Hi  All, 

Today I try to onboarding the Mac to Microsoft Defender, and following https://learn.microsoft.com/en-us/defender-business/mdb-onboard-devices?tabs=mac#intune-for-mac 

but find the CLI is

/usr/bin/unzip WindowsDefenderATPOnboardingPackage.zip
&& /bin/chmod +x MicrosoftDefenderATPOnboardingMacOs.sh
&& /bin/bash -c MicrosoftDefenderATPOnboardingMacOs.sh

and the onboarding package zip file name is GatewayWindowsDefenderATPOnboardingPackage.zip

in Line 1 Doesn’t seem to be the same as the file name.

in Line 3 This command is not recognized

I tried opening it directly GatewayWindowsDefenderATPOnboardingPackage.zip and get MicrosoftDefenderATPOnboardingMacOs.sh Then the device has onboarded 

may i know Is this a documentation wrong / not update or ?

​Hi  All,  Today I try to onboarding the Mac to Microsoft Defender, and following https://learn.microsoft.com/en-us/defender-business/mdb-onboard-devices?tabs=mac#intune-for-mac but find the CLI is/usr/bin/unzip WindowsDefenderATPOnboardingPackage.zip
&& /bin/chmod +x MicrosoftDefenderATPOnboardingMacOs.sh
&& /bin/bash -c MicrosoftDefenderATPOnboardingMacOs.sh and the onboarding package zip file name is GatewayWindowsDefenderATPOnboardingPackage.zip in Line 1 Doesn’t seem to be the same as the file name.in Line 3 This command is not recognizedI tried opening it directly GatewayWindowsDefenderATPOnboardingPackage.zip and get MicrosoftDefenderATPOnboardingMacOs.sh Then the device has onboarded may i know Is this a documentation wrong / not update or ?          Read More

​

Dear Sir

I have an issue about Teams Tab application, which is Tab web hosting from AWS server.

On Desktop Teams app, when I click “Reload app” button in “…” more button at the top-right side, my web hosting is working normally.

However, If I click “Refresh” in Teams mobile, I works differently.

AWS static hosting server has strict rule. 

If the URL is different, AWS shows “access denied” like below.

Can you advise why its’ differently worked ? 

I tried to test this many times with cache clean. 

Definitely it shows different result.

Many thanks

BH

​Dear SirI have an issue about Teams Tab application, which is Tab web hosting from AWS server.On Desktop Teams app, when I click “Reload app” button in “…” more button at the top-right side, my web hosting is working normally.However, If I click “Refresh” in Teams mobile, I works differently.AWS static hosting server has strict rule. If the URL is different, AWS shows “access denied” like below.Can you advise why its’ differently worked ? I tried to test this many times with cache clean. Definitely it shows different result. Many thanksBH  Read More

​

The excel format 2003 xml Spreadsheet is no longer supported in my company, can any one suggest what could be an alternative file format I can choose to save the file to get the same file type options as per new M365 standard

​The excel format 2003 xml Spreadsheet is no longer supported in my company, can any one suggest what could be an alternative file format I can choose to save the file to get the same file type options as per new M365 standard  Read More

​

October 23, 2024 is our special day for Small Basic.  The newest version of my program is SFF730-12.  Have fun and Happy 16th Birthday, Small Basic!!

​Microsoft Tech Community – Latest Blogs –Read More 

Building Generative AI applications can feel daunting for traditional app developers. What does the end-to-end application development cycle look like? What models should I use, and where do I find them? What tools should I be using for build, test, and deploy, my AI application? This blog post gives you a sneak peek at a week-long series of posts that were just published, that give you a hands-on journey through the process. Let’s learn more!

Kicking Off Azure AI Week!

This week we published a 5-part blog on the Build Intelligent Apps initiative’s #30DaysOfIA series. Our focus was application developers who wanted to build a custom copilot code-first on Azure AI, allowing them to have more control over various decisions made in the end-to-end workflow for generative AI applications. We did this by walking through two core samples (Contoso Chat and Contoso Creative Writer) from prompt to production. Along the way, we shared insights into key tasks and the developer tools to simplify them.

In this blog post, we’ll briefly introduce the two applications and give you an overview of what the series covers, with links to each post for deeper dives. Ready? Let’s Go!

