zachzang.bsky.social
@zachzang.bsky.social
max_tokens is a guillotine, not a hint

Generation stops dead at the limit, mid-sentence, and stop_reason tells you it happened.

#CCAOF #ClaudeAPI #LLM #APIDesign

Full CCAO-F explanation, free: https://navyduck.com/anthropic/ccao-f/q054-output-tokens-limit
September 26, 2026 at 5:00 PM
A tool result goes back as a user message

Append the assistant's tool_use block, then a user message carrying tool_result.

#CCARP #ClaudeAPI #LLM #ToolUse

Full CCAR-P explanation, free: https://navyduck.com/anthropic/ccar-p/q010-team-implementing-tool-use-function
September 25, 2026 at 5:00 PM
Kafka is the shock absorber in an observability backend

It decouples telemetry ingestion from processing so a traffic burst doesn't drop data.

#C1000189

Full C1000-189 explanation, free: https://navyduck.com/ibm/c1000-189/q085-during-instana-architecture-planning-a-solutions-architect-must
September 24, 2026 at 5:00 PM
Bronze to Silver is where raw data becomes trustworthy

Silver cleans, deduplicates and conforms to a schema. Aggregation is Gold's job.

#C1000190 #DataEngineering #Lakehouse

Full C1000-190 explanation, free: https://navyduck.com/ibm/c1000-190/q018-what-transformation-occurs-when-data-moves
September 23, 2026 at 5:00 PM
🚀 NEW on Udemy: AWS Certified AI Business Strategist (AIB-C01) Practice Tests. 200 questions, 4 timed exams, detailed explanation + original diagram on every answer. No coding. Beta exam.

👉 https://www.udemy.com/course/aws-certified-ai-business-strategist-aib-c01-practice/

#AWS #AICertification
September 22, 2026 at 8:34 PM
🚀 NEW on Udemy: Google Agentic Architect (GCP-PAA) Practice Tests. 200 questions, 4 timed exams, every answer with a detailed explanation and a diagram. Beta exam; knowledge portion only.

👉 https://www.udemy.com/course/google-agentic-architect-gcp-paa-practice-tests/

#GoogleCloud #AgenticAI
September 22, 2026 at 8:34 PM
'Cancel my order' and 'cancel my subscription' are two intents, not one

If the object changes what happens next, it belongs in the intent, not an entity.

#C1000180 #Chatbots

Full C1000-180 explanation, free: https://navyduck.com/ibm/c1000-180/q164-developer-needs-distinguish-between-cancel-my
September 22, 2026 at 5:00 PM
Approval rights belong to a role, not to a notification list

In OpenPages, restrict who can approve by assigning the Approver role to that group.

#C1000195 #AIGovernance #IBM

Full C1000-195 explanation, free: https://navyduck.com/ibm/c1000-195/q079-openpages-administrator-wants-ensure-that-only
September 21, 2026 at 5:00 PM
Learning rate and tree count do not simply trade off

Low shrinkage needs more trees but generalises better; 0.01x500 is not 0.1x50.

#SASAIML #MachineLearning #GradientBoosting #SAS

Full SAS-AIML explanation, free: https://navyduck.com/sas/aiml/q250-gradient-boosting-model-trained-shrinkage
September 20, 2026 at 5:00 PM
When a moderation model flags one dialect three times as often

A dialect-skewed false positive rate is linguistic bias from unrepresentative training data.

#AAIR #AIRisk #AIGovernance #Fairness

Full AAIR explanation, free: https://navyduck.com/isaca/aair/q113-ai-model-used-automated-content
September 19, 2026 at 5:00 PM
Copilot did not leak your HR files, SharePoint permissions did

Oversharing is a permissions problem Copilot makes visible; fix it with Purview DSPM for AI.

#AB650 #Microsoft365 #Copilot

Full AB-650 explanation, free: https://navyduck.com/microsoft/ab-650/ab-650-e1-q035-data-oversharing-copilot
September 18, 2026 at 5:00 PM
Perfect recall is trivially easy to fake

Recall 1.0 means zero false negatives, and says nothing at all about false positives.

