This just started happening – I have not added any new applications or changed any settings.  Everything has been working for WEEKS, Months!  

Anyone else experiencing this on Windows 11?

Faulting application name: OUTLOOK.EXE, version: 16.0.17726.20160, time stamp: 0x668c527b
Faulting module name: ucrtbase.dll, version: 10.0.22621.3593, time stamp: 0x10c46e71
Exception code: 0xc0000005
Fault offset: 0x000000000005137c
Faulting process id: 0x0x2AD8
Faulting application start time: 0x0x1DADEBB287F3844
Faulting application path: C:Program FilesMicrosoft OfficerootOffice16OUTLOOK.EXE
Faulting module path: C:WindowsSystem32ucrtbase.dll
Report Id: dde66e9a-ce52-4e4b-81d0-4f88bb746329
Faulting package full name:
Faulting package-relative application ID:

​This just started happening – I have not added any new applications or changed any settings.  Everything has been working for WEEKS, Months!   Anyone else experiencing this on Windows 11?  Faulting application name: OUTLOOK.EXE, version: 16.0.17726.20160, time stamp: 0x668c527bFaulting module name: ucrtbase.dll, version: 10.0.22621.3593, time stamp: 0x10c46e71Exception code: 0xc0000005Fault offset: 0x000000000005137cFaulting process id: 0x0x2AD8Faulting application start time: 0x0x1DADEBB287F3844Faulting application path: C:Program FilesMicrosoft OfficerootOffice16OUTLOOK.EXEFaulting module path: C:WindowsSystem32ucrtbase.dllReport Id: dde66e9a-ce52-4e4b-81d0-4f88bb746329Faulting package full name:Faulting package-relative application ID:   Read More

​

Developing and deploying AI applications at scale requires a robust and flexible platform that can handle the complex and diverse needs of modern enterprises. This is where Azure AI services come into play, offering developers the tools they need to create customized AI solutions grounded in their organizational data.

One of the most exciting updates in Azure AI is the recent introduction of serverless fine-tuning for Phi-3-mini and Phi-3-medium models. This feature enables developers to quickly and easily customize models for both cloud and edge scenarios without the need for extensive compute resources. Additionally, updates to Phi-3-mini have brought significant improvements in core quality, instruction-following, and structured output, allowing developers to build more performant models without additional costs.

Azure AI continues to expand its model offerings, with the latest additions including OpenAI’s GPT-4o mini, Meta’s Llama 3.1 405B, and Mistral’s Large 2. These models provide customers with greater choice and flexibility, enabling them to leverage the best tools for their specific needs. The introduction of Cohere Rerank further enhances Azure AI’s capabilities, offering enterprise-ready language models that deliver superior search results in production environments.

The Phi-3 family of small language models (SLMs) developed by Microsoft has been a game-changer in the AI landscape. These models are not only cost-effective but also outperform other models of the same size and even larger ones. Developers can fine-tune Phi-3-mini and Phi-3-medium with their data to build AI experiences that are more relevant to their users, safely and economically. The small compute footprint and cloud and edge compatibility of Phi-3 models make them ideal for a variety of scenarios, from tutoring to enhancing the consistency and quality of responses in chat and Q&A applications.

Microsoft’s collaboration with Khan Academy is a testament to the potential of Phi-3 models. Khan Academy uses Azure OpenAI Service to power Khanmigo for Teachers, an AI-powered teaching assistant that helps educators across 44 countries. Initial data shows that Phi-3 outperforms most other leading generative AI models in correcting and identifying student mistakes in math tutoring scenarios.

Azure AI’s commitment to innovation is further demonstrated by the introduction of Phi Silica, a powerful model designed specifically for the Neural Processing Unit (NPU) in Copilot+ PCs. This model empowers developers to build apps with safe, secure AI experiences, making Microsoft Windows the first platform to have a state-of-the-art SLM custom-built for the NPU.

The Azure AI model catalog now boasts over 1,600 models from various providers, including AI21, Cohere, Databricks, Hugging Face, Meta, Mistral, Microsoft Research, OpenAI, Snowflake, and Stability AI. This extensive selection ensures that developers have access to the best tools for their AI projects, whether they are working on traditional machine learning or generative AI applications.

