<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Rosemary J Thomas, PhD on Medium]]></title>
        <description><![CDATA[Stories by Rosemary J Thomas, PhD on Medium]]></description>
        <link>https://medium.com/@me.rosemary?source=rss-a9543806140d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*YpZtegMKOwziG77t2NYJlA.jpeg</url>
            <title>Stories by Rosemary J Thomas, PhD on Medium</title>
            <link>https://medium.com/@me.rosemary?source=rss-a9543806140d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 05:29:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@me.rosemary/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Unmasking Generative AI: Understanding Explainability Techniques]]></title>
            <link>https://medium.com/version-1/unmasking-generative-ai-understanding-explainability-techniques-cb7eaf27ff16?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb7eaf27ff16</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <category><![CDATA[ai-explainability]]></category>
            <category><![CDATA[xai]]></category>
            <category><![CDATA[genai]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Tue, 13 Aug 2024 08:32:51 GMT</pubDate>
            <atom:updated>2024-08-15T14:31:40.577Z</atom:updated>
            <content:encoded><![CDATA[<p>Generative AI models are often referred to as ‘Black Boxes’. And Explainable AI has been a long-standing point of discussion in the traditional AI models space. Have you wondered what Explainable AI means in GenAI space? Does it help to understand and interpret the complex decision-making processes of AI models?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HtEgB3OXzUVclZMEyS5UIg.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>Let’s dig in. In existing literature, there are <a href="https://arxiv.org/abs/2404.09554">four primary categories of General AI explainability techniques</a>. First, <strong>feature-based explainability techniques</strong> focus on the importance of individual features in the model’s predictions, highlighting model’s aspects of the data it considers most meaningful. Second,<strong> sample-based explainability</strong> uses representative samples to explain the model’s behaviour. Third, <strong>the mechanistic approach</strong> examines the inner workings of the model itself, often creating simpler, interpretable models that approximate the behaviour of more complex ones. Lastly, <strong>probing-based explainability</strong> involves examining the model’s responses to various inputs, providing insights into its internal representations and decision-making processes. Each of these approaches offers unique perceptions and using them together may steer to a broad understanding of GenAI models.</p><h3>Feature-based</h3><p>Feature attribution plays a vital role by providing a relevance score for each input feature, such as a word or pixel. This helps in understanding which features are most influential in the model’s decision-making process. Another approach is <strong>perturbation</strong>, which changes inputs partially, for example, by removing or changing features and then examining the output. This can provide insights into how sensitive the model’s predictions are to changes in specific features. The <strong>gradient-based method</strong> is illustrated by techniques like Grad-CAM <a href="https://datadance.ai/deep-learning/visualizing-model-insights-a-guide-to-grad-cam-in-deep-learning/">which stands for Gradient-weighted Class Activation Mapping</a> requires a backward pass from outputs to inputs to obtain derivatives, proposing a way to visualise the value of each feature.</p><p>The <strong>surrogate model</strong> <strong>approach</strong> uses much simpler models, such as <a href="https://github.com/marcotcr/lime">LIME </a>(Local Interpretable Model-agnostic Explanations), <a href="https://shap.readthedocs.io/en/latest/index.html">SHAP </a>(SHapley Additive exPlanations), <a href="https://klu.ai/glossary/effective-accelerationism">EAC </a>(Effective Accelerationism), <a href="https://segment-anything.com/">SAM </a>or <a href="https://github.com/facebookresearch/segment-anything-2">SAM 2</a> (Segment Anything Model 2) and others, to understand individual predictions. These models can approximate the complex model’s behaviour, providing a more explainable reason. The <strong>decomposition model</strong> <strong>approach</strong>, including techniques like <a href="https://link.springer.com/chapter/10.1007/978-3-031-04083-2_2">LRP </a>(Layer-wise Relevance Propagation), focuses on accrediting relevance from outputs towards inputs or decomposing vectors. It can also refer to decomposing the reasoning process and accrediting outputs to specific reasons. Lastly, <strong>attention-based</strong> <strong>approach</strong> provides an importance score for inputs, where inputs are not necessarily inputs to the network but those of a prior layer. The scores are commonly acquired between all input token pairs for a single attention layer, which can be visualised using a heatmap or bipartite graph. This approach is particularly useful in the context of neural networks, where it can provide insights into the model’s internal representations.</p><h3>Sample-based</h3><p>Sample-based techniques are methods that examine changes in outputs for different inputs. The focus is more on the sample in its entirety to understand the relationship between various inputs and their corresponding outputs.</p><p>Two key techniques in this category are the<strong> influence of training data </strong>and <strong>adversarial samples</strong>. The former measures the impact of a specific training sample on the model. The latter involves input changes that are small and invisible for people, that lead to a change in outputs.</p><p>The other two techniques are <strong>counterfactual explanations</strong> and <strong>contrastive explanations</strong>. Counterfactual explanations spot smallest changes to an input so that the model’s output changes from one class to another specific class. The most typical example is a loan approval model. A counterfactual explanation might show that increasing the applicant’s income by a certain amount could change the model’s decision from ‘deny’ to ‘approve’.</p><p>While contrastive explanations explore why a model made a certain prediction ‘A’ instead of another prediction ‘B’. These techniques provide a deeper understanding of the model’s decision-making process. For example, in a text generation model, a contrastive explanation might explore why the model chose to output one word over another, considering various aspects such as part of speech, tense, and semantics.</p><h3>Mechanistic approach</h3><p>The <a href="https://philsci-archive.pitt.edu/22747/1/Explaining%20AI%20through%20Mechanistic%20Interpretability_final_v2.pdf">mechanistic approach</a> is a method that studies neurons and their interconnections within a model. Primarily, it reverse-engineer the components of a model into algorithms that are understandable to people i.e. it provides a deeper understanding of the inner workings of AI models, making them more transparent and trustworthy.</p><p>First technique in this approach is <strong>circuit discovery</strong>. Typically, its manual workflow that includes the creation of the model, defining metrics, and dataset creation. The goal is to discover how different parts of the model, or <strong>circuits</strong>, interact and contribute to the final output.</p><p>Second is <strong>causal tracing</strong> which calculates the impact of intermediate activation on the output. In other words, it traces the effect of certain neurons or groups of neurons in the network on the final prediction, helping to understand which parts of the model are most influential in decision-making.</p><p>Lastly, the <strong>vocabulary lens</strong> technique is used to determine relations to the vocabulary space. This technique focuses on understanding how the model processes and uses different words or tokens. By examining these relationships, we can gain insights into how the model understands and generates language.</p><h3>Probing-based</h3><p>Probing-based methods helps to understand the knowledge that a Language Learning Model (LLM) has captured through queries. A common approach is to train a classifier, often referred to as a probe, on a model’s activation to distinguish different types of inputs and outputs.</p><p>There are few types of probing methods. First, <strong>knowledge-based probing</strong> involves training a classifier on the model’s outputs with the goal of identifying the presence of output properties or abilities that emerge from the inputs. For example, if the LLM is trained to translate languages, a knowledge-based probe might look for the presence of correct grammar in the translated output.</p><p><strong>Concept-based probing</strong> provides relevance scores for a given set of concepts within the inputs. This method must be designed carefully, as it primarily examines interactions among input variables.</p><p>Lastly, <strong>neuron activation-based probing</strong> seeks to understand neurons using their activations for inputs. This method provides insights into how individual neurons contribute to the model’s overall performance.</p><p>Each of these methods offers a unique perspective on the inner workings of an LLM.</p><p><a href="https://www.version1.com/webinars-and-events/version-1s-ai-webinar-series/">Click here to learn more about Version 1’s AI Webinar Series.</a></p><p><strong>About the Author:</strong></p><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a> You can read more of her <a href="https://www.version1.com/blog/ai-labs-publications/">publications here.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb7eaf27ff16" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/unmasking-generative-ai-understanding-explainability-techniques-cb7eaf27ff16">Unmasking Generative AI: Understanding Explainability Techniques</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI in EdTech: How to Plan and Validate Your Lessons with AI?]]></title>
            <link>https://medium.com/version-1/ai-in-edtech-how-to-plan-and-validate-your-lessons-with-ai-fc6ab7222ac9?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc6ab7222ac9</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[teaching]]></category>
            <category><![CDATA[gpt-4]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[edtech]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Thu, 22 Feb 2024 15:40:55 GMT</pubDate>
            <atom:updated>2024-02-22T16:00:15.919Z</atom:updated>
            <content:encoded><![CDATA[<p>It is high time we acknowledge that artificial intelligence (AI) plays a substantial role in education. Besides, it comes with the promise to tackle some of the prime challenges in education today. Furthermore, it can not only revolutionise teaching and learning practices but also accelerate momentum towards the <a href="https://www.unesco.org/en/digital-education/artificial-intelligence">Education 2030 Agenda</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K1nVxCuhwx7zQzbmLsrmtA.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><h3>What is the importance of AI-Enabled tools in global education technology (EdTech)?</h3><p>First, it has the promise to tackle some of the prime challenges in education today including inclusion and equity. Second, integrating these tools into education solves the problems arounds highly tailored teaching experiences, adaptive subject delivery, and immediate feedback processes. Third, it can be used to cut down on administration, releasing teachers from mundane tasks and granting them to focus more on education and supporting their students. Finally, AI has penetrated education, revolutionising the way one learns and teaches, spanning from student engagement to teaching styles.</p><p>AI-enabled tools can be used in education, for example, Smart Sparrow, Century Tech, and DreamBox. These tools can <a href="https://www.skillcombo.com/blog/ai-tools-for-education-the-complete-list-2023/">create tailored learning experiences</a> and provide student improvement lesson plans by adapting in real-time.</p><h3>What is the current state of EdTech market and the future trends?</h3><p>The EdTech market size was valued at <a href="https://straitsresearch.com/report/education-technology-market">USD 105.48 billion in 2021</a> and <a href="https://www.grandviewresearch.com/industry-analysis/education-technology-market">USD 142.37 billion in 2023</a>. It is projected to reach USD 436.68 billion by 2030, growing at a compound annual growth rate (CAGR) of 17.10% during the forecast period (2022 to 2030).</p><p>There are few prominent trends shaping the future of the EdTech market. First, the COVID-19 pandemic hastened the acceptance of online learning, with many students and educators directing to digital platforms to continue education without disruptions. As the world adjusts to new AI technologies in hybrid education, a mix of both online and offline learning seems to be predicted to become the norm.</p><p>Second, AI has made floods in the EdTech space, with ChatGPT being one of the most famous disruptors. ChatGPT, a large language (learning) model, has the competency to revolutionise tailored teaching styles, helping with creating lesson plans and validating it, and also enhance problem-solving across various subjects. It can help both teachers and students.</p><p>Third, immersive technologies like augmented reality (AR) and virtual reality (VR) are ready to change how educators can adopt different teaching styles. These cutting-edge tools offer simulated experiences that can enhance the teaching experience with appropriate lesson planning, particularly in fields like science and technology.