1. What are we building?

Our first application is Contoso Chat, a customer service chatbot that answers user questions about a retailer’s products,  using the Retrieval Augmented Generation pattern (RAG) to ground responses in both the product catalog and customer purchase history.
 

Our second application is Contoso Creative Writer, a content publishing assistant that uses the Multi-Agent Conversation pattern to coordinate and execute multiple tasks autonomously, on behalf of the user.

2. How are we building it?

The figure below shows the AI Application Architecture for the Contoso Chat retail copilot. User requests are received through an endpoint hosted in Azure Container Apps, then processed using a RAG-based workflow that uses Azure AI Search (product index) and Azure Cosmos DB (customer database) with Azure OpenAI Services (model deployment) to process user requests and return the response back to the UI.

The next figure shows the AI Application Architecture for the Contoso Creative Writer multi-agent copilot which follows a similar user interaction flow – except that processing now requires coordination across multiple agentic AI tasks before final output is generated.

3. What does the developer workflow look like?

We’re glad you asked! If you’ve explored generative AI application development before, you’re probably familiar with this GenAIOps application lifecycle which breaks down the developer workflow into 3 stages: ideation (build and validate a prototype), augmentation (iterate & evaluate with larger input datasets), operationalization (deploy to production).

In this blog series, we map this lifecycle to a very clear developer workflow as shown below, giving you an intuitive sense for the task to perform, and the tool to use to accomplish it, at each stage.

Get started reading the posts, in this order:

Kicking Off Azure AI Week – Learn about the app scenarios, architecture & lifecycle.
Provision with AZD – Provision Azure infrastructure & setup your dev environment.
Ideate with Prompty – Build an app prototype using Prompty assets and tooling.
Evaluate with AI – Build custom evaluators and use AI-assisted evaluation flows.
Deploy with ACA – Create a FastAPI server & deploy with Azure Container Apps.

Here’s a visual summary of what you’ll learn:

If you found this series valuable, please star the repos to help others discover them!

Contoso Chat – custom retail copilot with Retrieval Augmented Generation
Contoso Creative Writer – custom content copilot with Multi-Agent Collaboration

5. Next Steps

Want to get hands-on experience building these copilots? Take these actions today!

Register for Microsoft AI Tour – join an instructor-led workshop session.
Register for Microsoft Ignite – look for related lab & breakout sessions on Azure AI.
Browse the AI Templates Collection – explore samples for more scenarios.

Have a scenario you want to build a custom copilot for? Have questions about Prompty, Azure AI Studio, or the GenAI Ops workflow? Want to provide feedback on the samples? Leave us a comment here and let us know! Happy learning!

​Microsoft Tech Community – Latest Blogs –Read More 

Is there any possible way to achieve Auditing and encryption in Dataverse for teams?

​Is there any possible way to achieve Auditing and encryption in Dataverse for teams?   Read More

​

Hello,

I am trying to put together a spreadsheet that allows me to compare members of our team. We work in different “Sub” teams as you will. Just because you work on one team, does not mean your will not also be a part of another team. We encourage people to work on three different sub teams to allow them to work with others and allow for more free thinking ideas and strategies. I would like to be able to compare our team and see how well others work on different teams and try to objectively identify who really excels working with certain people and who does not. 

So my problem is when I try and use AVERAGEIFS or STDEV.S it only pulls the first available value. As the team and year changes I would Expect the average and the standard deviation to change as well. I have attached a spread sheet with my problem. Any clues with how to fix it would be greatly appreciated. 

Thank You,

Tripp

​Hello, I am trying to put together a spreadsheet that allows me to compare members of our team. We work in different “Sub” teams as you will. Just because you work on one team, does not mean your will not also be a part of another team. We encourage people to work on three different sub teams to allow them to work with others and allow for more free thinking ideas and strategies. I would like to be able to compare our team and see how well others work on different teams and try to objectively identify who really excels working with certain people and who does not.  So my problem is when I try and use AVERAGEIFS or STDEV.S it only pulls the first available value. As the team and year changes I would Expect the average and the standard deviation to change as well. I have attached a spread sheet with my problem. Any clues with how to fix it would be greatly appreciated.  Thank You,Tripp  Read More

​

Overcoming Asymmetrical Routing in Azure Virtual WAN: A Collaborative Journey

In the rapidly evolving landscape of cloud networking, professionals often encounter complex challenges that demand innovative solutions. This blog post delves into a recent scenario involving Azure Virtual WAN (VWAN), where a team embarked on a collaborative journey to address asymmetrical routing issues. This case study not only highlights the technical intricacies but also underscores the importance of collaboration and knowledge sharing in the tech community.

