<?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 Zsuzsanna Hornyák on Medium]]></title>
        <description><![CDATA[Stories by Zsuzsanna Hornyák on Medium]]></description>
        <link>https://medium.com/@zsuzsi.hornyak?source=rss-63190133ce57------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*exXTGAbaIMaPRXHMl-NMkQ.jpeg</url>
            <title>Stories by Zsuzsanna Hornyák on Medium</title>
            <link>https://medium.com/@zsuzsi.hornyak?source=rss-63190133ce57------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 22:26:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zsuzsi.hornyak/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[Own goal]]></title>
            <link>https://medium.com/@zsuzsi.hornyak/own-goal-d3366cd389cb?source=rss-63190133ce57------2</link>
            <guid isPermaLink="false">https://medium.com/p/d3366cd389cb</guid>
            <dc:creator><![CDATA[Zsuzsanna Hornyák]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 12:06:29 GMT</pubDate>
            <atom:updated>2026-11-18T12:06:29.262Z</atom:updated>
            <content:encoded><![CDATA[<p>(word vomit microblogs here I come)</p><p>Beautiful, responsible AI strategy / guidelines that start with: the human is responsible of what the AI coding agents produce.</p><p>But also, the human is not expected to write code at all. Or even read it for most of the time. AI coding agents do all that, reviewing it by automatic quality assurance tooling.</p><p>But, small print! Manual review and coding is still expected when it feels needed (who decides this? Unclear).</p><p>To put it another way:</p><p>Here I have this football team, they have been doing fine. But from now on, it will be robots playing on the field! Think how many more games we can play this way, they never get tired!</p><p>We will keep some of the previous players and coaches to give instructions to our robot team, and to adjust their movements with remote controls when needed.</p><p>Sometimes, when we have a very important match and a robot breaks down, we need to get the human soccer player back on the field. Surely they can pick up where they left of, even though they haven’t had daily training in months?</p><p>Right?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d3366cd389cb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Make Humans Smart]]></title>
            <link>https://medium.com/@zsuzsi.hornyak/make-humans-smart-64e6672417f3?source=rss-63190133ce57------2</link>
            <guid isPermaLink="false">https://medium.com/p/64e6672417f3</guid>
            <dc:creator><![CDATA[Zsuzsanna Hornyák]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 08:58:29 GMT</pubDate>
            <atom:updated>2026-11-17T08:58:29.684Z</atom:updated>
            <content:encoded><![CDATA[<p>Two years ago I held a presentation at the Make Data Smart conference(s) titled “Hype vs. ethics — a critical look at generative AI”.</p><p>I could literally hold the same presentation today. The question of “should we all rush to this technology considering all the backsides and shady practices?” is still relevant and burning every day in my mind. The only difference I would make is to showcase even more damning evidence against this technology and it’s providers.</p><p>In 2024 I finished my presentation with the stock photo of a scale and talked about how everything in life is complex with advantages and disadvantages, we can’t make perfect decisions, etc. That was mostly to appear PC and not like a raging technology-pessimist.</p><p>I guess my ending today would be different, as I have grown older, wiser and angrier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Va-DQTx6JNaovIf7lEGTMQ.png" /></figure><p>I have been fascinated with the field of Artificial Intelligence (in it’s 90s definition from Stuart Russell and co., not today’s narrowed mainstream understanding) for over 16 or so years now. What attracted me to it was the focus on learning, both for human and machine. Logic. Scavenge hunting in data.</p><p>So much of generative AI use is focused on a certain perceived productivity. At the danger of sounding like a snob, both the design, development and use of these tools lacks sophistication for me. My interest in the field is dwindling.</p><p>Make Data Smart is still a conference by humans, for humans (even if probably heavily AI-assisted), so I am not taking that from them. But I am kind of sick of making data and the machine smart and making us humans more stupid and exhausted.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=64e6672417f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Securely Building Azure ML Environments with Private Packages from Azure Artifacts]]></title>
            <link>https://medium.com/@zsuzsi.hornyak/securely-building-azure-ml-environments-with-private-packages-from-azure-artifacts-db30a11ca3c1?source=rss-63190133ce57------2</link>
            <guid isPermaLink="false">https://medium.com/p/db30a11ca3c1</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[azure-ml-studio]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[azure-devops]]></category>
            <dc:creator><![CDATA[Zsuzsanna Hornyák]]></dc:creator>
            <pubDate>Tue, 29 Apr 2025 12:29:06 GMT</pubDate>
            <atom:updated>2025-04-29T12:30:56.669Z</atom:updated>
