What a Model’s Written Reasoning Reflects
Written reasoning reads like an explanation. Whether it reflects what the model computed is a separate question.
Worth a read: devlinliles.com/what-a-model…
#LLM #Interpretability
22
How a Model Decides to Answer or Decline
A refusal isn’t a lookup against banned topics. Something inside the model made that choice. Know what your guardrails rely on.
Dive in: devlinliles.com/how-a-model-…
#AISafety #Guardrails
14
How Researchers Look Inside a Model
Interpretability methods show what happens inside a model while it runs. Some are better established than others.
Read the post: devlinliles.com/how-research…
#Interpretability #AIResearch
5
What a Seed Controls and What Batch Invariance Controls
A seed fixes the random choices, not everything that can change an output. Batch invariance covers the rest.
See it here: devlinliles.com/what-a-seed-…
#LLM #Reproducibility
14
Where the Variation Comes From When Temperature Is Zero
Set temperature to zero and you still don’t always get the same answer. The variation has real sources.
Full post: devlinliles.com/where-the-va…
#LLM #MLOps
6
Where a Language Model’s Memory Lives
A model has several kinds of memory: weights, context window, and what you bolt on. Each fails differently.
Check it out: devlinliles.com/where-a-lang…
#LLM #AIExplained
10
What Training Puts in the Weights
Training doesn’t store documents. It stores patterns, compressed into billions of numbers. That decides what a model recalls and what it approximates.
Worth a read: devlinliles.com/what-trainin…
#LLM #AIExplained
14
What One Forward Pass Computes
Every token comes from one forward pass with a fixed amount of computation. Knowing what it can’t do sets honest limits on a single step.
Dive in: devlinliles.com/what-one-for…
#LLM #MachineLearning
7
What a Language Model Reads: Tokens and Embeddings
A language model doesn’t read words. It reads tokens, converted into numbers that carry meaning. That explains a lot of odd behavior.
Read it here: devlinliles.com/what-a-langu…
#LLM #AIExplained
6
Who Owns the Permission on a Managed Endpoint
An AI tool asks for permission on a managed device. User, admin, or vendor? Unclear ownership is how access spreads unchecked.
See it here: devlinliles.com/who-owns-the…
#AIGovernance #Security
7
Three Beats After the Coalition
A coalition forms around an AI policy position. The three beats that follow tend to decide whether it holds.
Full post: devlinliles.com/three-beats-…
#AIPolicy #AIGovernance
10
The Number to Model Is the Cached Input Price
The headline token price isn’t what drives your bill. For agent workloads that reuse context, the cached input price is.
Check it out: devlinliles.com/the-number-t…
#AICosts #LLM
1
11
Seven Minutes and Twenty-Four Hours
Seven minutes to detect. Twenty-four hours to resolve. Those two numbers say more about maturity than any accuracy score.
Check it out: devlinliles.com/seven-minute…
#AIOps #MLOps
18
Scope Violations as a Release Blocker
If an agent acts outside its granted scope in testing, that is not a warning. It is a release blocker.
Worth a read: devlinliles.com/scope-violat…
#AIAgents #AIGovernance
12
Routing Moved From the Architecture Diagram to the Product
Model routing used to be a box on the diagram. Now a vendor makes the call for you. Decide what to keep control of.
See it here: devlinliles.com/routing-move…
#LLM #AIStrategy
1
1
24
Open Weights You Can Run and Training You Can’t Audit
You can run the weights on your own hardware. You still can’t see the training data. Open weights are not open provenance.
Full post: devlinliles.com/open-weights…
#OpenWeights #AIRisk
1
55
Matching the Sovereignty Tier to the Data Class
Not all data needs the same residency and control. Classify the data, then match each class to a sovereignty tier.
Dive in: devlinliles.com/matching-the…
#DataSovereignty #AIGovernance
25
Mark the Machine’s Paragraphs
When AI drafts part of a document, readers should be able to tell which parts. It supports review, accountability, and trust.
Check it out: devlinliles.com/mark-the-mac…
#ResponsibleAI #AITransparency
14
An Agent’s Outbound Traffic Is Someone Else’s Load
Your agent calls external services at machine speed. That lands on someone else’s infrastructure. Throttle and back off.
Read it here: devlinliles.com/an-agents-ou…
#AIAgents #APIDesign
10
A Second Reward Hack With Lower Stakes
The first reward hack gets the attention. The second, with lower stakes, is the cheaper place to catch the pattern.
Worth a read: devlinliles.com/a-second-rew…
#AISafety #RewardHacking
8
