Terry Miles

Miles treats expert guidance as another name for context, argues that reusable context from one initiative pays off in the next, and warns readers not to let AI manage their context for them.

• The memory post adds a security dimension: customer session context is confidential, and procedures from colleagues or agents are proprietary.
• Keeping control also means deciding what the model never holds.
• In the diagram post, shared understanding of context across product and systems colleagues is the payoff—a new full stack, as they call it.

The procurement post argues that starting with cost efficiency delivers context and fine-tuned models that carry into later initiatives.

• It names suppliers, transactions, payments and contracts as context to set.
• The rubric post makes the same point: expert guidance is context, and context is capital.

The memory post sharpens the ownership argument: control means not only authoring your own context, but also keeping confidential and proprietary context out of the model’s reach.

Read original · Original 2 · Original 3 · Original 4

A simple question reaches value quickly and is easier to fund.

• But choosing one whose answers require many stores forces the harder architecture, security and access work up front, avoiding rework later.
• The aim is not the first answer, but a capability—and reusable context—built across the whole estate.

Cost efficiency and unit economics over time provide a familiar procurement metric, making investment in visibility easier to sell.

• But answering the question requires contracts, invoices and purchase orders, moving the model beyond public internet access and into early joint work with IT and information security.
• Miles argues that small language models often suffice because repositories are limited, much of the data is structured and speed is not critical.
• They suggest using top-of-mind tasks such as tracking big spenders or time-sensitive renewals to get people and models working together.

Read original

They divide that memory into short-term, long-term semantic, long-term procedural and long-term episodic, then map the categories onto a familiar distinction between a concept to recognise and a procedure to carry out. The practical implications: short-term memory is a question of personalisation as much as time; short-term load can complicate long-term memory operations; and procedural knowledge lives in logic and algorithms, not in the prompt.

A gaming simulation architecture illustrates where personalisation demands are highest: interactions run long, requiring more short-term memory. Episodic memory—recalling prior interactions with a given user—is treated as a separate implementation, analogous to the human transfer from short- to long-term memory.

Read original

Instead, they turn someone else’s diagram or paper into an exercise a team can run: reorder elements chronologically into a value-stream checklist, attach them to services and traces, or use Wardley mapping to decide what to build, buy or partner.

• The aim is a shared vocabulary for systems, security and product teams before they argue about features.
• The teaching anchor belongs to the guest; the contribution is the application.

The diagram Miles uses—Gennaro Cuofano’s elaboration of Andrej Karpathy’s comment on a shift from ChatGPT to Claude—includes the MCP box, the orchestration and infrastructure layer, and MCP-connected tools and data sources.

• Miles suggests reordering these into a value stream for a feature or release, then using an observability platform to associate the items with services and resources.
• Tracing a varied set of prompts across stores and tools can then get the conversation started with spans and heat maps.

Read original

They highlight two details: separating essential from non-essential attributes of success helps a model prioritise and judge how far to reason; and checklists are the clearest format for those inputs because they lay out the elements of a finished piece of work.

The concrete example is a rubric’s definition-of-done logic: a checklist that marks what must be present and what merely helps, so the model can distinguish a complete answer from an elaborate one.

Read original

A newcomer can see how one diagram becomes a practical value-stream exercise and shared vocabulary for product, data, and infrastructure teams.

Terry Miles highlights Gennaro Cuofano's diagram, which elaborates on Andrej Karpathy's observation of a shift from ChatGPT to Claude.

• They value its clear explanation of context engineering and MCP, especially for system architecture, observability, information security, and infrastructure, while remaining accessible to teams focused on features rather than systems.
• They suggest two ways to use the diagram.
• First, take the blue MCP box, the Orchestration & Infrastructure Layer, and the MCP-Connected Tools & Data Sources section, then reorder the relevant items into a value stream or checklist for adding a new feature or major release.
• Second, use an observability platform to associate diagram items with services and resources, trace varied prompts across stores and tools, and use spans and heat maps to start conversations from a systems and services perspective.
• As a bonus, take traces, spans, or team-focused value streams into Wardley Mapping to view the business as capabilities to build, buy, or partner on.
• They call the result “a shared understanding of context engineering and a new full stack.”

For brands planning a campaign

For an enterprise software company, this post offers a ready-made workshop for aligning product, data, security, and infrastructure teams around integrations and personalization. Use the MCP, orchestration, and data-source layers as a checklist for where AI features touch customer data, then use observability traces to identify which integrations and memory stores need attention.

A practical starting point for turning expert judgment into model inputs.

Terry Miles connects Scale AI's paper, Rubrics as Rewards, to rubric-building skills from education, business analysis, and instructional design.

• The paper describes a framework for human-authored inputs to help language models with reinforcement learning.
• They flag two details: distinguishing essential from non-essential attributes of success helps models prioritize tasks and decide how far to go as they deduce; checklists work best because they clearly separate the elements of a definition of done.
• Their closing point: expert guidance is another name for context.
• “Context is capital,” and teams should continue to own and articulate their business context rather than let AI manage it for them.