Building AI solutions responsibly is at the core of AI development at Microsoft. Azure AI evaluations enable developers to iteratively assess the quality and safety of models and applications, informing mitigations and ensuring responsible AI deployment. Additional Azure AI Content Safety features, such as prompt shields and protected material detection, are now “on by default” in Azure OpenAI Service, providing an extra layer of security for developers.

Learn more about these recent exciting developments by checking out this blog: Announcing Phi-3 fine-tuning, new generative AI models, and other Azure AI updates to empower organizations to customize and scale AI applications | Microsoft Azure Blog

​Developing and deploying AI applications at scale requires a robust and flexible platform that can handle the complex and diverse needs of modern enterprises. This is where Azure AI services come into play, offering developers the tools they need to create customized AI solutions grounded in their organizational data.
 
One of the most exciting updates in Azure AI is the recent introduction of serverless fine-tuning for Phi-3-mini and Phi-3-medium models. This feature enables developers to quickly and easily customize models for both cloud and edge scenarios without the need for extensive compute resources. Additionally, updates to Phi-3-mini have brought significant improvements in core quality, instruction-following, and structured output, allowing developers to build more performant models without additional costs.
 
Azure AI continues to expand its model offerings, with the latest additions including OpenAI’s GPT-4o mini, Meta’s Llama 3.1 405B, and Mistral’s Large 2. These models provide customers with greater choice and flexibility, enabling them to leverage the best tools for their specific needs. The introduction of Cohere Rerank further enhances Azure AI’s capabilities, offering enterprise-ready language models that deliver superior search results in production environments.
 
The Phi-3 family of small language models (SLMs) developed by Microsoft has been a game-changer in the AI landscape. These models are not only cost-effective but also outperform other models of the same size and even larger ones. Developers can fine-tune Phi-3-mini and Phi-3-medium with their data to build AI experiences that are more relevant to their users, safely and economically. The small compute footprint and cloud and edge compatibility of Phi-3 models make them ideal for a variety of scenarios, from tutoring to enhancing the consistency and quality of responses in chat and Q&A applications.
 
Microsoft’s collaboration with Khan Academy is a testament to the potential of Phi-3 models. Khan Academy uses Azure OpenAI Service to power Khanmigo for Teachers, an AI-powered teaching assistant that helps educators across 44 countries. Initial data shows that Phi-3 outperforms most other leading generative AI models in correcting and identifying student mistakes in math tutoring scenarios.
 
Azure AI’s commitment to innovation is further demonstrated by the introduction of Phi Silica, a powerful model designed specifically for the Neural Processing Unit (NPU) in Copilot+ PCs. This model empowers developers to build apps with safe, secure AI experiences, making Microsoft Windows the first platform to have a state-of-the-art SLM custom-built for the NPU.
 
The Azure AI model catalog now boasts over 1,600 models from various providers, including AI21, Cohere, Databricks, Hugging Face, Meta, Mistral, Microsoft Research, OpenAI, Snowflake, and Stability AI. This extensive selection ensures that developers have access to the best tools for their AI projects, whether they are working on traditional machine learning or generative AI applications.
Building AI solutions responsibly is at the core of AI development at Microsoft. Azure AI evaluations enable developers to iteratively assess the quality and safety of models and applications, informing mitigations and ensuring responsible AI deployment. Additional Azure AI Content Safety features, such as prompt shields and protected material detection, are now “on by default” in Azure OpenAI Service, providing an extra layer of security for developers.
 
Learn more about these recent exciting developments by checking out this blog: Announcing Phi-3 fine-tuning, new generative AI models, and other Azure AI updates to empower organizations to customize and scale AI applications | Microsoft Azure Blog  Read More

​

Hello Team,

Currently, we are running a Large set of ML recommendation models in the Azure Compute Cluster while running this model it will take more than 5 days.

How can we run a large number of datasets in the Azure Compute cluster? For Example Around (5 million) records.