</p><p>It is worth observing that these trends are subject to shift as the field of EdTech continues to advance rapidly.</p><h3>What are some of the AI-enabled tools in Ed Tech?</h3><p>It is important to note that attributes such as tailored learning options, simple integration with your existing teaching practices, and the capability to track and evaluate students’ progress are necessary when choosing an AI tool for lesson planning and validation. These are some of factors that will help you as an educationalist to be more successful. There are numerous tools available for <a href="https://richardccampbell.com/ai-tools-for-lesson-planning/">AI-enabled lesson planning</a>. Education Copilot, LessonPlans.ai, ChatGPT, QuillBot, Google Gemini, Lex.page, Auto Classmate, and Curipod are to name some of them.</p><h4><strong>Education Copilot</strong></h4><p>It is ideal for creating quality, structured lesson plans for any subject, lesson, or concept.</p><p><a href="https://educationcopilot.com/">Lesson Planning</a>: At the kernel of Education Copilot is the AI lesson planner which can create lesson plans, presentations, and more. For example, proposing or outlining activities, resources, and assessments that align with your learning objectives.</p><p><a href="https://educationcopilot.com/ai-lesson-plans/">Validation</a>: You can easily to align or validate your lesson plans to your educational specifications and goals. This ensures that your roadmap meets the needs of your students and is aligned with the standards and objectives of your curriculum.</p><p>It offers an additional of more than 10 tools that can be used to create a variety of educational materials, further helping in the planning and validation process.</p><h4><strong>LessonPlans.ai</strong></h4><p>It is ideal for creating <a href="https://www.lessonplans.ai/about">high-quality lesson plans</a> designed to help teachers.</p><p><a href="https://www.lessonplans.ai/">Lesson Planning</a>: It creates complete lesson plans with engaging activities and resources. When teachers provide a detailed description of the lesson they want to create, the AI creates a tailored plan to their needs. It is quite simple and almost instant. In addition, teachers can then easily customise and adapt the AI-generated plan before using it in class.</p><h4><strong>ChatGPT</strong></h4><p>It is a Large Language Model capable of <a href="https://openai.com/chatgpt">answering questions, helping with tasks, and generating creative content</a>. Therefore, it can be used for crafting conversational educational content.</p><p><a href="https://www.techlearning.com/how-to/chatgpt-lesson-plan">Lesson Planning</a>: It can generate a general or specific lesson plan with learning objectives and activities depending on the details provided. You can also make ChatGPT use personas to create lesson plans. Then you can customise the plan to suit your needs.</p><p><a href="https://ditchthattextbook.com/chatgpt-lesson-plans/">Validation</a>: It can help ensure your lesson plans meet learning objectives and align with curriculum standards by using techniques like prompting i.e. by using extremely specific language.</p><h4><strong>QuillBot</strong></h4><p>It is designed to improve writing processes and well-known for its <a href="https://quillbot.com/">paraphrasing capabilities</a>. However, it also offers a range of AI writing tools and generative capabilities that can be <a href="https://quillbot.com/">beneficial in an educational context</a>.</p><p><a href="https://www.theknowledgeacademy.com/blog/what-is-quillbot/">Lesson Planning</a>: It can help teachers save time creating lesson plan materials such as worksheets, and assessments. For example, teachers can add a assessment outline into the tool, and it will swiftly generate alternative but more engaging and unique content.</p><p>Validation: Its <a href="https://quillbot.com/grammar-check">grammar checker tool</a> can be used to proof-read that is specifically useful when creating lesson plans. Thus, ensuring your content is clear, precise, and professional.</p><h4><strong>Google Gemini</strong></h4><p><a href="https://gemini.google.com/">Google Gemini</a> previously known as Google Brad is in a family of multimodal large language models. It can understand, operate across, and combine distinct types of information including text, image, audio, and video, therefore can be used in the educational setting.</p><p>Lesson Planning: It can supply you with ideas for shaping lesson plans. You can ask it to generate a detailed lesson plan tailored to your needs like ChatGPT by providing the necessary details.</p><p>Validation: It can help make sure your lesson plans meet learning objectives and align with curriculum standards by using techniques like prompting i.e. by using extremely specific language.</p><h4><strong>Lex.page or Lex</strong></h4><p>It combines a word processor editor with enhanced AI to help you write quicker, better, and more structured.</p><p><a href="https://lex.page/">Lesson Planning</a>: It can help you in creating ideas and content for your lesson plans. The ‘AI Feedback’ option help brainstorm ideas, and revise content. This can be specifically useful when you are having a writer’s block.</p><p>Validation: The ‘AI Feedback’ option also provides suggestions on your draft, helping you ensure that your lesson plans are clear, precise, and professional. You can collaborate with your colleagues and request your lesson plans be reviewed in real-time.</p><h4><strong>Auto Classmate</strong></h4><p>It is designed to help teachers in <a href="https://autoclassmate.io/">enhancing student learning</a> by providing a collection of AI tools that can be used to supplement human creativity creating collaborations, and helpful and significant learning opportunities.</p><p><a href="https://autoclassmate.io/tools/ai-powered-lesson-plan-generator/">Lesson Planning</a>: Its AI-Powered Lesson Plan Generator helps you to create highly tailored, revisable, and exclusive lesson plans. It is powered by OpenAI’s GPT-3.5 and GPT-4 that will transform your content creation methods for your lesson plans.</p><p>Validation: It allows you to easily align or validate your lesson plan with <a href="https://autoclassmate.io/tools/ai-powered-lesson-plan-generator/">state standards with a button click</a> (Align this Lesson to State Standards?). This ensures that your lesson plans meet the needs of your students and are aligned with the standards and goals of your curriculum. In addition, the Lesson Plan and Activity Forecast Tool predicts two possible outcomes each for positive and negative, permitting to review the merits and especially the demerits of your lesson plans. This helps planning for all possible scenarios. Furthermore, suggestions are proposed on how to prevent negative outcomes.</p><p>It offers a wide range of extra tools including AI-Powered Instructional Coach, an Activation and Engagement Activity Generator, and a ‘Would You Rather’ Question Generator that can be used to boost your lesson plans and materials. It also has a list of ‘Coming Soon’ tools of which Lesson Enrichment Toolkit looks promising.</p><h4><strong>Curipod</strong></h4><p>It is designed to help teachers in <a href="https://curipod.com/">creating interactive lesson plans and delivering them</a>. It uses presentation elements like polls, word clouds, drawings, Q&amp;As, and others, to offer an interactive experience for both teachers and students.</p><p>Lesson Planning: Its <a href="https://curipod.com/generators/Full-lesson/g/753d7c32-4f9b-42bf-bb53-12419c28425d">proprietary AI is a game-changer for lesson planning</a> which the teachers can then customise. It offers free lessons that <a href="https://curipod.com/discover">other teachers have created</a>, which you can use for motivation.</p><p>Validation: Content generation is aligned or validated against grades <a href="https://curipod.com/generators/Full-lesson/g/753d7c32-4f9b-42bf-bb53-12419c28425d">1st to 12th/13th (country depended) and for professional development</a>, covering diverse subjects. You can collaborate with your colleagues and request your lesson plans be reviewed in real-time.</p><h3>Takeaway</h3><p>The above-mentioned AI tools can substantially lower your time and workload required for lesson planning and validation. However, the helpfulness of these tools may change based on your own requirements and environments. For your own benefit, you should continually consider your own requirements as a priority when choosing an AI tool for lesson planning and validation.</p><blockquote><strong>Boost Your Teaching Game with the Perfect EdTech AI Tools for Lesson Planning and Validation!</strong></blockquote><p>For more intricate or specialised use cases that involve higher levels of complexity, it may be necessary to develop custom-tailored solutions.</p><p><strong>About the Author:</strong><br>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc6ab7222ac9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/ai-in-edtech-how-to-plan-and-validate-your-lessons-with-ai-fc6ab7222ac9">AI in EdTech: How to Plan and Validate Your Lessons with AI?</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meet Gemini Ultra: Google’s game-changer for the future of AI]]></title>
            <link>https://medium.com/version-1/meet-gemini-ultra-googles-game-changer-for-the-future-of-ai-7bcd57e6e44b?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/7bcd57e6e44b</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[gpt-4]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Tue, 13 Feb 2024 15:17:55 GMT</pubDate>
            <atom:updated>2024-02-21T14:58:44.255Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wms7OghGtVtrKZFyGCQvpw.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>On 8th February 2024, Google announced Gemini Advanced available <a href="https://support.google.com/gemini/answer/14517446?visit_id=638434136510511849-1265713743&amp;p=adv_countries&amp;rd=1#countries&amp;zippy=%2Cwhere-gemini-advanced-is-available">across 150 countries</a> and territories in English. Gemini Ultra 1.0 is available through Gemini Advanced. Ultra is Google’s most capable model that delivers contemporary performance across a comprehensive range of overly complex tasks, including reasoning and multimodal tasks. Google stated that more languages will be supported in the future.</p><p>Gemini Ultra was trained using multiple sets of <a href="https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf">TPUv4 accelerators</a>, across several data centres. For a recap on Google Gemini, check out our earlier <a href="https://medium.com/version-1/meet-gemini-googles-multimodal-masterpiece-that-can-push-ai-boundaries-dc16d23803a3">blog</a> post for a refresher.</p><p>Now, let us plunge into some of its <a href="https://blog.google/products/gemini/bard-gemini-advanced-app">notable features</a>:</p><p><em>· Massive Multitask Language Understanding (MMLU):</em> MMLU combines knowledge from 57 subjects, to test knowledge and problem-solving capabilities. The largest model, Ultra, <a href="https://blog.google/technology/ai/google-gemini-update-sundar-pichai-2024/">outruns human experts in MMLU</a> with an accuracy of 90.04% (Chain-of-Thought@32) while GPT-4 scored 87.29% (Chain-of-Thought@32 via API).</p><p><em>· Advanced Reasoning and Logic:</em> Gemini Ultra outruns the capabilities of the original Gemini model. It surpasses it in coding, logical reasoning, and following subtle instructions.</p><p><em>· Personalised Learning: </em>Gemini Ultra can help you as a personal tutor tailored to your learning style. It can adapt to your requirements, making learning more effective and systematic.</p><p><em>· Creative Collaboration:</em> Gemini Ultra goes beyond the conventional AI, helping you plan business strategies by being a creative collaborator for content strategy, and building content. It can broaden horizons for professionals seeking innovative solutions.</p><p>At this time, the only way to get access to Ultra is via Gemini Advanced. Users will have to sign up to a new <a href="https://one.google.com/about/plans">Google One AI Premium tier</a> that costs $18.99 per month. It is the Google One 2TB tier with the addition of Gemini Advanced. Shortly, users will also get gain access to Gemini in Google Workspace that is tending be a Microsoft Copilot comparable experience. The price seems a tad bit lower than the offerings by Microsoft and OpenAI. Google has a lead here in that it can tier up these extra features on top of their native cloud without suffering overly extra cost but adding extra revenue in the process.</p><p>Google is offering a two-month free trial to this tier to let users experience Gemini Advanced. It does not matter if you are a Google One subscriber, if you upgrade to the new AI Premium plan you will get the free trial. This might attract of lot of users.</p><p>In addition, Google has <a href="https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf">valued transparency</a> as AI becomes gradually more connected into our daily lives. Gemini Advanced had undergone multiple rounds of red-teaming, including safety and persona evaluations. It also included three types of external testing: priority (120) user program, power (50) users testing and a group user testing.</p><p>In summary, the Gemini Ultra stand for a substantial enhancement over the standard Gemini model, for professionals who expect more advanced functions. Its multimodal proficiency, exceptional performance, and emphasis on context understanding makes it a powerful AI collaborator. Gemini Ultra performed remarkably well in complex reasoning and writing tasks, while GPT-4 performed remarkably well in common sense reasoning. So there’s always capacity for improvement. Google’s wider future goal is to create a scalable, modular system with wide-ranging generalisation capabilities across diverse modalities using AI models.</p><p>Meanwhile, wait for our next blog where we assess the actual performance of Google Gemini Ultra.</p><p><strong>About the Author:</strong><br>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7bcd57e6e44b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/meet-gemini-ultra-googles-game-changer-for-the-future-of-ai-7bcd57e6e44b">Meet Gemini Ultra: Google’s game-changer for the future of AI</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Rise of Large Action Models, LAMs: How AI Can Understand and Execute Human Intentions?]]></title>