For brands planning a campaign

For a product team adding AI features, this is a concrete way to encode domain expertise in evaluation and reward criteria. Use checklist-style rubrics to define good answers on personalization, data integration, or procurement, and keep ownership of those criteria rather than outsourcing them to a model.

A useful starting point for the vocabulary needed to design personalized AI systems.

Terry Miles recommends Paul Iusztin's Decoding ML Substack and two posts on context engineering and memory.

• They define context engineering as maximizing a language model's output quality by presenting the right knowledge at the right time—effectively curating the memory the model has for the task.
• Short-term memory depends not just on time but also on personalization: models need to remember an ongoing session or interaction to personalize their outputs.
• The post's diagram shows an architecture for a gaming simulation system, where personalization demands are high, interactions run long, and more short-term memory is required.
• The cognitive load of short-term memory can complicate long-term memory operations.
• They map memory types to knowledge as concept or procedure.
• Short-term memory is almost purely conceptual, confidential when it comes from customers and proprietary when it comes from colleagues or agents.
• Long-term semantic memory is also conceptual and often proprietary.
• Long-term procedural memory captures how to engage and deduce, generally through logic or algorithms valuable and proprietary to a team.
• Long-term episodic memory recalls prior interactions with a given user; it is conceptual, confidential, and a separate implementation analogous to the human process of transferring information from short-term to long-term memory.

For brands planning a campaign

For an enterprise software company improving data integration and personalization, this post turns memory architecture into a set of decisions: what stays in-session, what becomes semantic or procedural knowledge, and what must be recalled from prior interactions. Use it to plan where customer and internal proprietary data live, how access is controlled, and which memory type a personalization feature needs.

A practical adoption blueprint connecting cost efficiency, sensitive data stores, small language models, and cross-functional allies.

Terry Miles advises procurement colleagues to start with a small set of simple questions that are ambitious in relation to the full scope of relevant data stores.

• They recommend tackling the complexity of context and legacy data, not avoiding it: keep questions simple to get value quickly, while addressing the full range of data for long-term success.
• The example is tracking cost efficiency and unit economics over time.
• Because cost efficiency is a common procurement metric, investments in visibility are easier to sell.
• Answering the question requires contracts, invoices, and purchase orders—sensitive data stores.
• That rules out public internet access and makes early collaboration with IT, information security, and other allies essential.
• Procurement data sets can be vast and messy, but modest Small Language Models can often do the job: speed is not necessarily critical, repositories are limited, and much of the data is structured.
• Teams should map the full scope of data stores to avoid rework, use lessons from cost efficiency to build context and fine-tune models for later initiatives, and include suppliers, transactions, payments, and contracts in the context they set.
• The human work of customizing and fine-tuning a model should reflect team priorities.
• High-priority tasks such as big spenders or time-sensitive renewals can get the team and robots working together.
• They return to the same starting point: use simple, valuable questions that require AI to reach all your data stores to build procurement AI capabilities.

For brands planning a campaign

For an enterprise software company assessing Terry Miles, this is the clearest procurement-specific playbook: pick one high-value question, map every relevant data store, plan for private or on-premises deployment, involve security and IT allies early, and expect small language models plus reusable context and fine-tuning rather than a single public model. Apply the same pattern to integrations and personalization involving contracts, invoices, purchase orders, suppliers, and renewals.

Drawing on Paul Iusztin's Decoding ML writing, they distinguish short-term, long-term semantic, long-term procedural, and long-term episodic memory, using a gaming simulation architecture to show how high personalisation puts pressure on short-term memory.

• They note that session context, colleague procedures, and agent logic are often confidential or proprietary and should be protected.
• Short-term memory, they argue, is about personalisation as much as time.

Applies instructional-design concepts to AI, distinguishing knowledge to recognise from procedures to carry out, then mapping that distinction to how models store and recall information.

For brands planning a campaign

Useful for teams building memory, personalisation, chat, or retrieval features that need a plain-language account of the vocabulary and tradeoffs. A campaign could show how a product separates session, semantic, procedural, and episodic memory, and how each is protected.

Examples: Source 1 · Source 2

They suggest reordering those items chronologically to form a value stream for a new feature or major release.

• Another approach uses an observability platform: associate the items with services and resources, trace varied prompts across stores and tools, and open the conversation using spans and heat maps.
• The post recommends Wardley mapping to help product and systems colleagues agree on what to build, buy, or partner.

Turns a third-party diagram into a checklist, tracing exercise, and workshop agenda.

The teaching anchor is someone else's diagram; the value lies in the application exercise.

For brands planning a campaign

Fits observability, platform, and infrastructure vendors whose buyers need a way to bring product colleagues into systems conversations. Sponsorship could take the form of a walkthrough that turns the vendor's tracing data into a value stream for a named feature.

Examples: Source 1

Cost efficiency and unit economics over time is the example: it is a familiar metric with an easier investment case, but answering it requires contracts, invoices, and purchase orders, shifting the work away from public internet access and into collaboration with IT and information security.