#CDLS #MachineLearning #DataScience #ModelEvaluation

Full CDLS explanation, free: https://navyduck.com/deep-learning/cdls/q173-recall-of-one-meaning
September 17, 2026 at 5:00 PM
A feedforward network has no idea what order things came in

FNNs assume inputs are independent, so sequence order is invisible to them.

#CAID #DeepLearning #NeuralNetworks #NLP

Full CAID explanation, free: https://navyduck.com/blockchain-council/caid/q095-if-you-attempt-to-use
September 16, 2026 at 5:00 PM
Parent and child agents do not share variables

Copilot Studio child agents get context through declared input parameters, not shared memory.

#AB620 #CopilotStudio

Full AB-620 explanation, free: https://navyduck.com/microsoft/ab-620/ab-620-e3-q022-child-agent-context-passing-variable-handoff
September 15, 2026 at 5:00 PM
'Not used recently' and 'not valuable' are different things

Prune agent memory on recency x retrieval frequency x uniqueness, never on age alone.

#GH600 #AIAgents #LLM #VectorDatabase

Full GH-600 explanation, free: https://navyduck.com/github/gh-600/gh-600-e6-q023-memory-expiration-pruning
September 14, 2026 at 5:00 PM
AI projects start with a decision, not a model

Define the decision you want to change, then ask whether ML is even the right tool.

#AIEG #AI #MachineLearning #DataScience

Full AI-EG explanation, free: https://navyduck.com/ai-essentials/ai-eg/ai-eg-e6-q010-at-the-start-of-an-applied
September 13, 2026 at 5:00 PM
Why RNNs need backpropagation through time

BPTT unrolls the loop so early steps get credit for later errors.

#AIEG #DeepLearning #RNN #MachineLearning

Full AI-EG explanation, free: https://navyduck.com/ai-essentials/ai-eg/ai-eg-e3-q005-why-do-recurrent-neural-networks-use
September 12, 2026 at 5:00 PM
Two GPU tools for data too big to look at

cuGraph does graph analytics on the GPU; Omniverse does real-time 3D simulation.

#NCAAIIO #AI #DataScience #GPUComputing

Full NCA-AIIO explanation, free: https://navyduck.com/nvidia/ai-infrastructure/nca-aiio/q117-a-research-team-needs-to
September 11, 2026 at 5:00 PM
Your Pandas code, running on a GPU

Pandas, Scikit-learn and NetworkX are excellent and they are all built for a CPU, which is why they stall on genuinely large data.

#NCAAIIO

Full NCA-AIIO explanation, free: https://navyduck.com/nvidia/ai-infrastructure/nca-aiio/q028-a-data-science-team-wants
September 10, 2026 at 5:00 PM
September 9, 2026 at 5:00 PM
Owning versus renting your GitHub

GitHub Enterprise comes in two shapes and the difference is who runs the hardware.

#GH100 #GitHubEnterprise #DevOps #SelfHosted

Full GH-100 explanation, free: https://navyduck.com/github/gh-100/q001-which-github-enterprise-product-runs-entirely
September 8, 2026 at 5:00 PM
What Spark's Imputer actually does to your NaNs

Imputer is the Spark ML stage that fills missing numeric values, and the default strategy is the column mean.

#DBML #DataScience #MachineLearning #ApacheSpark

Full DBML explanation, free: https://navyduck.com/databricks/dbml/q232-imputer-spark-ml
September 7, 2026 at 5:00 PM
The four stages of an MLflow model's life

Every model in the MLflow Model Registry moves through the same four stages: None, Staging, Production, Archived.

#DBML #MLflow #MLOps #ModelRegistry

Full DBML explanation, free: https://navyduck.com/databricks/dbml/q003-mlflow-model-registry-stages
September 6, 2026 at 5:00 PM
Least privilege applies to GRC admin roles too

External vendor accounts should never hold platform admin by default either.

#CISRC #GRC #RiskManagement #Compliance

Full CIS-RC explanation, free: https://navyduck.com/cis-rc/implementation-planning/q046
September 5, 2026 at 5:00 PM
APT means capability and persistence, not just any breach

Zero trust is a defensive architecture and C2 is a channel, neither is a threat actor label.

#SPLK5001 #CyberSecurity

Full SPLK-5001 explanation, free: https://navyduck.com/splk-5001/threat-and-attack-types-motivations-and-tactics/q061
September 4, 2026 at 5:00 PM