            <link>https://medium.com/version-1/the-rise-of-large-action-models-lams-how-ai-can-understand-and-execute-human-intentions-f59c8e78bc09?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/f59c8e78bc09</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Tue, 16 Jan 2024 16:41:28 GMT</pubDate>
            <atom:updated>2024-09-23T08:08:08.899Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5lNBhLqdr-UrA5zEyXMQgQ.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p><strong>A hot topic and development in the realm artificial intelligence (AI) is Large Action Models, also referred as Large Agentic Models or LAMs in short. </strong>LAMs is the spanning out of Large Language Models (LLMs) which most of us are familiar now. LLMs can generate text by predicting the next word or token based on an input. <a href="https://blog.salesforceairesearch.com/large-action-models/">LAMs take this a stage forward</a> by enhancing these LLMs to turn into ‘agents’. Agents are software units capable of running tasks by themselves, so instead of plainly answering human user queries, they are eventually helping to achieve a goal. This combines the language fluency of an LLM with the capacity to complete tasks and decision-making autonomously, which involves a substantial change.</p><p>The structure of Large Action Models is based on the composition of the applications and human actions they are designed to mock-up. <a href="https://www.rabbit.tech/research">LAMs can clearly mock-up</a> the composition of various applications and human actions executed on them without a transient demonstration, such as text. This is empowered by developments in neuro-symbolic programming. We don’t have access to model to verify this.</p><p>Large Language Models (LLMs) and Large Agentic Models (LAMs) are both types of artificial intelligence models, <a href="https://analyticsindiamag.com/lam-makes-llm-sweat/">but they serve different purposes and have different capabilities</a>. Figure 1 illustrates the core differences.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*L-gmxWvkFCJkerGK" /><figcaption>Figure 1. <a href="https://www.linkedin.com/posts/superagi_large-language-models-llms-vs-large-agentic-activity-7130503954433527809-8L0y?utm_source=share&amp;utm_medium=member_desktop">Differences between LLMs and LAMs</a></figcaption></figure><p>In terms of the overview on their working, LAMs interact with the real world through integration with external systems, such in <a href="https://interestingengineering.com/innovation/tiny-ai-device-ditches-apps">IoT devices</a> and others. By connecting to these systems, LAMs can perform physical actions, control devices, retrieve data, or manipulate information. This permits LAMs to <a href="https://blog.salesforceairesearch.com/large-action-models/">mechanise complete processes and smartly interact with the world</a>, converse with people, adjusting as conditions vary, and even working together with other LAMs.</p><p>LAMs have numerous capabilities that make them a dominant tool in the field of artificial intelligence. First, LAMs are developed to <a href="https://www.rabbit.tech/research">understand complex human goals expressed in natural language</a>, translate these intentions into actionable steps and respond in real time. Second, LAMs can interact with the world with intelligence, including people, circumstances change adaption, and other LAMs. Third, LAMs interact with the real world through integration with external systems. Finally, LAMs promote generative AI from a docile tool to a functional collaborator in getting work done in real time.</p><p>The potential <a href="https://www.peacemongernetwork.com/unveiling-the-potential-of-large-action-model-lam-in-ai-a-comprehensive-guide">use cases of Large Action Model</a> can be applied in different domains. In healthcare, LAM can transform patient care through modern diagnostics and tailored treatment strategy. In the financial sector, LAM can help in risk measurement, fraud discovery, and algorithmic transactions. In automotive sector, LAM can aid in producing self-governing vehicles and improving vehicle safety systems.</p><p>LAMs can be used in practical products, with one noteworthy instance is the <a href="https://www.rabbit.tech/">Rabbit r1 device</a> that currently retails $199 available for pre-order. The Rabbit r1 is a standalone device, around half the size of an iPhone, and features a touchscreen and a unique 360-degree rotating camera for capturing photos and videos. A scroll wheel simplifies navigating through the device, allowing users to interact (far field mic, push to talk button etc.) with the built-in assistant effortlessly.</p><p>Rabbit is an AI company that has created a tailored operating system (OS) through a natural language interface and a dedicated hardware to host the OS. The core product of Rabbit is the <a href="https://www.rabbit.tech/research">Rabbit OS</a>, which is operated by their private LAM, allowing Rabbit r1 device to identify and reproduce human actions on various technology interfaces, modernise navigation through applications naturally. This signifies a striking advancement toward effortless online user interaction without the need for any applications.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F22wlLy7hKP4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D22wlLy7hKP4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F22wlLy7hKP4%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/07b2aeb339fa3710b6f54b86fada1e9c/href">https://medium.com/media/07b2aeb339fa3710b6f54b86fada1e9c/href</a></iframe><p>Large Action Models are prepared to play a large role in shaping the future of AI. By strengthening language models to become ‘agents’ that can execute tasks on their own crafts generative AI into a real time action companion. Real-world applications like Rabbit are already harnessing the power of LAMs. This spreads a total cosmos of new prospects and denotes a big shift in the development of AI. As we continue to discover and transform, LAMs will absolutely showcase a paramount role in influencing the outlook of AI.</p><p>The voyage has just commenced. Stay tuned for more updates on this exciting frontier in AI technology.</p><blockquote><a href="https://www.version1.com/application-modernisation-integration/ai-powered-code-analysis-and-documentation-decipher/">AI-powered code analysis and documentation — Decipher | Version 1</a></blockquote><p><strong>About the Author:</strong></p><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f59c8e78bc09" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/the-rise-of-large-action-models-lams-how-ai-can-understand-and-execute-human-intentions-f59c8e78bc09">The Rise of Large Action Models, LAMs: How AI Can Understand and Execute Human Intentions?</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Small but Powerful: A Deep Dive into Small Language Models (SLMs)]]></title>
            <link>https://medium.com/version-1/small-but-powerful-a-deep-dive-into-small-language-models-slms-b793bdb002f2?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/b793bdb002f2</guid>
            <category><![CDATA[llama-2]]></category>
            <category><![CDATA[fine-tuning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[language-model]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Wed, 13 Dec 2023 12:05:50 GMT</pubDate>
            <atom:updated>2024-09-23T08:09:09.412Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0x5w1ooQva7D2QHp_HNEOw.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>Large Language Models (LLMs) have been popular for quite some time now. Lately, <strong>Small Language Models (SLMs) have enhanced our capacity to handle and communicate with various natural and programming languages.</strong> However, some user queries require more accuracy and domain knowledge than what the models trained on the general language can offer. Also, there is a demand for custom Small Language Models that can match the performance of LLMs while lowering the runtime expenses and ensuring a secure and fully manageable environment.</p><p>In this article, <strong>we explore Small Language Models, their differences, reasons to use them, and their applications. We also use fine-tuning methods on Llama-2–13b, a Small Language Model, to address the above-mentioned issues.</strong></p><p>Furthermore, our goal is to examine the possibility of making this process platform independent. For this purpose, we selected Databricks as a platform that can be transferred among the Azure, Amazon Web Services (AWS) or Google Cloud Platform.</p><blockquote><a href="https://www.version1.com/application-modernisation-integration/ai-powered-code-analysis-and-documentation-decipher/">AI-powered code analysis and documentation — Decipher | Version 1</a></blockquote><p>In the context of artificial intelligence and natural language processing, <a href="https://www.techopedia.com/definition/small-language-model-slm#:~:text=Because%20SLMs%20are%20trained%20on%2cresponses%20compared%20to%20larger%20models.&amp;text=Can%20have%20less%20than%2015%20million%20parameters.">SLM</a> can stand for ‘Small Language Model’. It is a lightweight generative AI model. The label “small” in this context refers to a) the size of the model’s neural network, b) the number of parameters and c) the volume of data the model is trained on. There are several implementations that can run on a single GPU, and over 5 billion parameters, including <a href="https://deepmind.google/technologies/gemini/#capabilities">Google Gemini Nano</a>, Microsoft’s <a href="https://www.microsoft.com/en-us/research/blog/orca-2-teaching-small-language-models-how-to-reason/">Orca-2–7b</a>, and <a href="https://huggingface.co/microsoft/Orca-2-13b">Orca-2–13b</a>, Meta’s<a href="https://huggingface.co/meta-llama/Llama-2-13b"> Llama-2–13b</a> and <a href="https://www.lakera.ai/blog/open-source-llms">others</a>.</p><p>There are <a href="https://www.businessgo.hsbc.com/en/article/what-are-large-language-models-llms-and-why-they-matter">some differences</a> between SLMs and LLMs. First, the LLMs are bigger in size and have undergone more widespread training when weighed with SLMs. Second, the LLMs have notable natural language processing abilities, making it possible to capture complicated patterns and outdo in natural language tasks, for example complex reasoning. Finally, the LLMs can understand language more thoroughly while, SLMs have restricted exposure to language patterns. This does not put SLMs at a disadvantage and when used in appropriate use cases, they are more beneficial than LLMs.</p><p>There are several reasons to use these models. They are gaining popularity and relevance in various applications especially with regards to sustainability and amount of data needed for training. From the hardware point of view, it is cheaper to run i.e., SLMs require less computational power and memory and it is suitable for on-premises and on-device deployments making it more secure. From the usage point of view, these are small language models, that are trained or fine-tuned for particular domains or tasks, hence they can have specialized lingo and knowledge from legal jargons to medical diagnoses protecting intellectual property. Depending on the scenario, SLMs would be cheaper and efficient.</p><p>SLMs find applications in a wide range of sectors, spanning healthcare to technology, and beyond. The common use cases across all these industries include summarizing text, generating new text, sentiment analysis, chatbots, recognizing named entities, correcting spelling, machine translation, code generation and others.</p><p>Language model fine-tuning is a process of providing additional training to a pre-trained language model making it more domain or task specific. This process involves updating the model’s parameters with additional training data to improve its performance in specific areas or applications such as text generation, question answering, language translation, sentiment analysis, and others. We are interested in ‘domain-specific fine-tuning’ as it is especially useful when we want the model to understand and generate text relevant to specific industries or use cases.</p><h4><strong>Hardware Requirements</strong></h4><p>The hardware requirements may vary based on the size and complexity of the model, the scale of the project, and the dataset. It’s a good practice to start with a small-scale and then scale up as necessary. However, here are some general guidelines for fine-tuning a private language model.</p><ol><li>GPUs (Graphics Processing Units) for processing. It could be cloud-based.</li><li>A fast and reliable internet connection for data transfer.</li><li>A powerful multi-core CPU for data pre-processing, and managing distribution steps.</li><li>Sufficient memory, and fast and ample storage.