            <content:encoded><![CDATA[<p>With the increasing adoption of Machine Learning across projects, maintaining code quality and reusability becomes paramount. Packaging common logic into Python packages significantly improves maintainability, facilitates collaboration, and ensures consistency across different ML experiments and deployments. This approach is key to building robust and scalable ML systems.</p><p>In our scenario, we are training multiple models within Azure Machine Learning studio. These models share common data processing and utility functions, which we’ve consolidated into a Python package published on an Azure Artifact feed within the same Azure DevOps project. A critical requirement is to build reusable Azure ML environments that include this private package, ensuring our model training and deployment processes are standardized and reproducible.</p><h3>Initial version</h3><p>Azure Machine Learning environments can be defined using <a href="https://learn.microsoft.com/en-us/azure/machine-learning/how-to-manage-environments-v2?view=azureml-api-2&amp;tabs=cli#create-an-environment-from-a-conda-specification">Conda files</a>, which specify Python dependencies. This allows for reproducible environments by listing exact package versions. Note that you have to explicitly define the address for your Azure Artifact Feed Python repository.</p><pre>name: training-environment<br><br>channels: <br><br>- anaconda <br><br>- defaults <br><br>dependencies: <br><br>- python&gt;=3.10 <br><br>- pip&gt;=22.2.2 <br><br>- pip: <br><br>- --extra-index-url https://pkgs.dev.azure.com/[..]<br><br>- azure-ai-ml==1.13.0 <br><br>- azure-identity==1.15.0 <br><br>- our-cool-reusable-package==1.0</pre><p>Initially, we explored using Azure ML Connections (a preview feature at the time of writing) to authorize access to the Azure Artifact feed directly from within Azure ML. While this feature simplified connecting to external resources, it primarily supported authentication via Personal Access Tokens (PATs), which are less secure and harder to manage at scale compared to identity-based authentication methods.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CYvy66j0wIys3vjQmQsMNw.png" /></figure><h3>Proper feed authentication</h3><p>To move beyond the limitations of PATs and adopt a more secure and sustainable authentication method, we aimed for an identity-based solution, similar to how Managed Identities are used elsewhere in Azure for service-to-service authentication.</p><p>Our first attempt involved assigning a Managed Identity to the Azure ML Compute Instance used for building environments. This identity was granted read access to the Azure DevOps feed. However, we discovered that while the identity worked correctly for authenticating package access <em>at runtime</em> within an environment, it did <em>not</em> automatically authenticate access during the <em>build phase</em> when Docker images were being constructed, even though the build process ran on the same compute. This approach proved unreliable and unsustainable for consistently building environments.</p><p>We also reached out to Microsoft Support to hear about when/if they will support Managed Identity for authentication in Connections but have not heard back from them,.</p><h3>Final solution with DevOps Pipelines</h3><p>The final, robust solution leverages an Azure DevOps Pipeline. Azure DevOps pipelines have built-in, secure mechanisms for authenticating to Azure Artifact feeds within the same project, eliminating the need for PATs or complex identity configurations during the pipeline execution. The pipeline’s role is to first securely download the required package version from the feed and then package it along with the environment definition files before sending everything to Azure ML for the environment build process.</p><pre># Example pipeline, could be parameterized to support several environments with similar definition<br>stages:<br>  - stage: DownloadPackages<br>    displayName: &quot;Download packages from requirement&quot;<br>    jobs:<br>      - job: DownloadPackages<br>        steps:<br>          - task: UsePythonVersion@0<br>            inputs:<br>              versionSpec: &quot;3.10&quot;       # In sync with AML environment<br>              architecture: &quot;x64&quot;<br>              addToPath: true<br><br>          - task: PipAuthenticate@1<br>            displayName: &#39;Pip Authenticate&#39;<br>            inputs:<br>              artifactFeeds: &#39;My DevOps Project/My Artifact Feed&#39;<br>              onlyAddExtraIndex: true   # Force pip to fallback to external index<br><br>          - script: |<br>              cd training_environment/docker-context<br>              mkdir pip_packages<br>              pip download --platform manylinux2014_x86_64 --only-binary=:all: --no-binary=:none: -r requirements.txt -d pip_packages<br>              ls -lah pip_packages<br>            displayName: &quot;Download environment dependencies&quot;<br><br>          - script: |<br>              python -m pip install -r requirements.txt<br>            displayName: &quot;Install dependencies for publishing job to AML&quot;<br><br>          - task: AzureCLI@2<br>            displayName: &#39;Create AML environment with Python SDK&#39;<br>            continueOnError: false<br>            inputs:<br>              azureSubscription: &quot;AML Service connection&quot;<br>              scriptType: bash<br>              workingDirectory: $(System.DefaultWorkingDirectory)/code<br>              scriptLocation: inlineScript<br>              inlineScript: |<br>                my_env_building_script.py<br> </pre><p><strong>Note:</strong> When downloading the package within the pipeline, ensure that the correct platform version is selected that is compatible with the target platform and Python version specified for the Azure ML environment.