Find the sample Code :

import os
import pickle
import argparse
import pandas as pd
import json
from azureml.core import Workspace, Datastore, Run
from azureml.data.dataset_factory import TabularDatasetFactory
import tempfile

# Load environment variables
from dotenv import load_dotenv
load_dotenv()

# Parse arguments
parser = argparse.ArgumentParser(“model_training”)
parser.add_argument(“–model_training”, type=str, help=”Model training data path”)
parser.add_argument(“–interaction”, type=str, help=”Interaction type”)
args = parser.parse_args()

# Workspace setup
workspace = Workspace(subscription_id=os.environ.get(“SUBSCRIPTION_ID”),
resource_group=os.environ.get(“RESOURCE_GROUP”),
workspace_name=os.environ.get(“WORKSPACE_NAME”))

print(‘Workspace:’, workspace)

# Get the datastore from the Azure ML workspace
datastore = Datastore.get(workspace, datastore_name=’data_factory’)
print(‘Datastore:’, datastore)

# Define the path to your Parquet files in the datastore
datastore_path = [(datastore, ‘sampless_silver/’)]

# Create a TabularDataset from the Parquet files in the datastore
dataset = TabularDatasetFactory.from_parquet_files(path=datastore_path)
print(‘Dataset:’, dataset)

# Convert the TabularDataset to a Pandas DataFrame
training_dataset = dataset.to_pandas_dataframe()
print(‘Training Dataset:’, training_dataset)

# Sample data
training_dataset = training_dataset.head(25000000)
training_dataset = training_dataset.sample(frac=1).reset_index(drop=True)
training_dataset[“views”] = pd.to_numeric(training_dataset[‘views’], errors=’coerce’)

df_selected = training_dataset.rename(columns={‘clientId’: ‘userID’, ‘offerId’: ‘itemID’, ‘views’: ‘views’})
df_selected = df_selected[[‘userID’, ‘itemID’, ‘views’]]
print(‘Selected Data:’, df_selected)

# Create and fit model
from lightfm import LightFM
from lightfm import cross_validation

dataset = Dataset()
dataset.fit(users=df_selected[‘userID’], items=df_selected[‘itemID’])
(interactions, weights) = dataset.build_interactions(df_selected.iloc[:, 0:3].values)

user_dict_label = dataset.mapping()[0]
item_dict_label = dataset.mapping()[2]

train_interactions, test_interactions = cross_validation.random_train_test_split(
interactions, test_percentage=0.25, random_state=np.random.RandomState(2016))

model = LightFM(loss=’warp’, no_components=1300, learning_rate=0.000001,
random_state=np.random.RandomState(2016), user_alpha=0.000005, max_sampled=100, k=100,
learning_schedule=’adadelta’, item_alpha=0.000005)
print(‘Model:’, model)

model.fit(interactions=train_interactions, epochs=2, verbose=True, num_threads=8)

user_dict_label = {str(key): value for key, value in user_dict_label.items()}
item_dict_label = {str(key): value for key, value in item_dict_label.items()}

# Save and upload model
with tempfile.TemporaryDirectory() as tmpdirname:
recommendation_model_offer = os.path.join(tmpdirname, “sample_recommendation_model.pkl”)
with open(recommendation_model_offer, ‘wb’) as f:
pickle.dump(model, f)

model_intersection = os.path.join(tmpdirname, “sample_training_intersection.pkl”)
with open(model_intersection, ‘wb’) as f:
pickle.dump(interactions, f)

model_user_dict = os.path.join(tmpdirname, “users_dict_label.json”)
with open(model_user_dict, ‘w’) as f:
json.dump(user_dict_label, f)

model_item_dict = os.path.join(tmpdirname, “items_dict_label.json”)
with open(model_item_dict, ‘w’) as f:
json.dump(item_dict_label, f)

datastore.upload_files(
files=[recommendation_model_offer, model_intersection, model_user_dict, model_item_dict],
target_path=’SAMPLE_MODEL_TRAINING/’,
overwrite=True
)
print(‘Files uploaded to datastore’)

# Register the model
register_name = f”{args.interaction}_light_fm_recommendation_model”
Model.register(workspace=workspace, model_path=tmpdirname, model_name=register_name,
tags={‘affinity’: args.interaction, ‘sample’: ‘recommendation’})
print(‘Model registered’)

Please share the feedback. Thanks!