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/676/1*mcwrzXI44hMUwUoiV7HJxQ.png" /><figcaption>Figure 1. The virtual machine used for our fine-tuning process.</figcaption></figure><h4><strong>Data Preparation</strong></h4><p>The quality and feasibility of your dataset significantly impact the performance of the fine-tuned model. For our goal in this phase, we need to extract text from PDF’s, to clean and prepare the text, then we generate question and answers pairs from the given text chunks. Finally, proceed with the fine-tuning process.</p><p>It is to be noted that, we used a LLM such as GPT-3.5 for generating the Q&amp;A pairs (which might defeat the purpose here), however, we could try with SLMs here as well to generate these pairs depending on the use case.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/491/1*iUzCHcUejxkJNJ_e69Q2Zg.png" /><figcaption>Figure 2. Key steps involved in preparing a dataset for fine-tuning.</figcaption></figure><h4><strong>Fine-tuning Process</strong></h4><p>We used HuggingFace, its full suite of components, and integrated them to accomplish this task.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/443/1*g-TX1zBLiS9MeEx4VXE-ew.png" /><figcaption>Figure 3. Components integrated for fine-tuning.</figcaption></figure><p>The pre-trained language <strong>model</strong> Llama-2–13b-chat-hf model was chosen. For the domain-specific <strong>dataset</strong>, we converted into HuggingFace datasets type and used the <strong>tokenizer</strong> accessible through the HuggingFace API. In addition, <strong>quantization</strong> used to reduce the precision of numerical values in a model allowing, data compression, computation and storage efficiency and noise reduction. <strong>Performance configuration</strong> was also enabled for efficient adaptation of pre-trained models. Finally, <strong>training arguments</strong> were used for defining particulars of the training process and the <strong>trainer</strong> was passed parameters, data, and constraints.</p><h4><strong>Training Process</strong></h4><p>We conducted a 50-epoch fine-tuning of the model. An epoch refers to one full cycle through the training dataset. It required about 16 hours to complete, and our CPU and RAM resources were not fully utilized during the process. It’s possible that a machine with limited CPU and RAM resources might suit the process. Our GPU usage aligns with the stated model requirements; perhaps increasing the batch size could accelerate the training process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZFxxqi9MUjmfsiaAW6N7Wg.png" /><figcaption>Figure 4. CPU and RAM usage.</figcaption></figure><p>Overall, despite the initial challenges of understanding the interconnections and facing several unsuccessful attempts, the fine-tuning process appeared to run smoothly and consistently. The monetary cost for this fine-tuning process was around $100/£83. However, this cost above did not include the cost of all trials and errors that concluded to the final fine-tuning process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2lOJCYOTYYr6Rz8tVDHz6g.png" /><figcaption>Figure 5. Cost of fine-tuning in GBP.</figcaption></figure><h4><strong>Results and Observations</strong></h4><p>Please note that we used GPT-3.5 to generate questions and answers from the training data. The model that we fine-tuned is Llama-2–13b-chat-hf has only 13 billion parameters while GPT-3.5 has 175 billion. In other words, we are expecting a small model to perform as well as a large one. Therefore, due to GPT-3.5 and Llama-2–13b-chat-hf difference in scale, direct comparison between answers was not appropriate, however, the answers must be comparable.</p><p>Embedding were created for the answers generated by the SLM and GPT-3.5 and the cosine distance was used to determine the similarity of the answers from the two models.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tkRWVGMHLjHCO9U6U6awbQ.png" /><figcaption>Figure 6. Similarity distribution of GPT-3.5 answers and Llama-2–13b-chat-hf answers.</figcaption></figure><p>According to Figure 6, 0.5 was established as the cut-off for quality and 0.6 represents the average quality of the result produced by Llama-2–13b-chat-hf. Anything above 0.5 was considered acceptable and below unacceptable. This is because, as the similarly ranges from -1 being opposite, 1 being an exact match, and 0 being unrelated to the value of 0.5, which seems reasonable argument.</p><p>For the fine-tuning process, we use about 10,000 question-and-answer pairs generated from the Version 1’s internal documentation. But for evaluation, we selected only questions that are relevant to Version 1 and the process. Further analysis of the results showed that, over 70% are strongly similar to the answers generated by GPT-3.5, that is having similarity 0.5 and above (see Figure 6). In total, there are 605 considered to be acceptable answers, 118 somewhat acceptable answers (below 0.4), and 12 unacceptable answers.</p><p>The fine-tuned model seems to competent at extracting and maintaining knowledge while demonstrating the ability to generate answers to the specific domain. A platform agnostic approach allowed us to execute the same fine-tuning processes on AWS and achieve almost identical results without any changes to the code.</p><h3><strong>Conclusions</strong></h3><p>There are some drawbacks to SLMs. A single constant running instance of this system will cost approximately $3700/£3000 per month. The knowledge bases are more limited than their LLM counterparts meaning, it cannot answer questions like who walked on the moon and other factual queries. Due to the narrow understanding of language and context it can produce more restricted and limited answers. Although, the future looks somewhat bright for SLMs in their own merits. The voyage of language models highlights a fundamental message in AI, i.e., small can be impressive, assuming that there is constant advancement and modernization. In addition, there is an understanding that efficiency, versatility, environmentally friendliness, and optimized training approaches grab the potential of SLMs.</p><p>We would have to wait and see how popular SLMs will become compared to their LLM counterparts especially with the recent launch of SLMs, such as Gemini Nano, <a href="https://mistral.ai/news/mixtral-of-experts/">Mixtral</a>, <a href="https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/">Phi-2</a>, and others.</p><p><strong>Note to the reader:</strong></p><p>Alexander Suvorov, our Senior Data Scientist conducted the fine-tuning processes of Llama 2.</p><p><strong>About the Author:</strong></p><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b793bdb002f2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/small-but-powerful-a-deep-dive-into-small-language-models-slms-b793bdb002f2">Small but Powerful: A Deep Dive into Small Language Models (SLMs)</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meet Gemini: Google’s Multimodal Masterpiece that can Push AI Boundaries]]></title>
            <link>https://medium.com/version-1/meet-gemini-googles-multimodal-masterpiece-that-can-push-ai-boundaries-dc16d23803a3?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/dc16d23803a3</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gemini]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Fri, 08 Dec 2023 16:12:00 GMT</pubDate>
            <atom:updated>2024-09-23T08:16:56.581Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XB0Pnr08nBbFwWn1EryopQ.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>On 6th December 2023, Google announced <a href="https://blog.google/technology/ai/google-gemini-ai/#introducing-gemini">Gemini</a>, the next generation foundation model. It looks like Google is grabbing the market advantage against their competitor, OpenAI to challenge GPT-4 and build a better a model. Gemini claims to be the most versatile model can be the efficiently be deployed in wide range of devices from mobile to data centres. Google claims to be <a href="https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf">responsible</a> while developing all their AI models.</p><p><a href="https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf">Gemini</a> is a new family of multimodal and multilingual models from Google DeepMind that achieve state-of-the-art performance across text, image, audio, and video understanding. Its family consists of Ultra, Pro and Nano sizes suitable for different and wider range of applications. In short, Ultra is the most capable, Pro is optimised for performance and scale, and Nano is for on-device deployment. These models are trained on a large dataset of web documents, books, code, images, audio, and video. These models can process interleaved inputs across modalities.</p><p>Gemini is built upon the advancements made in PaLM 2, their previous model. There are some major improvements reflected in Gemini. First, Gemini is designed to be multimodal and multilingual from the foundation upwards which means it has the ability to understand and generate not just text, but also images and other types of media. Second, Gemini is highly competent at integrating with various tools and APIs. This allows it to interact with a wide range of applications and devices, enhancing its versatility and usability. Commencing December 13, Gemini Pro can be used by developers and enterprise customers via the <a href="https://blog.google/technology/ai/google-gemini-ai/#availability">Gemini API</a> in Google AI Studio or Google Cloud Vertex AI. Third, Gemini is designed with future inventions in mind, such as memory and planning which suggests that it might be capable of more intricate tasks and reasoning, building upon PaLM’s reasoning capabilities. Finally, before the public release, Gemini has been rigorously tested in line with responsible AI practices which makes sure that it can be implemented across various products, applications, and devices for wider societal good.</p><blockquote><a href="https://www.version1.com/application-modernisation-integration/ai-powered-code-analysis-and-documentation-decipher/">AI-powered code analysis and documentation — Decipher | Version 1</a></blockquote><h4>The Making of Gemini</h4><p>Gemini was trained using multiple sets of TPUv5 accelerators, across several data centres. A surplus of in-memory copies of the model state were operated, and on any unplanned hardware breakdowns, recovery was made directly from a complete model duplication. This is a massive expansion and upgrade compared to their prior flagship model PaLM-2. The end-product being a general goodput for the bulkiest training job swelling from 85% to 97%. The magnitude of Gemini’s operations means that Silent Data Corruption (SDC) incidents can be anticipated to crash training every week or two. Their completely deterministic infrastructure permits rapid detection of the source (eg. hardware breakdowns) during the development which made a massive contribution concerning stable training.</p><p>Impact assessments, mitigation research, model policies, and external evaluations around safety, fairness, and robustness were included in the responsible AI development process of Gemini.</p><h4>Main Features of Google Gemini</h4><p><em>Mastering Human-Style Conversations, Language, Content, and Image Interpretation: </em>Gemini Ultra exceeds human-expert achievement by securing 90% on the test benchmark Massive Multitask Language Understanding (MMLU) demonstrating Gemini’s competence to master human-style conversations, language, and content. MMLU was used to determine progress for Language Models since it was announced in 2020. Additionally, Gemini models can describe intricate images including, charts, infographics etc. and explain interleaved chains of images, audio, and text, demonstrating its ability to interpret images.</p><p><em>Driving Data Analytics and Creating New AI Apps and APIs:</em> There is a blanket of possibilities of new apps with the innovative capabilities of Gemini models. Gemini can foster new advances in themes including education, everyday problem-solving, multilingual communication, information summarisation, extraction, and creativity. This shows Gemini’s possibility in driving data analytics. Developers can use these models when available to create new AI apps and APIs, leading to AI for good.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FUIZAiXYceBI&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DUIZAiXYceBI&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2FUIZAiXYceBI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/3ea92f7dbe4388247d611937103898a4/href">https://medium.com/media/3ea92f7dbe4388247d611937103898a4/href</a></iframe><h4>Gemini vs. Other LLMs</h4><p>To date, OpenAI’s GPT models, like GPT-3 and GPT-4, and Claude are some of the largest language models. Gemini establishes new state-of-the-art (SOTA) performance results across major language and text benchmarks, often with zero-shot evaluation. In direct parallels, Gemini models has a close match or go beyond GPT performance on major language benchmarks while having additional multimodal capabilities (see Figure 1 below). Gemini does markedly well on more intricate reasoning tasks compared to GPT. For example, in math and coding tests. Gemini is also developed to be safer, more helpful and aligned with human preferences. Early findings suggest advantages over GPT here. So, in brief, Gemini aims to be competitive with the best existing LLMs like GPT-4, while having native multimodality, stronger reasoning, and more oversight into safety. The goal is to advance towards more generally capable AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mCy9y6z3c-wo8-zN6SVdTQ.png" /><figcaption>Figure 1. <a href="https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf">Gemini performance on text benchmarks with other models</a></figcaption></figure><h4>Future of Gemini</h4><p>It looks like <a href="https://blog.google/technology/ai/google-gemini-ai/#availability">Gemini was developed</a> to boost Google’s enterprise-based products. For example, products like Google Docs, Slides, Chrome, Meet and more could gain from Gemini’s sophisticated language comprehension and generation capabilities. This could spearhead more supportive processes, tailored customer dealings, and elevated modernisations. When developers will be able to access the Gemini AI app via Google Cloud, it could lead to the inventions and revolution of new AI apps in the market.</p><p>The future work looks to revolve around developing Gemini models into a modularised system with general cross-modal reasoning and understanding capabilities. The reported benchmarks seem to show its extraordinary capabilities, but limitations around LLMs reasoning and hallucinations linger. More robust evaluations and research to advance LLMs is needed due to overloaded current benchmarks.</p><p>With the broad range of applications of Google Gemini in the field of AI we are yet to uncover its real-world capacities that is, how it can achieve intelligent solutions, promote scientific progress, and improve human welfare.</p><h4>About the author</h4><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher here at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dc16d23803a3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/meet-gemini-googles-multimodal-masterpiece-that-can-push-ai-boundaries-dc16d23803a3">Meet Gemini: Google’s Multimodal Masterpiece that can Push AI Boundaries</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Really Matters for Green Calculations: A Practical Perspective]]></title>