The Challenge:

The core issue revolved around asymmetrical routing within an Azure VWAN architecture, which included two hubs located in different regions. The primary goal was to ensure seamless connectivity between a Palo Alto NGFW in one hub and Panorama in another, without disrupting the existing VPN default routes. The asymmetry in routing was particularly problematic for traffic intended to reach Panorama from the NGFW, as the return traffic defaulted through the VPN, deviating from the desired path.

The configuration issue with the firewall is acknowledged as a known issue. The only mitigation provided is documented in Microsoft’s official documentation. https://learn.microsoft.com/en-us/azure/virtual-wan/whats-new#known-issues

Possible solutions:

Exposing the Panorama to the public IP and creating a relay subnet for routing / advertising summary route from NCUS to SCUS for NCUS subnet (10.193.0.0/16)

​Microsoft Tech Community – Latest Blogs –Read More 

So I recenly took an udemy course part of which reinforcement learning was included as part of a broad AI course. So based on what I learnt about Q Learning(a type of reinforcement learning, under stereotypical Artificial Intelligence) , even though the implementation was not discussed in any full details, I had to go through a provided book to understand how it’s implemented in practice. And YouTube was not helping matters.

I need any additional opinions I can get from you, thanks.

Now in order for me to apply Q Learning to TicTacToe, I have to make the AI (Agent)always play X (makes the first move) for simplicity in my AI software development.

Q Learning algorithm is based on Bellmans Equation. Literally it’s about rewarding Favourable actions taken at different states and punishing unfavorable actions taken at different states also. Rewarding and punishing are one and the same number(variable) in the Bellmans equation. There is also the learning rate and discount factors, which are both variables too in the Bellmans equation.

For every action the agent takes at every state it always get a tiny reward (positive or negative), then after winning or losing it gets a much larger reward (positive or negative).

So how does the agent remembers all it has learnt, by looking at the q table,and checking for the action with the highest Q value. These Q values are updated by Bellmans equation whether positively or negatively.

Now the first challenge is I have to pair all the possible valid board configurations (states space) to all the possible valid actions that can be taken for each states in a Q table dictionary, and map all the pairs to Q values of 0 (for initialization). I will write a Python code that will generate this mapping and remove all impossible states (where X is greater than 0 by more than 1 is definitely invalid). Also make Q values for actions where ever is occupied by either X or O as – 1.0 to prevent agent from making such moves.

I will make 4 different players in different classes of the game software who would play with the agent at different stages of it’s learning automatically to update the Q values of every actions taken by agent in each game state, instead of waiting for the final result before updating Q value(my initial mistake when I was still learning about Q Learning) .

Below is for any of the 5 agents (Balanced, Quick Myopic, Quick Overplanner, Slow Myopic and Slow Overplanner) selected at the start of the training games. These agents have different combinations of hyperparameters (learning rate and discount factor)

To train the agent I will make it play against 4 heuristics players (players programed to play only in a certain way) all using different playing strategies.

For starting stage, agent will play with random player 1 for 2000 iterations of games and update it’s Q values for all the state action pairs it encountered.

Then for the next 2000 iterations, agent will play with a player 2 that always favor the center if available for its first move otherwise plays at any corner piece, otherwise plays any random available space.

Then for the next 2000 iters, agent will play against a player 3 that plays randomly until Agent is about to make a winning move, then blocks it. Not really trying to move, just blocking agent winning moves.

Then for the next 2000 iters, agent will play with a player 4 that tries to complete a line as soon as possible, by playing in corners that are impossible for agent to block, that is playing in triangular corners that leads to a definite win if agent doesn’t win on time.

Now create separate classes (.NET MAUI) for these four players that would train a selected agent chosen with options to pick the desired iterations of games for the training of the agent with that player.

For the reward system, +1 reward on completing a line. – 1 for allowing opponent win and not blocking it. +0. 5 for playing a position that can lead to a win in its next move. – 0.5 by playing in a position that cannot lead to a win in its next move (that is a move that doesn’t form a straight line of three with an empty cell anywhere in the line) and – 0.5 for playing along a line that is played by an opponent. This is the reward system rules.

