncovers the truth about his parents’ deaths — and about the villain who’s to blame.

Title: Dumb and Dumberer: When Harry Met LloydOverview: This wacky prequel to the 1994 blockbuster goes back to the lame-brained Harry and Lloyd’s days as classmates at a Rhode Island high school, where the unprincipled principal puts the pair in remedial courses as part of a scheme to fleece the school.

Title: Harry Potter and the Prisoner of AzkabanOverview: Harry, Ron and Hermione return to Hogwarts for another magic-filled year. Harry comes face to face with danger yet again, this time in the form of an escaped convict, Sirius Black — and turns to sympathetic Professor Lupin for help. What is ‘Dumb and Dumberer: When Harry Met Lloyd’ doing here you may wonder? Thankfully, this issue was not reproduced on parameters with more parameters. Calculating the distance between the query and documents The search output should consist of documents sorted in descending order by relevance. To achieve this, we should be aware of the distance between the current query and each document. The shorter the length, the comparatively more relevant the output. Then, after a defined maximum reach, we should stop considering the relevance of the remaining documents. In the aforementioned example, we used to calculate distance due to the high dimensionality of the vector space. With the babbage model, we have 2048 parameters. cosine similarity Distance calculation algorithms tend to represent this similarity (difference) between a query and a document with a single number. However, we cannot rely on the because of the — distances will be too similar. This is because Euclidean distance becomes highly impractical beyond around seven dimensions — at this point, distances fall into the same buckets and become almost identical. Euclidean distance curse of dimensionality If you would like, you may check out the resulting repository here: https://github.com/weblab-technology/gpt3-embeddings-movies-search?embedable=true Alternatively, you can play with it in Google Colab . here Time complexity We used the brute force approach to sort the documents. Let’s determine: ● n: number of points in the training dataset ● d: data dimensionality The search time complexity for brute force solutions is . Parameter depends on the model (in the case of Babbage, it is equal to 2048) while we have block due to the sorting step. O(n * d * n * log(n)) d O(nlog(n)) It is critical to remind ourselves at this stage that smaller models are faster and cheaper. For instance, in the search case similarity calculation step, the Ada model is two times faster, while the Davinci model is six times slower. Cosine similarity calculations between the query and 4803 documents of 2048 dimensions took 1260 ms on my M1 Pro. In the current implementation, time required to calculate would grow linearly to the total number of documents. Simultaneously, this approach supports computation parallelization. Alternatives to the brute force solution In search solutions, queries should be completed as quickly as possible. And this price is usually paid on the side of training and pre-caching time. We can use data structures like a k-d tree, r-tree, or ball tree. Consider the article from about the computational complexity analysis of these methods: They all lead to computational complexity close to , where is the number of elements we would like to return within a single batch. Towards Data Science O(k * log(n)) k K-d trees, ball trees, and r-trees constitute data structures that are used to store and efficiently search for points in N-dimensional space, such as our meaning vectors. K-d and ball trees are tree-based data structures that use an iterative, binary partitioning scheme to divide the space into regions, with each node in the tree representing a subregion. K-d trees are particularly efficient at searching for points within a specific range or finding the nearest neighbor to a given point. Similarly, r-trees are also used to store points in N-dimensional space, however, they are much more efficient at searching for points within a specific region or finding all points within a certain distance of a given point. Importantly, r-trees use a different partitioning scheme to k-d trees and ball trees; they divide the space into ‘rectangles’ rather than binary partitions. Tree implementations fall outside the scope of this article, and different implementations will lead to different search outputs. Query time Perhaps, the most significant disadvantage of the current search solution is that we must call an external OpenAI API to retrieve the query embedding vector. No matter how quickly our algorithm is able to find the nearest neighbors, a sequential blocking step will be required. Text-search-babbage-query-001