            <link>https://medium.com/version-1/what-really-matters-for-green-calculations-a-practical-perspective-0bc0f5c7540c?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/0bc0f5c7540c</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[ethical-ai]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Wed, 29 Nov 2023 09:50:25 GMT</pubDate>
            <atom:updated>2023-11-29T12:52:02.654Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XCw4fpPkbM4Z7dwkcEse5Q.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>Since joining the UN Global Compact, Version 1 has been proud to be part of a global movement of sustainable companies. We are integrating a principles-based approach to sustainability into all our business operations and our AI Labs are no exception.</p><p>So how green is our AI? This is a question we are asking ourselves, together with many business leaders, researchers, practitioners, and policymakers are asking. AI models, algorithms and systems have increasingly become powerful and universal in various domains, such as generative AI, natural language processing, computer vision, speech technology, and more. However, these models also come with a significant environmental cost, as they need large datasets and computational resources to train and run.</p><p>For instance, <a href="https://www.scientificamerican.com/article/a-computer-scientist-breaks-down-generative-ais-hefty-carbon-footprint/">to create GPT-3</a>, a huge AI model with 175 billion parameters, 1,287 megawatt hours of electricity were used and 552 tons of carbon dioxide equivalent were emitted, which is the same as driving 123 petrol-powered cars for a year. Can you imagine how much would it be for the latest Claude model with 200k token limit? Therefore, it is important to consider the sustainable aspects of AI models, such as their energy efficiency, carbon emissions, and environmental impact. We aim to provide a brief overview of the current green calculations for AI models and their progress so far, which are essential for the long-term sustainability and development of AI and humanity.</p><p>Various tools and methods exist to help us calculate environmental impact. Green Software Foundation (GSF) is a non-profit organisation that strives to establish a reliable network of people, standards and best practices for building software and AI that are environmentally friendly. First, let us investigate some of the projects of measurements by GSF: Impact Framework, Software Carbon Intensity specification and Green Software Maturity Matrix.</p><p>The <a href="https://if.greensoftware.foundation/">Impact Framework</a> is a project by GSF that allows anyone running a model, application or process to estimate their energy and carbon impact. It stresses that measurement is not the end-result but should be able to guide where we should make advances and how quickly. By providing some basic configuration data in a simple manifest file, you can get energy and carbon usage estimates. This is in the alpha stage and a promising development towards transparent sustainability measurements with extreme flexibility for any organisation.</p><p>The GSF also developed the <a href="https://learn.greensoftware.foundation/measurement/#the-sci-equation">Software Carbon Intensity</a> (SCI) specification to score any model, software or application, providing a common language to describe its carbon emissions behaviour. The SCI equation is simple and applicable to various scenarios. The equation is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/883/1*frsBWofiCNx4F2hMHsV6Dw.png" /></figure><p>The equation SCI=RC quantifies the Sustainability of Carbon Intensity (SCI), a measure of carbon emissions for a given process. In this equation, R represents a core characteristic of the SCI, serving as a functional unit that converts the SCI from a total measure to an intensity measure. Meanwhile, C stands for the Carbon emissions per unit of R. Thus, this equation provides a more detailed view of carbon emissions, focusing on their intensity rather than their total amount.</p><p>A unique feature of the SCI is that it cannot be reduced by purchasing carbon offsets such as neutralisations, compensations, or renewable energy credits. This means that organisations can’t simply buy carbon credits to lower their SCI. They must make legitimate efforts to reduce their emissions. The SCI promotes the development of models that are more energy-efficient, hardware-efficient, or carbon-aware. This equation has been found to be effective in <a href="https://sci-guide.greensoftware.foundation/CaseStudies/">case studies</a> and practical applications, demonstrating its utility in promoting sustainability.</p><p>GSF has also developed the <a href="https://github.com/Green-Software-Foundation/green-software-maturity-matrix">Green Software Maturity Matrix</a>, which is a self-assessment tool in the alpha stage designed to help organisations understand the extent to which they have applied green principles, patterns, and processes for building and operating their models. The tool addresses four key questions and attempts to respond to: ‘Where do I start?’, ‘Where am I now?’, ‘What am I heading towards?’, and ‘How do I know I am making progress?’. This seems to be a good foundation tool to start the sustainability conversation in your organisation.</p><p>Next, let’s investigate <a href="https://medium.com/bcggamma/ai-computing-emits-co%E2%82%82-we-started-measuring-how-much-807dec8c35e3">CodeCarbon</a>, an open-source tool developed to track the carbon footprint of energy consumed during the execution of AI algorithms, models and systems. It is a light package that integrates smoothly into Python code. By recording power usage and estimating carbon emissions based on the energy grid’s carbon intensity, the tool provides developers with visibility into the sustainability of their code. It also includes a dashboard to weigh carbon emissions across geographies where cloud infrastructure is hosted. This is a step towards greater transparency and consciousness in the AI community, enabling users to measure and report the emissions generated by various AI algorithms, models and systems.</p><p>Finally, let’s investigate the two significant equations that are <a href="https://www.linkedin.com/pulse/reducing-carbon-footprint-generative-ai-boris-gamazaychikov">instrumental in understanding the energy consumption and carbon emissions</a> associated with computational processes.</p><p>The first equation calculates the energy consumption in kilowatt-hours (kWh) for training models over a certain period. This is specifically used to calculate energy consumption during the training of models. However we could use this for inference too i.e. ‘Hours to Train’ can be replaced by ‘Hours of Execution’. The former equation is:</p><pre>kWh = Hours to Train × Number of Processors × Average Power per Processor (in Watts) × PUE / 1000</pre><blockquote>Where:</blockquote><blockquote>Hours to Train is the time taken for the training process in hours.</blockquote><blockquote>Number of Processors is the total number of processors used.</blockquote><blockquote>Average Power per Processor is the average power consumption of each processor in watts.</blockquote><blockquote>PUE (Power Usage Effectiveness) is a measure of how efficiently a computer data centre uses energy; specifically, how much energy is used by the computing equipment.</blockquote><blockquote>The division by 1000 is to convert the power from watts to kilowatts, as 1 kilowatt is equal to 1000 watts.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/745/1*va1mN264AOMBqnkzWZE4LA.png" /><figcaption>Figure 1. <a href="https://www.linkedin.com/pulse/reducing-carbon-footprint-generative-ai-boris-gamazaychikov">Emissions from Prominent Language Models during Training</a></figcaption></figure><p>The second equation calculates the carbon dioxide equivalent emissions (tCO2e) resulting from energy consumption measured in kilowatt-hours (kWh).</p><p>The equation is:</p><pre>tCO2e = kWh × kg CO2e per kWh / 1000</pre><blockquote>Where:</blockquote><blockquote>kWh is the energy consumption in kilowatt-hours.</blockquote><blockquote>kg CO2e per kWh is the carbon intensity factor, which represents the amount of CO2 equivalent emissions produced per kilowatt-hour of energy consumed.</blockquote><blockquote>The division by 1000 is to convert the emissions from kilograms to metric tonnes, as 1 metric tonne is equal to 1000 kilograms.</blockquote><p>Both these equations provide a simplified model, and actual energy consumption and carbon emissions can be influenced by various other factors. Nonetheless, they offer a valuable starting point for understanding the environmental impact of computational processes.</p><p>There is an important factor to consider when investigating into these tools and equations is that organisations can decide what to include and exclude in the calculation depending on their sustainability objective or plans. However, when we need data that is publicly available in order to make these calculations. The data availability might be restricted to certain publicly available models for now. We have to wait and see how open businesses like Microsoft, AWS and others are likely to share this information. However, things might change when new regulations and laws come into place.</p><p>Let’s conclude with <a href="https://hbr.org/2023/07/how-to-make-generative-ai-greener">some strategies</a> that can be exercised to create greener models. By implementing these strategies, we can make strides towards a more sustainable future with AI.</p><p><em>Develop Computationally Efficient Models:</em> Strive to create models that require less computational power, thereby reducing energy consumption.</p><p><em>Reuse Existing Models:</em> Whenever possible, use pre-existing models instead of training new ones from scratch. This approach can save both time and energy.</p><p><em>Optimise Model Parameters:</em> Efficient use of parameters can lead to optimised models that perform well without excessive energy use.</p><p><em>Choose Energy-Efficient Hardware:</em> Select cloud providers who utilise renewable energy sources. This choice can significantly reduce the carbon footprint of your AI operations.</p><p><em>Track and Report Carbon Footprints:</em> Use AI to monitor and report variables that affect carbon footprints. This data can provide valuable insights for further optimisation.</p><h4>About the author</h4><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0bc0f5c7540c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/what-really-matters-for-green-calculations-a-practical-perspective-0bc0f5c7540c">What Really Matters for Green Calculations: A Practical Perspective</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Exploring AWS Bedrock: Data Storage, Security and AI Models]]></title>
            <link>https://medium.com/version-1/exploring-aws-bedrock-data-storage-security-and-ai-models-6a22032cee34?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a22032cee34</guid>
            <category><![CDATA[aws-bedrock]]></category>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Tue, 17 Oct 2023 08:18:17 GMT</pubDate>