• The post recommends small language models for structured work across a limited number of repositories, scoping all data stores up front to avoid rework, and using top-of-mind tasks such as tracking big spenders or time-sensitive renewals to bring people and models together.
• The resulting context and fine-tuned models can then carry into later initiatives.

For brands planning a campaign

Directly relevant to procurement software, data integration, and private-deployment vendors. A campaign could ask Miles to apply this starting-question method to the vendor's buyer set or write a first-question playbook for the category.

Examples: Source 1

They highlight how separating essential from non-essential attributes of success helps a model prioritise and judge how far to reason.

• Checklists, they argue, are the clearest format for that guidance because they define a finished piece of work.
• The post also argues that expert guidance is another name for context, and teams should keep owning and articulating their business context rather than letting AI manage it.

Compares human assessment practice with reinforcement-learning inputs, arguing that the same drafting skills transfer between the two.

For brands planning a campaign

Speaks to evaluation, model-training, and learning-platform buyers. A partnership could pair the vendor's rubric or evaluation tooling with Miles' framing of checklists as a usable interface for expert input.

Examples: Source 1

The procurement post links starting with cost efficiency to reusable context and fine-tuned models; the memory post notes that customer and colleague context is confidential or proprietary; the diagram post calls for shared understanding across product and systems teams; and the rubric post frames expert guidance as context that teams should not let AI manage for them.

For brands planning a campaign

This argument could anchor a campaign for vendors selling data governance, private deployment, enterprise search, or security. Because the claim is close to a slogan, tie it to a specific decision—such as where a model runs or which data store it may reach—rather than restating it.

Examples: Source 1 · Source 2 · Source 3 · Source 4

Ask whether Miles has a sponsorship, client, employment, or partnership relationship with Scale AI.

View example

Ask whether Miles has a commercial relationship with Iusztin or Decoding ML.

View example

Paul's breakdown of agentic memory into working, episodic, semantic, procedural, entity, taxonomic, long-term and shared layers covers much of the ground Miles covers with short-term, semantic, procedural and episodic memory. Both make the point that most agent frameworks ship only a fraction of it, and everything breaks once the context window fills.

Paul writes as a builder organising a product schema: the memory post ends by mapping his company's products to each layer, and numbers the layers as a development checklist.

• Miles writes as an instructional designer, turning someone else's diagram or paper into an exercise a team can run.
• Paul's posts are narrower and less consistent in kind, including a career update from enterprise software sales rather than AI commentary.

Read original · Original 2

The two make sensible companion reading for someone trying to understand why context, not the model, is the hard part.

Daali writes original long-form essays, using her own examples to argue for a research direction — world models, joint embedding architectures, embodied intelligence — and is building a subscriber community around it. Miles writes much shorter posts anchored to another person's diagram or paper, then turns the material into a team exercise rather than making a research claim of their own.

Read original · Original 2

That overlaps with Miles's MCP, orchestration-and-infrastructure and procurement material.

Bhattacharya writes as a founder reporting on his own company's outbound and GTM workflows, and points readers to a whitepaper.

• The frame is operational infrastructure and pipeline.
• Miles approaches systems, security and procurement through an instructional-design lens, arguing that context is capital and the team should keep owning it.

Read original · Original 2

In his post on agentic memory, he breaks the problem into eight layers, from working and episodic memory to entity and shared memory. The framework helps a technical audience see why treating memory as one component can leave an agent forgetting or confusing context—and gives a memory-platform company a clear way to explain its role.

This is an implementation-focused taxonomy, not a guide to cross-functional decisions such as choosing an initial use case, coordinating access to sensitive business data, or involving security and IT. The other post is a career update, so these posts do not say much about a broader campaign.

Examples: Source 1 · Source 2

Her article on agents explains agents through familiar ideas—planning, tools, memory, and feedback—and frames them as systems that can act toward goals.

• Her world-models article contrasts predicting words with learning from a changing environment.
• These pieces suit an educational campaign aimed at readers who need the concepts before they can weigh architecture choices.

The work is conceptual and introductory. Product teams making implementation decisions should ask for operational guidance on enterprise data stores, deployment boundaries, and access controls.

Examples: Source 1 · Source 2

In his post about the AI-driven outbound arms race, he argues that automating execution does not guarantee better pipeline: trust, positioning, and distribution still matter.

• His digital-worker post makes the operational case through a concrete coordination problem—people manually coordinating disconnected systems—and identifies orchestration, permissions, reliability, and oversight as deployment concerns.
• That perspective is a closer fit for revenue and operations teams than a campaign focused on memory or personalization in AI products.

The examples focus on outbound and go-to-market processes, not context engineering, memory design, or handling confidential data in personalized products.

• The digital-worker post also points readers to a whitepaper, so the campaign would need to account for that promotional element.
• Before booking, ask for audience composition and commercial terms.

Examples: Source 1 · Source 2