So bells formula would be used for updating the Q value for every action taken for every state in the Q Table already defined in their respective Q Table json file for the particular agent being trained. We would use both learning rate, discount factor and reward from reward system for every action taken at every state.

Below is the link to the Q table (the brain of the AI);

https://www.kaggle.com/datasets/adedapoadeniran/tictactoe-q-learning-table

And below is the link to the code that generated the Q table.

https://www.kaggle.com/code/adedapoadeniran/reinforcement-learning-for-tictactoe-ai/

Thanks for your attention.

This was really mentally tasking to come up with and figure out.

​So I recenly took an udemy course part of which reinforcement learning was included as part of a broad AI course. So based on what I learnt about Q Learning(a type of reinforcement learning, under stereotypical Artificial Intelligence) , even though the implementation was not discussed in any full details, I had to go through a provided book to understand how it’s implemented in practice. And YouTube was not helping matters. I need any additional opinions I can get from you, thanks. Now in order for me to apply Q Learning to TicTacToe, I have to make the AI (Agent)always play X (makes the first move) for simplicity in my AI software development.Q Learning algorithm is based on Bellmans Equation. Literally it’s about rewarding Favourable actions taken at different states and punishing unfavorable actions taken at different states also. Rewarding and punishing are one and the same number(variable) in the Bellmans equation. There is also the learning rate and discount factors, which are both variables too in the Bellmans equation.For every action the agent takes at every state it always get a tiny reward (positive or negative), then after winning or losing it gets a much larger reward (positive or negative).So how does the agent remembers all it has learnt, by looking at the q table,and checking for the action with the highest Q value. These Q values are updated by Bellmans equation whether positively or negatively. Now the first challenge is I have to pair all the possible valid board configurations (states space) to all the possible valid actions that can be taken for each states in a Q table dictionary, and map all the pairs to Q values of 0 (for initialization). I will write a Python code that will generate this mapping and remove all impossible states (where X is greater than 0 by more than 1 is definitely invalid). Also make Q values for actions where ever is occupied by either X or O as – 1.0 to prevent agent from making such moves. I will make 4 different players in different classes of the game software who would play with the agent at different stages of it’s learning automatically to update the Q values of every actions taken by agent in each game state, instead of waiting for the final result before updating Q value(my initial mistake when I was still learning about Q Learning) .  Below is for any of the 5 agents (Balanced, Quick Myopic, Quick Overplanner, Slow Myopic and Slow Overplanner) selected at the start of the training games. These agents have different combinations of hyperparameters (learning rate and discount factor)To train the agent I will make it play against 4 heuristics players (players programed to play only in a certain way) all using different playing strategies.For starting stage, agent will play with random player 1 for 2000 iterations of games and update it’s Q values for all the state action pairs it encountered.Then for the next 2000 iterations, agent will play with a player 2 that always favor the center if available for its first move otherwise plays at any corner piece, otherwise plays any random available space.Then for the next 2000 iters, agent will play against a player 3 that plays randomly until Agent is about to make a winning move, then blocks it. Not really trying to move, just blocking agent winning moves.Then for the next 2000 iters, agent will play with a player 4 that tries to complete a line as soon as possible, by playing in corners that are impossible for agent to block, that is playing in triangular corners that leads to a definite win if agent doesn’t win on time. Now create separate classes (.NET MAUI) for these four players that would train a selected agent chosen with options to pick the desired iterations of games for the training of the agent with that player.For the reward system, +1 reward on completing a line. – 1 for allowing opponent win and not blocking it. +0. 5 for playing a position that can lead to a win in its next move. – 0.5 by playing in a position that cannot lead to a win in its next move (that is a move that doesn’t form a straight line of three with an empty cell anywhere in the line) and – 0.5 for playing along a line that is played by an opponent. This is the reward system rules. So bells formula would be used for updating the Q value for every action taken for every state in the Q Table already defined in their respective Q Table json file for the particular agent being trained. We would use both learning rate, discount factor and reward from reward system for every action taken at every state. Below is the link to the Q table (the brain of the AI);https://www.kaggle.com/datasets/adedapoadeniran/tictactoe-q-learning-tableAnd below is the link to the code that generated the Q table.https://www.kaggle.com/code/adedapoadeniran/reinforcement-learning-for-tictactoe-ai/Thanks for your attention.This was really mentally tasking to come up with and figure out.  Read More

​