​Hello Team,Currently, we are running a Large set of ML recommendation models in the Azure Compute Cluster while running this model it will take more than 5 days.How can we run a large number of datasets in the Azure Compute cluster? For Example Around (5 million) records.Find the sample Code :import osimport pickleimport argparseimport pandas as pdimport jsonfrom azureml.core import Workspace, Datastore, Runfrom azureml.data.dataset_factory import TabularDatasetFactoryimport tempfile# Load environment variablesfrom dotenv import load_dotenvload_dotenv()# Parse argumentsparser = argparse.ArgumentParser(“model_training”)parser.add_argument(“–model_training”, type=str, help=”Model training data path”)parser.add_argument(“–interaction”, type=str, help=”Interaction type”)args = parser.parse_args()# Workspace setupworkspace = Workspace(subscription_id=os.environ.get(“SUBSCRIPTION_ID”),resource_group=os.environ.get(“RESOURCE_GROUP”),workspace_name=os.environ.get(“WORKSPACE_NAME”))print(‘Workspace:’, workspace)# Get the datastore from the Azure ML workspacedatastore = Datastore.get(workspace, datastore_name=’data_factory’)print(‘Datastore:’, datastore)# Define the path to your Parquet files in the datastoredatastore_path = [(datastore, ‘sampless_silver/’)]# Create a TabularDataset from the Parquet files in the datastoredataset = TabularDatasetFactory.from_parquet_files(path=datastore_path)print(‘Dataset:’, dataset)# Convert the TabularDataset to a Pandas DataFrametraining_dataset = dataset.to_pandas_dataframe()print(‘Training Dataset:’, training_dataset)# Sample datatraining_dataset = training_dataset.head(25000000)training_dataset = training_dataset.sample(frac=1).reset_index(drop=True)training_dataset[“views”] = pd.to_numeric(training_dataset[‘views’], errors=’coerce’)df_selected = training_dataset.rename(columns={‘clientId’: ‘userID’, ‘offerId’: ‘itemID’, ‘views’: ‘views’})df_selected = df_selected[[‘userID’, ‘itemID’, ‘views’]]print(‘Selected Data:’, df_selected)# Create and fit modelfrom lightfm import LightFMfrom lightfm import cross_validationdataset = Dataset()dataset.fit(users=df_selected[‘userID’], items=df_selected[‘itemID’])(interactions, weights) = dataset.build_interactions(df_selected.iloc[:, 0:3].values)user_dict_label = dataset.mapping()[0]item_dict_label = dataset.mapping()[2]train_interactions, test_interactions = cross_validation.random_train_test_split(interactions, test_percentage=0.25, random_state=np.random.RandomState(2016))model = LightFM(loss=’warp’, no_components=1300, learning_rate=0.000001,random_state=np.random.RandomState(2016), user_alpha=0.000005, max_sampled=100, k=100,learning_schedule=’adadelta’, item_alpha=0.000005)print(‘Model:’, model)model.fit(interactions=train_interactions, epochs=2, verbose=True, num_threads=8)user_dict_label = {str(key): value for key, value in user_dict_label.items()}item_dict_label = {str(key): value for key, value in item_dict_label.items()}# Save and upload modelwith tempfile.TemporaryDirectory() as tmpdirname:recommendation_model_offer = os.path.join(tmpdirname, “sample_recommendation_model.pkl”)with open(recommendation_model_offer, ‘wb’) as f:pickle.dump(model, f)model_intersection = os.path.join(tmpdirname, “sample_training_intersection.pkl”)with open(model_intersection, ‘wb’) as f:pickle.dump(interactions, f)model_user_dict = os.path.join(tmpdirname, “users_dict_label.json”)with open(model_user_dict, ‘w’) as f:json.dump(user_dict_label, f)model_item_dict = os.path.join(tmpdirname, “items_dict_label.json”)with open(model_item_dict, ‘w’) as f:json.dump(item_dict_label, f)datastore.upload_files(files=[recommendation_model_offer, model_intersection, model_user_dict, model_item_dict],target_path=’SAMPLE_MODEL_TRAINING/’,overwrite=True)print(‘Files uploaded to datastore’)# Register the modelregister_name = f”{args.interaction}_light_fm_recommendation_model”Model.register(workspace=workspace, model_path=tmpdirname, model_name=register_name,tags={‘affinity’: args.interaction, ‘sample’: ‘recommendation’})print(‘Model registered’)Please share the feedback. Thanks!  Read More

​

Part of our fleet remains Entra Hybrid Join (as computers are refreshed, they are Entra Joined instead). We apply Windows Security Baselines through both Group Policy and Intune. Recently, we evaluated the differences between the two baselines and determined they are nearly identical. Accordingly, we decided to disable GPO based security baselines for Entra Hybrid Joined devices and let Intune push security settings for the baseline instead.