            <atom:updated>2024-09-23T08:12:18.669Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/297/0*i-L9c2IgC3N3XxzM" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>Amazon Web Services (AWS), a global leader in cloud computing, has recently launched a new service known as <a href="https://aws.amazon.com/bedrock/">Amazon Bedrock</a>. This service is fully managed and designed to streamline the process of developing generative AI applications.</p><p>Amazon Bedrock provides access to a selection of high-performance foundation models (FMs). These <a href="https://aws.amazon.com/about-aws/whats-new/2023/09/amazon-bedrock-generally-available/">models are the product of collaboration with leading AI companies</a> such as AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon itself. The service offers a unified API, making it easier for developers to leverage these models in their applications.</p><p>Despite simplifying the development of generative AI application process, it does not compromise on privacy and security aspects. This ensures that developers can create powerful AI applications while maintaining the integrity and confidentiality of their data.</p><p>In this article, we delve into the fundamental aspects of AWS’s security measures, and the availability and pricing overview of the models.</p><blockquote><a href="https://www.version1.com/application-modernisation-integration/ai-powered-code-analysis-and-documentation-decipher/">AI-powered code analysis and documentation — Decipher | Version 1</a></blockquote><h3><strong>Security Implications</strong></h3><p>Security is a chief concern when it comes to AWS Bedrock. The platform is designed with <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/security.html">security and privacy at its core</a>, leveraging the robust security features of other AWS services. This ensures the safety and privacy of data, with comprehensive data encryption guaranteeing the secure management and protection of sensitive information.</p><p>Amazon Bedrock employs <a href="https://aws.amazon.com/bedrock/faqs/">rigorous encryption measures and storage practices within the AWS Region where it operates</a>, safeguarding user data at rest. Crucially, data inputs and model outputs are not disclosed to third-party model providers for training, ensuring confidentiality like Microsoft Azure.</p><p>Amazon Bedrock empowers users to <a href="https://aws.amazon.com/bedrock/pricing/">engage in private customisation of only Titan Text Lite and Titan Text Express FMs</a>, granting them control over how their data is utilised and encrypted. The system creates an isolated copy of the base FM, on which private model training takes place. Notably, user data such as prompts, supplementary information, FM responses, and customised FMs are retained within the region where the API requests are processed.</p><p>What sets Amazon Bedrock apart is its ability to refine and fine-tune models for specific tasks without the need for extensive data annotation. It then generates a distinct private version of the FM, accessible exclusively by the user, and trains it independently. Importantly, none of the user data contributes to the training of the original base models. Amazon Bedrock can be configured with Amazon Virtual Private Cloud (VPC) and AWS Identity and Access Management (IAM).</p><p>Moreover, users do not have to manage any infrastructure as Amazon Bedrock is serverless. This means that the <a href="https://www.aboutamazon.com/news/aws/how-aws-partners-are-advancing-generative-ai-for-government-health-care-and-other-public-sector-organizations">integration and implementation of generative AI capabilities</a> into desired applications can be done securely using established AWS services.</p><p>Bedrock’s <a href="https://docs.aws.amazon.com/bedrock/latest/userguide/abuse-detection.html">abuse detection systems are automated</a>, ensuring no human oversight is involved in reviewing or accessing user inputs or model outputs. However, as of now, there is no available information about opting out of data usage. In contrast, Azure’s abuse monitoring processes are similarly automated, but they <a href="https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy">do offer their customers the option to o</a>pt-out.</p><h3><strong>Models Availability</strong></h3><p>You’ll find a summary table of AWS Bedrock FMs at the end for reference. If you’re keen on delving deeper into the FMs (Amazon Titan, AI21 Labs Jurassic, Anthropic Claude, Cohere Command, Meta Llama 2, and Stable Diffusion XL), the section below provides more details.</p><h4><a href="https://aws.amazon.com/bedrock/titan/">Amazon Titan</a></h4><p>There are three versions of <a href="https://aws.amazon.com/bedrock/titan/">Amazon Titan</a>.</p><p>Titan Text Express (preview) offers an offset between price and performance. It supports a maximum of 8K tokens and over 100 languages. The supported use cases include retrieval augmented generation, open-ended text generation, brainstorming, summarisation, code generation, table creation, data formatting, paraphrasing, chain of thought, rewrite, extraction, Q&amp;A, and chat.</p><p>Titan Text Lite (preview) is reasonably priced and small, ideal for basic tasks and fine-tuning. It supports English and has a maximum of 4K tokens. The supported use cases include open-ended text generation, brainstorming, summarisation, code generation, table creation, data formatting, paraphrasing, chain of thought, rewrite, extraction, Q&amp;A, and chat.</p><p><a href="https://www.aboutamazon.com/news/aws/aws-amazon-bedrock-general-availability-generative-ai-innovations">Titan Embeddings (generally available)</a> that translates text into a numerical representation. It supports a maximum of 8K tokens, over 25 languages, and 1536 embeddings. This is the only embedding model available on AWS Bedrock presently. The supported use cases include text retrieval, semantic similarity, and clustering.</p><h4><a href="https://aws.amazon.com/bedrock/jurassic/">AI21 Labs Jurassic</a></h4><p>There are two versions of the <a href="https://aws.amazon.com/bedrock/jurassic/">Jurassic model</a> optimised for zero-shot. It can be accessed through Amazon Bedrock to build generative AI applications and is available through AI21 Studio. It is the <a href="https://www.prnewswire.com/news-releases/ai21-labs-introduces-jurassic-2-the-worlds-most-advanced-language-model-301767781.html">world’s most customisable model with 5 APIs</a> i.e. APIs are designed for effortless customisation, in stark contrast to other models that rely solely on the base API and prompt engineering. It supports a maximum of 8192 tokens and languages including English, Spanish, French, German, Portuguese, Italian, and Dutch.</p><p>Jurassic-2 Ultra is AI21’s most powerful model, offering remarkable quality. The supported use cases include question-answering, summarisation, draft generation, advanced information extraction, and ideation for tasks requiring intricate reasoning and logic.</p><p><a href="https://aws.amazon.com/marketplace/pp/prodview-ivk4wjg6gbr26">Jurassic-2 Mid</a> is a mid-sized model that strikes the true balance between remarkable quality and affordability. The supported use cases include question-answering, summarisation, draft generation, advanced information extraction, and ideation. It is also available on the AWS Marketplace and optimised to follow natural language instructions and context without requiring any instances. It’s a pinnacle in Stanford’s Holistic Evaluation of Language Models (HELM), standing notably advanced than models up to 30x bigger in size. This allows users to optimise production costs and speed without forgoing quality.</p><h4><a href="https://aws.amazon.com/bedrock/claude/">Anthropic Claude</a></h4><p>There are three versions of <a href="https://aws.amazon.com/bedrock/claude/">Anthropic Claude</a>. It supports a maximum of 100K tokens and multiple languages including English.</p><p>Claude 1.3 is an earlier version of their general-purpose model. The supported use cases include searching, writing, editing, outlining, and summarising text; coding; and providing helpful advice about a broad range of subjects.</p><p>Claude Instant is their faster, reasonably priced yet very efficient model. The supported use cases include casual dialogue, text analysis, summarisation, and document comprehension.</p><p><a href="https://www.anthropic.com/index/claude-2">Claude 2</a> is the most efficient system released up to date. The supported use cases include thoughtful dialogue, content creation, complex reasoning, creativity, and coding. Claude 2 excels at the core capabilities of Claude 1.3 and Claude Instant.</p><p>We have an upcoming blog on this with our analysis.</p><h4><a href="https://aws.amazon.com/bedrock/cohere-command/">Cohere Command</a></h4><p><a href="https://aws.amazon.com/bedrock/cohere-command/">Command </a>is Cohere’s only model with 52B parameters. It supports a maximum of 4K tokens and English language. The supported use cases include chat, text generation, and text summarisation. According to the HELM benchmark, Command <a href="https://cohere.com/models/command">stands midst the world’s finest language models</a>. It is also available through Cohere’s subscription for enterprises. They have a <a href="https://www.oracle.com/news/announcement/oracle-to-deliver-powerful-and-secure-generative-ai-service-for-business-2023-06-13/">partnership with Oracle</a>.</p><h4><a href="https://aws.amazon.com/bedrock/llama-2/">Meta Llama 2</a></h4><p>There are two versions of the fine-tuned <a href="https://aws.amazon.com/bedrock/llama-2/">Meta Llama 2</a>. It supports a maximum of 4K tokens and English. The supported use case is an assistant-like chat. The only difference between Llama-2–13b-chat and Llama-2–70b-chat is that the <a href="https://aws.amazon.com/about-aws/whats-new/2023/09/amazon-bedrock-generally-available/">former has a parameter size of 13B while the latter parameter size of 70B</a>.</p><h4><a href="https://aws.amazon.com/bedrock/stable-diffusion/">Stable Diffusion XL</a></h4><p>There are two versions of <a href="https://aws.amazon.com/bedrock/stable-diffusion/">Stable Diffusion XL</a> with a <a href="https://stability.ai/press-articles/stable-diffusion-xl-1-featured-amazon-aws-bedrock?format=amp">3.5B parameter base model and a 6.6B parameter model ensemble pipeline</a>. It supports English and has a maximum of 77 tokens. The supported use cases include advertising and marketing, media and entertainment, gaming and metaverse. It’s also available through Stability AI API, Stability AI’s Github page and its Clipdrop and DreamStudio consumer applications.</p><p>Stable Diffusion XL 1.0 (preview) is the most advanced text-to-image model from Stability AI while Stable Diffusion XL 0.8 (preview) is an earlier version of the text-to-image model from Stability AI.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kx3ntIVAo7ozaW823azrdA.png" /><figcaption>Table 1. Summary of AWS Bedrock FMs</figcaption></figure><p>All these models are pre-trained on large datasets making them powerful tools for a wide range of applications.</p><h3>Models Pricing</h3><p>There are two pricing consumption plans for inference:</p><p><em>On-Demand:</em> With this plan, access to FMs is on a flexible pay-as-you-go basis without the need for long-term commitments.</p><p><em>Provisioned Throughput:</em> This plan enables the allocation of the necessary throughput resources to meet your application’s performance needs in exchange for committing to a specific time-based term.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/935/1*4H4T7MBR_2XyoIrOueyedQ.png" /><figcaption>Table 2. Overview of key pricing elements (full pricing details available <a href="https://aws.amazon.com/bedrock/pricing/">here</a>)</figcaption></figure><h3><strong>Conclusion</strong></h3><p>Amazon Web Services’ Bedrock represents an immense advancement in the orbit of generative AI. It provides a secure environment for deploying generative AI models, with robust data encryption, secure access management via IAM, and a serverless architecture that eliminates the need for users to manage infrastructure. Importantly, user inputs and model results are kept confidential and not shared with third-party model providers. However, there is currently abuse monitoring is automated, and there is currently no information available about the ability to opt out of data usage.</p><p>The service’s rigorous security protocols are designed to uphold data privacy, a feature that is increasingly important in today’s digital age. Furthermore, Bedrock provides access to an array of potent foundation models. These models, developed by leading AI companies, can address a wide range of use cases.</p><p>Given these features, AWS Bedrock emerges as an optimal choice for businesses aiming to harness the power of AI. It offers a comprehensive solution that balances ease of use with advanced capabilities, making it an invaluable tool in the modern business landscape.</p><p>This article was written with the help of ChatGPT and Bing Chat.</p><h4>About the author</h4><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a22032cee34" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/exploring-aws-bedrock-data-storage-security-and-ai-models-6a22032cee34">Exploring AWS Bedrock: Data Storage, Security and AI Models</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Moral Compass in the Machine: Ethics in AI]]></title>
            <link>https://medium.com/version-1/the-moral-compass-in-the-machine-ethics-in-ai-64374b30d4b4?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/64374b30d4b4</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[explainable-ai]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Fri, 13 Oct 2023 07:07:30 GMT</pubDate>