Number of dimensions: 2048
Number of queries: 100
Average duration: 225ms
Median duration: 207ms
Max duration: 1301ms
Min duration: 176ms Text-search-ada-query-002
Number of dimensions: 1536
Number of queries: 100
Average duration: 264ms
Median duration: 250ms
Max duration: 859ms
Min duration: 215ms Text-search-davinci-query-001
Number of dimensions: 12288
Number of queries: 100
Average duration: 379ms
Median duration: 364ms
Max duration: 1161ms
Min duration: 271ms If we take the median as a referencing point, we can see that ada-002 is +28% slower and davinci-001 is +76% slower. Limitations of GPT-3 Search Embedding Referring to , we may conclude that GPT-3 does not provide an exceptional performance or output quality and requires dependence on external API that is rather slow. GPT-3 has the capability to support inputs of up to 4096 tokens (approximately 3072 words), however, there is no truncation service available through the API and attempting to encode text longer than 4096 tokens will result in an error. Thus, it is the responsibility of the user — you — to determine how much text can actually be encoded. Nils Reimer’s article about dense text embeddings model comparison Also, training costs are relatively high with OpenAI. Alternatively, you may consider trying or . TAS-B multi-qa-mpnet-base-dot-v1 Approximate nearest neighbor search in Elasticsearch Elasticsearch 8.0 supports efficient approximate nearest neighbor search (ANN) that may be used to solve our problem magnitudes faster than any linear KNN could. Elasticsearch 8.0 utilizes an ANN algorithm called Hierarchical Navigable Small World graphs (HNSW) to organize vectors into graphs based on similarity. Tested on a dataset of 10 million embedded vectors, we achieved an impressive performance of 200 queries per second with ANN on a single compute-focused machine, while only 2 queries per second using KNN. Both were provided by Elasticsearch. According to ElasticSearch’s and the , the cost of this performance is 5% of recall. This means that around 1 in 10 of the closest vectors found are not genuinely relative. We applied a hybrid approach to improve this result: Request k+l instead of k and apply filtering to remove outliers. Sadly, there are no pagination capabilities. So, this approach will only work in some use cases. blog post about the ANN release benchmark for dense vector searches As you will have hopefully seen, GPT-3 Embeddings is not the perfect solution for any and all search problems due to its indexing complexity, cost, as well as the high computational complexity of the search operation, even of approximates. Nevertheless, GPT-3 Embeddings remains an excellent choice for anyone looking for a powerful backbone for a search solution with infrequent querying and modest indexing requirements. Moreover, it’s also worth adding that recently that the Bing search engine now uses the new upgraded version of GPT 3.5, called ‘Prometheus’ and developed initially for search. According to the announcement, the new Prometheus language model allows Bing to increase relevance, more accurately annotate snippets, provide fresher results, understand geolocation and improve security. This may open up new possibilities for using the autoregressive language model for search solutions, which we will definitely keep an eye on going forward. Microsoft announced References: New and Improved Embedding Model k nearest neighbors computational complexity scipy.spatial.KDTree k-d tree OpenAI GPT-3 Text Embeddings — Really a new state-of-the-art in dense text embeddings? Introducing approximate nearest neighbor search in Elasticsearch 8.0 About the author Oleksandr Knyga is a Solution Architect and open-source believer with a big passion for Software Development. Oleksandr lives the life of a digital nomad. When not tapping on the keyboard, he can be found hiking, snowboarding, fishing, and exploring the city. Oleksandr has an MBA and a Computer Science Master’s degree. Also published here.

