What's changed in Ultra Transcenders AI-300

Kindle readers: Amazon sends updates to existing copies only for corrections. To get the latest version, turn on automatic book updates in Amazon's Manage Your Content and Devices. Paperbacks printed after an update include it.

Version 1.1: 25 October 2026 (kindle, paperback)

• Retirement: Ch 7: Phi-3-mini (128k context) retired on 30 August 2025; Phi-4-mini-instruct replaces it.
• Retirement: Ch 7: Key takeaways now note that Phi-3-mini is retired.
• Retirement: Ch 11: davinci-002 retired on 27 January 2025.
• Retirement: Glossary: davinci-002 marked as retired on 27 January 2025.
• Correction: Ch 11: Training types now include Data zone (US data zone only), alongside Standard, Global and Developer.
• Retirement: Ch 1: Azure Machine Learning data import (preview) and data connections (preview) retired on 30 September 2026; bring SQL-like sources in through Microsoft Fabric and use a OneLake or Blob/ADLS Gen2 datastore instead.
• Correction: Ch 3: delay_evaluation delays the first early-termination check until the first multiple of evaluation_interval at or after it, so with 1 and 5 the first check is at interval 5 (not after skipping five).
• Correction: Ch 4: The Responsible AI dashboard is created for an already registered MLflow (scikit-learn) model, so you register first and use the dashboard before promotion, not before registration.
• Correction: Ch 4: Chapter introduction corrected to say the Responsible AI dashboard checks a registered model before promotion.
• New: Ch 1: Job data access modes: ro_mount, download and direct for inputs; rw_mount and upload for outputs.
• New: Ch 1: Creating a compute cluster in the SDK or CLI, and which settings (node counts, idle time) can be updated later.
• New: Ch 1: Connecting to a workspace with MLClient and MLClient.from_config (config.json).
• New: Ch 1: The three ways to define an environment: image, image plus conda_file, or a Dockerfile build context.
• New: Ch 1: Which credential each datastore type takes (ADLS Gen2: service principal or identity) and why identity-based access gives per-user permissions.
• New: Ch 1: Spark compute options: serverless Spark compute or an attached Synapse Spark pool.
• New: Ch 1: Kubernetes compute (AKS or Arc-enabled clusters) for training and online inference, with instance types.
• New: Ch 2: SSH setup order on a compute instance: ssh-keygen, add the public key to the Git host, clone with the SSH URL.
• New: Ch 2: What the hbi_workspace (confidential data) flag does.
• New: Ch 3: Bayesian sampling doesn’t support early termination policies.
• New: Ch 3: The wider MLflow logging API: log_metrics, log_text, log_dict, log_artifact(s), log_figure and log_image.
• New: Ch 3: Query and compare runs with mlflow.search_runs; it returns only the latest value of each metric.
• New: Ch 3: Adding a Jupyter kernel on a compute instance (conda environment plus ipykernel) and using %pip in notebooks.
• New: Ch 3: AutoML image data is JSONL wrapped in an MLTable; NER uses CoNLL format.
• New: Ch 3: AutoML image primary metrics: accuracy, iou and mean_average_precision.
• New: Ch 3: Defining a command job with command() and submitting it with ml_client.jobs.create_or_update().
• New: Ch 3: Creating components from YAML with load_component or from a Python function with @command_component.
• New: Ch 4: Registered model tags (editable) and properties (fixed) can store metrics for comparing versions.
• New: Ch 4: What counterfactual what-if and causal analysis answer in the Responsible AI dashboard.
• New: Ch 5: Online endpoint authentication modes: key (default), aml_token and aad_token, with aad_token on managed online endpoints only.
• New: Ch 5: MLflow models deployed to an online endpoint without outbound internet access must be packaged first (with_package=True or –with-package).
• New: Ch 5: Debug online deployments locally in Docker with –local or local=True before deploying to Azure.
• New: Ch 5: How production inference data is captured for monitoring: the data collector (model_inputs/model_outputs) writes to workspace blob storage.
• New: Ch 5: Trigger retraining from model monitoring via the Event Grid Run status changed event filtered on azureml_modelmonitor_threshold_breached, handled by Logic Apps or Functions.
• New: Ch 5: Batch deployments need a registered model and a compute cluster.
• New: Ch 10: The semantic ranker needs a semantic configuration on the index.
• New: Ch 11: Fine-tuning job inputs and hyperparameters: n_epochs, batch_size, learning_rate_multiplier and seed.

Version 1.0: 4 October 2026 (kindle, paperback)

• New: First edition.

About the book · AI-300 terms in the glossary