Here’s the expected behavior:

Security baseline settings are set by both Intune and GPO. By default, GPO wins, so the Intune setting is not applied.When the GPO settings are removed, at some point in the next 24 hours (I believe it happens every 8) all Intune policies are reapplied whether or not they have changed. With the GPOs gone, MDM policies that were once blocked by group policy are applied.The end result: all security policies are applied, but most of them are coming from Intune (MDM) instead of from GPOs.

However, this is not what is happening. While Intune claims the security baseline have applied, the settings that were once overridden by GPOs never apply and the computer effectively has no security baseline.

Here’s what I’ve done to try to fix this:

Make a copy of the existing baseline with a new name and assign it to the computers, unassign the original baseline. This does not work. The policies claim to have applied, but never apply on the endpoint.Change a single setting in the baseline hoping the change triggers the whole configuration reapplying. The endpoint only applies the changed setting, other settings in the baseline do not get applied.Unassign the baseline entirely, wait for the computer to sync and reassign the baseline. This works, but is not a viable solution for a large fleet of computers. This would be fine if all of our computers were receiving GPO updates regularly, but they’re not (they are remote). This only works if the computer syncs one time while no settings are applied and again after the configurations are reassigned. We can’t negotiate the timing on this for our whole fleet of computers.Apply the policy that makes MDM policies take precedence over GPOs. This did not work.

Here’s what we’re not willing to try (I’m preempting some of Microsoft’s usual boilerplate responses):

We will not reset the computers – there are too many for this to be a scalable solution.We will not unjoin and rejoin the computers from MDM – there are too many for this to be a scalable solution.

While I’m tempted to open a support case with Microsoft, this has only ever been a time-consuming and frivolous process. I expect they would pass the ticket around and eventually apologize to me when they decide this is a support case I should actually pay for.

Why would MDM policies not apply even after the group policies that once conflicted with them have been removed? This is impacting all Entra Hybrid Joined computers, the vast majority of which are running the latest build of Windows 11 23H2. Some of these computers have sat for 48 hours in this state, so I don’t think this is something that will be resolved with time.

Any advice would be greatly appreciated!

​Part of our fleet remains Entra Hybrid Join (as computers are refreshed, they are Entra Joined instead). We apply Windows Security Baselines through both Group Policy and Intune. Recently, we evaluated the differences between the two baselines and determined they are nearly identical. Accordingly, we decided to disable GPO based security baselines for Entra Hybrid Joined devices and let Intune push security settings for the baseline instead. Here’s the expected behavior:Security baseline settings are set by both Intune and GPO. By default, GPO wins, so the Intune setting is not applied.When the GPO settings are removed, at some point in the next 24 hours (I believe it happens every 8) all Intune policies are reapplied whether or not they have changed. With the GPOs gone, MDM policies that were once blocked by group policy are applied.The end result: all security policies are applied, but most of them are coming from Intune (MDM) instead of from GPOs.However, this is not what is happening. While Intune claims the security baseline have applied, the settings that were once overridden by GPOs never apply and the computer effectively has no security baseline. Here’s what I’ve done to try to fix this:Make a copy of the existing baseline with a new name and assign it to the computers, unassign the original baseline. This does not work. The policies claim to have applied, but never apply on the endpoint.Change a single setting in the baseline hoping the change triggers the whole configuration reapplying. The endpoint only applies the changed setting, other settings in the baseline do not get applied.Unassign the baseline entirely, wait for the computer to sync and reassign the baseline. This works, but is not a viable solution for a large fleet of computers. This would be fine if all of our computers were receiving GPO updates regularly, but they’re not (they are remote). This only works if the computer syncs one time while no settings are applied and again after the configurations are reassigned. We can’t negotiate the timing on this for our whole fleet of computers.Apply the policy that makes MDM policies take precedence over GPOs. This did not work.Here’s what we’re not willing to try (I’m preempting some of Microsoft’s usual boilerplate responses):We will not reset the computers – there are too many for this to be a scalable solution.We will not unjoin and rejoin the computers from MDM – there are too many for this to be a scalable solution.While I’m tempted to open a support case with Microsoft, this has only ever been a time-consuming and frivolous process. I expect they would pass the ticket around and eventually apologize to me when they decide this is a support case I should actually pay for. Why would MDM policies not apply even after the group policies that once conflicted with them have been removed? This is impacting all Entra Hybrid Joined computers, the vast majority of which are running the latest build of Windows 11 23H2. Some of these computers have sat for 48 hours in this state, so I don’t think this is something that will be resolved with time. Any advice would be greatly appreciated!  Read More

​