Harnessing GPT-3 for Improved Search Solutions

Aprovechar GPT-3 para mejorar las soluciones de búsqueda

Exploiter GPT-3 pour améliorer les solutions de recherche

बेहतर खोज समाधानों के लिए GPT-3 का उपयोग करना

GPT-3 を利用して検索ソリューションを改善する

Aproveitando o GPT-3 para soluções de pesquisa aprimoradas

Khai thác GPT-3 cho các giải pháp tìm kiếm được cải thiện

利用 GPT-3 改进搜索解决方案

GPT-4, arguably the most powerful AI model ever, . has been released The expectation was that this was going to be massive. OpenAI’s announcement does not focus on model size, but on capabilities. They seem impressive, and GPT-4 is purportedly being used at the likes of and Stripe Morgan Stanley. The question I posed just before it was released was — does it matter? My unpopular opinion: probably not as much as ChatGPT, probably not for the reasons you think. I’ve been covering AI since 2016, initially for , then also for . I also happen to have a background and experience in the field. Some people have been wondering about my take on ChatGPT / the Generative AI explosion, as I’ve been kind of silent about it. They’re kind of right. ZDNet VentureBeat I have not been posting either on ZDNet or VentureBeat as of late. The reasons for that are a different, albeit related story. Regardless, if you know or have been following me, my take would probably be not that hard to guess. Here it goes. Do you remember how you joined ? I do, very clearly. It was 2007, after hearing about it from a friend. As someone who has been in technology his entire life, I was not particularly impressed. Yes, Facebook had a user-friendly interface and it worked well. But the thing that stood out about it the most was – I heard about it from a friend. Facebook I’m a digital native, but it’s not because I’m very young. It’s because I came across computers as a kid, adopted them immediately for work and play, and decided that’s what I wanted to do in life. I’ve been writing code and documents since the 80s, using the internet, and communicating via email and apps since the 90s. I have always thought of technology as a force for good – raising productivity, providing access to knowledge, connecting people, and all that. No wonder then, when Facebook came along I thought – hey, looks fun and could make more people go digital. I even invited some of my friends to join. In all honesty, I had forgotten about this embarrassing detail. It was one of my old techie friends who reminded me a couple of weeks back. Being the early adopter among early adopters that I am, my friend was curious to know what I thought of ChatGPT, OpenAI’s recently released AI language model. I’m also a semantics geek who . Maybe that’s why I don’t buy the “AI” hype. When we say “AI”, . It really is tech companies using massive data sets to train models to perform – sometimes useful and/or impressive – imitations of human intelligence. happens to be keeping track of AI for a few years it really is a convention, not something to be taken at face value Still, my friend was right in saying that ChatGPT is AI’s Facebook moment – for better or worse. There was AI before ChatGPT, and there will be AI after it. What made ChatGPT stand out over predecessors like GPT-3 is similar to what made Facebook stand out over predecessors like MySpace. Qualitative and quantitative differences may exist, but that’s not what kept my friends from raving about GPT-3 or MySpace. OpenAI, the makers of ChatGPT, does not lack technical sophistication, cash, or good PR. GPT-3 was already accessible to select members of the press and can be used via an offered and monetized . A first wave of apps built on top of this API is . But it was making ChatGPT easily accessible to the general public that made the difference. API in collaboration with Microsoft already out there Facebook did not become a household name by offering APIs to build on. Facebook offered direct, unfettered access to features like photo sharing and newsfeed. In addition to getting the coverage lots of cash and good PR will get you, opening up platforms like Facebook or ChatGPT to the public gets the early adopter crowd excited and talking about it. This gets the media flywheel turning. Whether genuinely impressed or out of FOMO, . Organic coverage is great and works wonders for growing your user base, but there are some significant ways in which Facebook and ChatGPT are different..or are there? editors will dispatch and journalists will write their takes Facebook made sure that users flock in and are then locked in. After all, as Cory Doctorow put it, once a critical mass of people you cared about were on Facebook, it became effectively impossible to leave. Despite the growing realization of the of Facebook and , not much has changed. ‘enshittification’ efforts to fight back via interoperability protocols Facebook is free, and always will be. That was the slogan it sported on its homepage, . Perhaps the realization that if something is free, it means the product is you hit home. But while a growing part of users may have come to accept the fact that their data is harvested on and off Facebook, , they don’t necessarily understand what the data is used for. until public sentiment changed using every trick in the book and beyond Besides targeted advertising or the , data is also used to develop features and products. Everything from your timeline to is data-driven and AI-powered. More data, more power, more products, more data. occasional genocide image cropping That’s also true for language models like ChatGPT which are trained on huge amounts of data. This is where the really interesting parallels begin. ChatGPT is free to use at the moment, although a subscription plan also exists in addition to API monetization. Microsoft, which has a , is rushing to . multi-billion investment in OpenAI integrate ChatGPT in its Bing search engine Although a mix of and , a plausible reason for the rush is to leverage user interaction to fine-tune ChatGPT. While some guardrails have been added to ChatGPT by , they and more user exposure will help strengthen them. blunders, limitations uncanniness is being exposed in the “New Bing aka Sydney” manually fine-tuning can be sidelined You may be fine with helping build someone else’s product for free, again. But there are more things that Facebook and contemporary AI have in common: funding and agendas. The so-called was . It’s also here to give AI a boost via the Effective Altruism initiative. PayPal Mafia there to give Facebook a boost Insiders like Émile P Torres and Timnit Gebru are issuing stark warnings on Effective Altruism as a and the influence it’s exerting, . Others like Dan McQuillan and Rajiv Malhotra are raising critical voices, calling for a resistance against in the . dangerous secular credo specifically in AI AI as another weapon in Big Tech’s arsenal global battleground for power and control If you still have doubts about the thinking behind this, reading should give you pause. Phrases like “Our mission is to ensure that artificial general intelligence—AI systems that are generally smarter than humans—benefits all of humanity”, especially delivered by and have no problem at all making U-turns on their commitments, read more like a thinly veiled threat than anything else. OpenAI’s “manifesto” for AGI people who think we are all “stochastic parrots” Shouldn’t we, the early digital natives, have known better in 2007? We should. Yet, we still saw technology in a positive light. We underestimated Big Tech’s potential for harm and overestimated people’s ability to see through the lies and manipulation to make technology their own. Isn’t there any ray of hope today? Maybe – but that’s a topic for a different note. One thing is certain: we should not make that . It’s going to be a long and winding road, but we’ll never get anywhere if we don’t know where we stand. mistake of being “dumb f***s” again Originally published on Linked Data Orchestration

DataGPT

datagpt.com

11-50 employees

Since 2021

COMPANY RANKING

DATAGPT

EVERGREEN INDEX #187

DataGPT WIKI

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

DataGPT

datagpt.com

11-50 employees

Since 2021

COMPANY RANKING

DATAGPT

EVERGREEN INDEX #187

RELATED COMPANIESRANK

STORIES ABOUT DATAGPT

DataGPT WIKI

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