            <atom:updated>2024-12-17T09:51:06.976Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/297/0*UHMCENDc2dWf7rAM" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>AI ethics, or machine ethics, is an <a href="https://onlinelibrary.wiley.com/doi/pdf/10.1609/aimag.v28i4.2065">interdisciplinary field</a> focused on addressing the moral aspects of artificial intelligence, involving study of <a href="https://scholarsmine.mst.edu/cgi/viewcontent.cgi?article=1356&amp;context=bio_inftec_facwork">ethical theories, guidelines, policies, and regulations specific to AI</a>. Additionally, it explores the concept of ethical AI, which involves artificial intelligence systems capable of upholding ethical standards and behaving in a conscientious and trustworthy manner.</p><p>Establishing an ethical foundation for AI is necessary for both <a href="https://www.gov.uk/guidance/understanding-artificial-intelligence-ethics-and-safety">developing systems that adhere to ethical norms and ensuring that AI operates in an ethically sound manner</a>. This endeavour hinges on defining the ethical and moral values and principles that discern what is deemed moral or harmful. With <a href="https://www.coursera.org/articles/ai-ethics">high-tech innovations in AI ethics</a> and through the application of suitable ethical frameworks, one can construct and apply systems that exhibit ethical behaviour, using various methodologies and technologies.</p><p>In this article, we explore the primary ethical challenges arising from AI and examine best practices for effectively addressing them.</p><h3>Major ethical issues caused by artificial intelligence.</h3><h4><strong>Biases</strong></h4><p>Some of the prominent types of biases are discussed below.</p><p><a href="https://dl.acm.org/doi/pdf/10.1145/3466132.3466134">Sampling bias</a> occurs when the training data used for machine learning algorithms does not adequately represent the entire population it is meant to serve, lacks diversity, resulting in inaccurate and biased predictions.</p><p><a href="https://www.isaca.org/resources/isaca-journal/issues/2022/volume-4/bias-and-ethical-concerns-in-machine-learning">Algorithmic bias</a> arises from unequal treatment of datasets, due to biased assumptions during AI algorithm development or inherent prejudices in the training data. It results from the algorithm design and execution, favouring certain attributes, ultimately leading to unjust outcomes.</p><p><a href="https://www.forbes.com/sites/forbestechcouncil/2023/09/06/navigating-the-biases-in-llm-generative-ai-a-guide-to-responsible-implementation">Confirmation bias</a> can manifest either in the training data or in the way prompts are framed, influencing the AI’s responses. This means that when users request information or responses from AI, it may tend to generate content that aligns with the users’ existing viewpoints.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/297/0*GH1HZ5X_U_e_Ks2O" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p><a href="https://atrium.ai/resources/ethical-ai-real-world-examples-of-bias-and-how-to-combat-it/">Measurement bias</a> is associated with flawed sensors or devices, can also arise when human judgment and perspectives influence data collection, it favours or underrepresents specific groups.</p><p><a href="https://www.forbes.com/sites/forbestechcouncil/2023/06/23/the-ethics-of-ai-navigating-bias-manipulation-and-beyond">Interaction bias</a> occurs when biased data is employed to train AI models, the outputs generated by these models will inherently carry those biases It can magnify these biases, and they can inadvertently perpetuate societal prejudices.</p><h4><strong>Transparency and explainability</strong></h4><p><a href="https://www.mdpi.com/2673-2688/4/1/3">Transparency is a key issue in AI ethics</a>, particularly in deep learning algorithms that often operate as “black boxes,” obscuring their decision-making processes. This opacity becomes more problematic when AI models are trained on biased data, potentially leading to unfair outcomes. Trust in AI, especially in sensitive fields like healthcare, relies on transparency because it affects public and patient trust and the adoption of AI technologies.</p><p><a href="https://www.mdpi.com/2673-2688/4/1/3">Explainability</a> is another ethical concern, closely related to transparency. Likewise, it’s hindered by the black-box nature of many AI algorithms, making it <a href="https://www.peertechzpublications.com/articles/ARA-5-111.php">difficult to explain the relationship between input data and results</a>. Deep learning models, like Neural Networks, are often considered “black boxes” due to their complex, less transparent processes.</p><h4><strong>Privacy</strong></h4><p>The <a href="https://dataconomy.com/2023/01/16/artificial-intelligence-security-issues/">main AI privacy and security concerns</a> revolve around personal data handling, which poses risks such as intentional breaches or accidental leaks, potentially leading to identity theft, fraud, and other forms of abuse. Additionally, as AI systems become more complex and autonomous, they are vulnerable to hacking or manipulation, potentially leading to harmful decisions.</p><p><a href="https://onlinedegrees.sandiego.edu/ethics-in-artificial-intelligence/">Facial recognition</a>, a contentious aspect of AI, has various applications such as for airport security, unlocking smartphones, and in federal law enforcement databases. While it aids in finding missing individuals and identifying criminals, critics argue that it can lead to mistaken identity and privacy invasion. Several cities and states have banned or regulated its use. The challenge for lawmakers is to strike a balance between protecting individual privacy and promoting AI development without getting entangled in broader social and political issues.</p><p>When assessing the impact of AI privacy, it’s essential to <a href="https://www.brookings.edu/articles/protecting-privacy-in-an-ai-driven-world/">differentiate between data-related concerns common to all AI</a>, like false positives and overfitting, and those particular to personal information use.</p><h4><strong>Autonomy</strong></h4><p>AI’s capability to either hinder or support human autonomy and is a significant issue. There are <a href="https://www.frontiersin.org/articles/10.3389/frai.2021.705164/full">different forms of interpersonal disrespect by AI systems</a> varying in extent, including direct interference, coercion, manipulation, deception, nudging, paternalism, delegation (by human) of cognitive tasks to AI, and biased AI recommendation systems. The degree of impact depends on factors like transparency, consent, alignment with users’ values, and the nature of AI recommendations and interactions.</p><p>There are <a href="https://iep.utm.edu/ethics-of-artificial-intelligence/">ethical challenges posed by autonomous systems</a>, focusing on autonomous vehicles and weapons systems. Autonomous weapons systems have the capability to replace human soldiers, potentially reducing war crimes if equipped with ethical governance, but concerns include increased conflicts, AI’s interpretation of ethical principles, and hacking risks. Though self-driving cars offer benefits like improved traffic safety, concerns arise about accidents and the need “ethics settings” to determine their behaviour, posing real-life machine ethics challenges. Fatal accidents involving autonomous vehicles have already occurred.</p><p>In addition, philosophers explore the ethical decision-making principles, focusing on “<a href="https://iep.utm.edu/ethics-of-artificial-intelligence/">meaningful human control</a>” in situations involving harm to humans. Furthermore, there are concerns about responsibility gaps in assigning blame for outcomes generated by autonomous systems, especially morally questionable ones, presenting complex challenges in allocating responsibility.</p><h4><strong>Sustainability</strong></h4><p>There <a href="https://link.springer.com/article/10.1007/s43681-021-00043-6">significant financial and environmental costs associated with training deep learning models</a>, particularly in natural language processing (NLP), which involve extensive energy consumption and carbon emissions. The argument that as AI models are developed and used, it’s crucial to evaluate their impact on climate change. Although some energy for AI model training can come from renewable sources or carbon offset resources, the high energy demands remain a concern, especially in regions without access to carbon-neutral energy. It highlights the ethical dilemma of allocating energy and resources to AI model training rather than providing electricity to millions of people lacking access to modern amenities.</p><p>Additionally, the fine-tuning AI models can have a more expensive environmental impact than initial training. This raises ethical questions about the proportionality of using AI methods for certain tasks, particularly those with ethically charged implications. Independent studies suggest that <a href="https://www.emerald.com/insight/content/doi/10.1108/JICES-11-2021-0106/full/html">AI’s contribution to climate change can be substantial</a>, with some estimates indicating that information and communication technology emissions, including AI, may exceed 14% of global emissions by 2040.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/297/0*8W_kCm_0LtBQOek6" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><h3>Optimal practices for addressing AI ethical issues.</h3><p>AI can unintentionally perpetuate human biases, leading to unfair decisions or recommendations. Technical tools exist that can detect and mitigate AI bias across various data types— Explainability tools play a crucial role in identifying bias in AI models.</p><p>In complex AI scenarios, full explainability may not always be realistic or necessary. <a href="https://www2.deloitte.com/us/en/insights/focus/cognitive-technologies/design-principles-ethical-artificial-intelligence.html">Trustworthiness</a> becomes a key concept. Users might not fully grasp the inner workings of AI systems, but trust can be built through factors like training, safety regulations, and the reputation of the system’s manufacturer. For instance, how drivers trust semiautonomous vehicles without fully understanding their algorithms. It’s vital to establish this sense of trust and map out under what conditions AI systems can be relied upon to ensure safe and ethical use.</p><p>Improved explainability can boost trust, especially in medicine where professionals need to understand and trust AI-driven decisions. Some AI models, like Decision Trees, are <a href="https://www.peertechzpublications.com/articles/ARA-5-111.php">inherently interpretable</a>.</p><p>AI developers need to <a href="https://www.weforum.org/agenda/2021/01/how-to-address-artificial-intelligence-fairness/">be educated about their biases</a> and how they can inadvertently introduce it into AI systems development. Creating a culture of fairness is essential, involving all decision-makers who understand AI bias issues and their impacts. AI systems should reduce the digital gap, improve accessibility, and address issues related to gender, disabilities, geography, and ethnicity. AI fairness should have a global dimension.</p><p>Additionally, <a href="https://link.springer.com/article/10.1007/s13347-022-00507-5">promoting AI autonomy and positive independence</a> involves measures such as transparency and informed consent. Freedom and autonomy values are crucial in guiding AI development and use, aiming to make AI technology that inspires and liberates, free from external interference and systemic power dynamics.</p><p>AI creators should establish internal principles and governance frameworks to ensure their AI systems are fair, accurate, transparent, and robust. External institutions, such as governments and industry bodies, can also contribute by co-creating AI frameworks.</p><p>Achieving AI fairness necessitates governance structures that detect and mitigate bias in data collection and processing. Data protection law requires that <a href="https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/how-do-we-ensure-fairness-in-ai/what-about-fairness-bias-and-discrimination/">personal data be processed fairly</a> without unjustified adverse effects on individuals. Frameworks should define oversight specific to each use case, and fairness should be context-dependent and defined through multi-stakeholder consultations. Further insights on <a href="https://medium.com/version-1/ai-law-regulation-a-synopsis-of-the-global-perspective-8ad720c6bca2">AI Law &amp; Regulation</a> can be found in the linked blog.</p><p>Furthermore, AI sustainability metrics, i.e., measuring carbon emissions during model training and acknowledging regional variations is vital. The energy efficiency of AI algorithms, revealing drastic differences in energy consumption shows the importance of incorporating energy efficiency metrics alongside traditional performance measures. The ultimate objective is to encourage the development of energy-efficient and ethically responsible AI solutions, with <a href="https://dl.acm.org/doi/abs/10.1145/3604930.3605715">standardized reporting of energy and carbon data</a> to foster a culture of social responsibility in AI research and development. Further insights on <a href="https://medium.com/version-1/the-green-algorithm-measuring-sustainability-in-ai-4c775e811c2a">Green Algorithm</a> can be found in the linked blog.</p><p>The drive to tackle AI ethical issues extends beyond preserving human values; it is also a socio-economic obligation. Companies that fail to establish ethical standards in their AI products may encounter obstacles in gaining broader technology acceptance. Consequently, every stakeholder within the AI ecosystem should allocate resources and <a href="https://www2.deloitte.com/us/en/pages/consulting/articles/ai-ethicist-and-ai-bias.html">work diligently to actively promote comprehensive AI ethics</a>.</p><h3>About the author:</h3><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=64374b30d4b4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/the-moral-compass-in-the-machine-ethics-in-ai-64374b30d4b4">The Moral Compass in the Machine: Ethics in AI</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Green Algorithm: Measuring Sustainability in AI]]></title>