</p><p>For defining the Azure ML environment itself, especially when including local files like a downloaded private package, using a Docker context is the most suitable approach. The <a href="https://learn.microsoft.com/en-us/azure/machine-learning/how-to-custom-docker-image-environment?view=azureml-api-2&amp;tabs=cli#create-an-environment-from-a-docker-context%5D">Docker context</a> includes the Dockerfile and all necessary files (like requirements.txt and the downloaded package) that are sent to Azure ML for building the custom Docker image that forms the environment.</p><p>Folder structure for our environment definition:</p><pre>training_environment/ <br>│   ├── docker-context/ <br>│   │   ├── Dockerfile <br>│   │   ├── requirements.txt <br>│   │   └── pip_packages/  # Downloaded by DevOps pipeline <br>│   └── training-environment.yml # AML environment definition pointing to Docker-context </pre><p>Dockerfile:</p><pre>FROM mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu22.04 <br><br>COPY pip_packages /pip_packages <br><br>COPY requirements.txt requirements.txt <br><br>RUN pip install --no-index --find-links /pip_packages -r requirements.txt </pre><p>AML environment definition (training-environment.yml):</p><pre>$schema: https://azuremlschemas.azureedge.net/latest/environment.schema.json <br><br>name: training_environment <br><br>description: Dependencies for training<br><br>build: <br><br>path: docker-context </pre><p>Simplified python script for environment build (my_env_building_script.py):</p><pre>from azure.ai.ml import MLClient, load_environment <br>from azure.identity import DefaultAzureCredential <br><br>client = MLClient() # initializing connection to Azure ML workspace<br>env = load_environment(f&quot;training-environment/training-environment.yml&quot;)<br>client.environments.create_or_update(environment=env)</pre><p>Standardizing the environment definitions and parameterizing the building script and pipeline can help you create a robust MLOps pipeline for continuous build and deploy of your environments with private packages.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db30a11ca3c1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Thoughts of cognitive dissonance in the summer of 2024]]></title>
            <link>https://medium.com/@zsuzsi.hornyak/thoughts-of-cognitive-dissonance-in-the-summer-of-2024-bb40c831f133?source=rss-63190133ce57------2</link>
            <guid isPermaLink="false">https://medium.com/p/bb40c831f133</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[climate-change]]></category>
            <category><![CDATA[cognitive-dissonance]]></category>
            <dc:creator><![CDATA[Zsuzsanna Hornyák]]></dc:creator>
            <pubDate>Fri, 28 Jun 2024 18:35:15 GMT</pubDate>
            <atom:updated>2024-06-28T18:36:57.476Z</atom:updated>
            <content:encoded><![CDATA[<p>This post is an attempt at cleaning my head of some conflicting thoughts, so that I can try to take my summer vacation without their constant rumbling in the back of my head. They are not thoughts I want to cast from myself, but this my plea for a pause so I can use my holidays for recharging.</p><p>I like summer and look forward to time off with my family, but the ever-increasing heat has become a constant reminder for me of the climate changes we experience. I find myself in a kind of cognitive dissonance —I do very much enjoy being out in the sun and the warm, feeling light after the long winter in Norway, and yet there is a trepidation with each warmth record and long streak. Can I really be joyous over the increase in warm days this year, when I know it comes with the price of more extreme climate events worldwide?</p><p>Such reflections weigh on me when it comes to my work in the field of data science and artificial intelligence as well. The excitement, the hype, the possibilities with generative AI — it is fun to try out this new, promising technology. But it is not free, and the costs are complex and loom beyond the price tag of the actual services. Do I, by a way of using these products, increase the demand for LLMs, and thus incentivize their further, incredibly energy greedy development? How do we measure the value of early adaption against concerns of security, environmental and societal impact?</p><p>At one point I made a choice to stop buying new clothes from shops which produce them in neither sustainable nor ethical ways, to stand up against fast fashion. I do wonder if it is time to make a stand against “fast technology” in a way.</p><p>I don’t really have any answers, but I do have a lot of questions. This post shall be my attempt of putting them on the shelf for later inspection. Hopefully in a few weeks I will come back refreshed to learn and write more about how we can approach the newest AI wave in responsible, sustainable ways.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bb40c831f133" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>