            <link>https://medium.com/version-1/the-green-algorithm-measuring-sustainability-in-ai-4c775e811c2a?source=rss-a9543806140d------2</link>
            <guid isPermaLink="false">https://medium.com/p/4c775e811c2a</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[green-ai]]></category>
            <category><![CDATA[responsible-ai]]></category>
            <dc:creator><![CDATA[Rosemary J Thomas, PhD]]></dc:creator>
            <pubDate>Mon, 09 Oct 2023 15:21:44 GMT</pubDate>
            <atom:updated>2024-12-17T09:54:33.355Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qu2So_UQVtUKPYHosv3X2w.jpeg" /><figcaption><a href="https://www.bing.com/images/create">Created using Microsoft Bing Image Creator</a></figcaption></figure><p>In the realm of AI, assessing carbon emissions using CO2-equivalents (CO2eq) during model training while considering regional variations is pivotal for sustainability. The significant variability in energy efficiency among AI algorithms underscores the need to incorporate energy metrics alongside conventional performance measures. Notably, large language models like ChatGPT and Llama 2 exhibit substantial carbon footprints and energy consumption, prompting a critical examination of sustainability within their context. <a href="https://www.cell.com/one-earth/pdf/S2590-3322%2821%2900299-2.pdf">Policy recommendations</a> underscore the importance of selecting eco-conscious cloud providers and strategic data centre locations. The overarching goal is to encourage the development of energy-efficient and environmentally responsible AI solutions, with a strong emphasis on standardized reporting of energy and carbon data to promote social responsibility within AI research and development.</p><blockquote><a href="https://www.version1.com/application-modernisation-integration/ai-powered-code-analysis-and-documentation-decipher/">AI-powered code analysis and documentation — Decipher | Version 1</a></blockquote><h3><strong>Sustainability metrics</strong></h3><p>Let’s investigate the some of the metrics: CO2-equivalents, electricity consumption, Carburacy and Universal Sustainability Metrics. You’ll find a summary table at the end for reference.</p><h4><strong>CO2-equivalents</strong></h4><p>The environmental impact of <a href="https://arxiv.org/abs/1910.09700">Machine Learning (ML) model training can be measured and addressed through key metrics</a>. CO2eq provide a standardized measure to quantify carbon emissions associated with ML model training, simplifying the assessment of greenhouse gas (GHG) impact. It’s important to note significant regional differences in carbon emissions, with North American servers emitting anywhere from 0 tCO2eq/kWh (20g) in Quebec, Canada, to 0.0008 tCO2eq/kWh in Iowa, USA. These variations are due to differences in energy sources and grids. While <a href="https://www.cell.com/one-earth/pdf/S2590-3322%2821%2900299-2.pdf">AI currently contributes a relatively small 1.4% to global GHG emissions within Information and Communication Technologies</a>, concerns arise about potential rapid increases if current AI research trends continue.</p><p>There is a focus on sustainability metrics for <a href="https://www.jair.org/index.php/jair/article/view/14340">evaluating the environmental impact of Automated Machine Learning</a> (AutoML) research. Key metrics include runtime, which relates to energy consumption and can be used to estimate environmental impact, especially when combined with data on energy consumption and the energy mix. CPU/GPU (Central/Graphical Processing Units) hours are suggested as a quantifiable measure that can be converted into CO2eq emissions with knowledge of the energy mix. Energy consumption, while dependent on hardware, is considered a direct measure of the environmental footprint. It is important to consider CO2eq as a direct measure of environmental impact, even if it’s challenging to measure directly, as it can be computed based on energy consumption and energy mix data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DKqoAORaftHYsNnd.png" /><figcaption>Figure 2. <a href="https://www.bbc.co.uk/news/science-environment-52719662">Carbon emission reduction options</a></figcaption></figure><p>In addition, sustainability metrics are crucial when assessing the <a href="https://arxiv.org/abs/2305.03123">environmental impact of large language models (LLMs) like ChatGPT</a> and Llama 2. This involves considering the carbon footprint associated with LLMs at various stages, including training, inference, and real-time requests, considering factors like energy source, training location, and GPU energy consumption. For example, <a href="https://ieeexplore.ieee.org/abstract/document/10251541">training GPT-3 with 175 billion parameters</a> resulted in 608 tCO2eq emissions. <a href="https://huggingface.co/meta-llama/Llama-2-13b">Llama 2’s developers disclosed its carbon footprint</a>, which was estimated at 539 tCO2eq due to 3.3 million GPU hours of computation. Importantly, Meta’s sustainability program has offset 100% of Llama 2’s emissions, making the model carbon neutral.</p><p>The intersection of <a href="https://www.cell.com/one-earth/pdf/S2590-3322%2821%2900299-2.pdf">AI and the climate emergency</a> presents valuable insights. It delves into the carbon footprint of AI training processes, highlighting the exponential increase in GHG emissions tied to ML model development due to rising computing demands. A striking estimate for a single training run of the GPT-3 model, amounting to 247 tCO2eq, serves as a sobering comparison to emissions from typical passenger cars in the United States.</p><h4><strong>Electricity consumption</strong></h4><p>The <a href="https://arxiv.org/abs/1910.09700">FLOPS/W (Floating Point Operations per Second per Watt) metric</a> is crucial for evaluating the energy efficiency of hardware devices used in ML. It shows substantial differences, with CPUs being significantly up to ten times less efficient, compared to GPUs, while TPUs (Tensor Processing Units) are impressively 4 to 8 times more efficient than GPUs. These metrics provide a solid basis for assessing and reducing carbon emissions in the field of ML, promoting environmentally responsible practices.</p><p>An independent study reveals <a href="https://ieeexplore.ieee.org/abstract/document/9830097">significant differences in energy efficiency among AI algorithms</a>. K-Nearest Neighbour (KNN) is the most energy-efficient, using only 0.01 Joules on average, while Random Forest is the least energy-efficient, with an average consumption of about 1.98 Joules per run. The energy consumption disparities range widely, from a 20% reduction (Bagging Classifier vs. Support-Vector Machine) to a substantial 99.49% reduction (Random Forest vs. KNN). These findings stress the importance of considering energy efficiency as a critical factor when choosing ML models, in addition to traditional performance metrics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/434/1*B5oeOj8Z9xxaI3dG4rL4oQ.png" /><figcaption>Figure 2. <a href="https://ieeexplore.ieee.org/abstract/document/9830097">Energy efficiency among AI algorithms</a></figcaption></figure><p>Another study highlights the importance of evaluating AI technologies with a dual focus on accuracy and energy consumption. While accuracy metrics are well-established, measuring energy usage is more challenging. The study introduces <a href="https://ieeexplore.ieee.org/abstract/document/10012292">various metrics to assess the environmental impact of AI</a>, including model size, number of parameters, elapsed time, FLOPs, and FLOPs per watt (efficiency ratio). It stresses the need to understand the trade-offs between accuracy and energy, especially in AI training, where larger datasets and complex models can yield diminishing returns in accuracy improvement.</p><p><a href="https://huggingface.co/meta-llama/Llama-2-13b">Llama 2</a> discloses its power consumption during training, with GPU usage ranging from 350 to 400W. While this power usage is significant due to the model’s size, efforts have been made to enhance efficiency. Meta has also initiated a sustainability program to offset the carbon emissions from the model’s training, showcasing their commitment to reducing the environmental impact of AI. They have further shared their research and a Responsible Use Guide, advocating for ethical and responsible model usage. It’s crucial to recognize that AI models like Llama 2 can have a substantial environmental impact due to their computational demands for training. Nevertheless, these measures to mitigate emissions, provide transparency on energy usage, and promote responsible use are steps toward more sustainable AI development.</p><h4><strong>Carburacy</strong></h4><p><a href="https://ojs.aaai.org/index.php/AAAI/article/view/26686">Carburacy</a> is introduced as a novel carbon-aware accuracy measure for evaluating the eco-sustainability of Natural Language Processing (NLP) models. It combines two essential factors: ‘effectiveness’ (as ‘R’), which measures model accuracy using ROUGE (package for automatic summaries evaluation) scores, and ‘cost’ (as ‘C’), encompassing resource requirements like energy consumption, hardware costs, and carbon emissions. Efficiency amalgamates effectiveness and cost. Carburacy aims to strike a balance between achieving high accuracy and minimising environmental impact.</p><p>It explores the impact of input size on sustainability, revealing larger inputs can improve effectiveness and increase energy consumption, affecting Carburacy scores. Training batch size is another consideration, with larger batches reducing backpropagation calls but potentially hindering Carburacy by decreasing effectiveness. Decoding strategy choice, such as beam search, presents a trade-off between effectiveness and energy consumption. Few-shot learning is highlighted as an eco-sustainable approach, achieving good effectiveness with minimal environmental impact. Model type also matters, with linear models proving more eco-sustainable in long document summarization tasks compared to quadratic models, achieving higher Carburacy scores with larger input lengths.</p><h4><strong>Universal Sustainability Metrics</strong></h4><p>In the context of edge computing and AI, the emphasis is on developing “Green AI” solutions by <a href="https://www.sciencedirect.com/science/article/pii/S2210537921000718">introducing crucial sustainability metrics</a>. These metrics include Recognition Efficiency (RE), which balances accuracy, complexity, and energy consumption during inference, and Training Efficiency (TE), which prioritises energy efficiency during AI model training. Deep Learning Lifecycle Efficiency (DLLCE) is introduced to assess model efficiency throughout their lifecycle, considering factors beyond inference energy. Life Cycle Recognition Efficiency considers efficiency across the complete model lifecycle, including re-training and deployment phases. A sustainability classification system (Class A, B, C, D) based on RE values categorises models, libraries, and platforms into different sustainability classes with specific threshold ranges.</p><p>There are valuable insights into <a href="https://arxiv.org/abs/1910.09700">reducing the environmental impact</a> of ML model training and addressing carbon emissions. It highlights significant energy consumption, particularly when using fossil fuels, and recommends quantifying emissions with CO2eq while considering server location. Efficiency improvements in infrastructure and training practices, such as optimising GPUs and hyperparameter search, are suggested. Advocacy for sustainable cloud providers and data centre locations is stressed. The importance of minimising resource waste and selecting energy-efficient hardware is stressed, while acknowledging practical constraints.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VnyKKIx5-0cC7Mc181UBow.png" /><figcaption>Table 1. Summary of sustainability metrics</figcaption></figure><h3><strong>Conclusion</strong></h3><p>This overview shows the need for <a href="https://dl.acm.org/doi/abs/10.1145/3604930.3605715">standardization in measuring and reporting power, energy, and carbon consumption data in AI</a>, emphasizing the <a href="https://www.cell.com/one-earth/pdf/S2590-3322%2821%2900299-2.pdf">lack of comprehensive sustainability metrics in academic and industry domains</a>. It advocates for a holistic approach that integrates human, social, and environmental values in AI solutions. It introduces sustainability metrics to encourage the development of energy-efficient AI systems, offering frameworks to evaluate and compare efficiency and sustainability across platforms and use cases. It particularly emphasises the importance of considering sustainability metrics in models like ChatGPT and Llama 2, using estimates and comparisons to illustrate the significance of this issue. The call for systematic reporting of carbon emissions and model details by researchers and organizations is made to promote sustainability and social responsibility in AI, encouraging further community engagement on this vital topic.</p><h3>About the author:</h3><p>Rosemary J Thomas, PhD, is a Senior Technical Researcher at the <a href="https://www.version1.com/it-service/innovation-labs/">Version 1 AI Labs.</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4c775e811c2a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/version-1/the-green-algorithm-measuring-sustainability-in-ai-4c775e811c2a">The Green Algorithm: Measuring Sustainability in AI</a> was originally published in <a href="https://medium.com/version-1">Version 1</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>