Amy Daali is most useful for accessible explanations of AI agents, world models, and human-centered AI in manufacturing.
Best for
A robotics campaign on embodied agents with persistent memory and a professional audience.
Not for
A consumer wellness or electronics launch needing broad reach, short-form video, and paid sponsorships.
Key takeaways
AI agents are digital employees, not advanced tools.
Language models predict words; they do not remember, plan, or understand the physical world.
Manufacturing AI should elevate operators and engineers, not replace them.
Check first: Confirm partnership intake, rates, and recent LinkedIn engagement on comparable AI architecture posts.
Main ideas
AI agents are digital employees, not advanced tools · Language models predict words; they do not remember, plan, or understand the physical world · Manufacturing AI should elevate operators and engineers, not replace them
Amy Daali writes about artificial intelligence, especially AI agents, world models and human-centered AI in manufacturing.
Their work connects technical ideas to practical questions: what agents can do, where language models fall short, and how AI might support rather than replace people in factories.
In their article on world models, they argue that persistent memory, physical experience and planning matter for AI that must operate in changing environments.
A LinkedIn post applies a similar human-centered lens to manufacturing, emphasizing the value of operators’ and engineers’ expertise.
Readers interested in accessible explanations of these ideas and their possible applications can start with their AI Minds articles.
Daali’s distinctive contribution in these articles is connecting big AI concepts to concrete situations.
Their agent article explains the difference between chatbots, copilots and agents through examples such as planning an event or remembering grocery preferences.
Their world-model article shifts from everyday tasks to robotics and other settings where memory, physical grounding and long-term planning matter; they suggest that early deployment is more plausible in bounded environments such as labs, factories and hospitals than as general-purpose autonomy.
The manufacturing post adds an important qualification: AI should elevate operators and engineers, not displace their expertise.
Together, these pieces give readers an accessible introduction to agent concepts and a clear account of their stated priorities.
The articles synthesize ideas and examples, including work associated with Yann LeCun.
Amy Daali argues that AI agents should be treated as digital employees rather than software tools, and that today's language models, which predict words rather than understand the world, will not get there on their own.
Read more
Their two AI Minds Substack articles set out that vocabulary and its limits, while a LinkedIn post applies the same stance to manufacturing, where they argue AI should elevate operators and engineers instead of replacing them. Their teaching style is consistent: explaining AI through comparisons with human roles, learning, and development.
AI agents are digital employees, not advanced tools
Daali rejects the common view of AI agents as software that answers queries or schedules a meeting, and proposes treating them as autonomous employees.
Read more
The argument rests on a three-way distinction.
Chatbots hold conversations and retrieve information with limited decision-making.
Copilots suggest and collaborate but do not act independently.
Agents plan, execute, adapt, decide, and reflect with minimal supervision.
Daali identifies four components that make an agent work: tool use, planning, reflection, and memory.
Memory covers both short-term handling of immediate tasks and long-term storage that an agent draws on later.
Agents also operate in networks, where specialization and delegation, coordinated problem-solving on a shared problem, autonomous decisions without human input, and shared feedback loops make the group more capable than a single agent.
The priority is management rather than purchase: agents need goals, frameworks, and feedback.
The open questions are what work to delegate, how to handle safety and ethics, and how to measure agent performance.
A vacation assistant pulls flight options from Google Flights and lodging from Airbnb in one request; an event agent sequences venue booking, invitations, catering, and reminders around deadlines; a writing assistant reviews and edits its own draft; a shopping assistant remembers past orders and anticipates forgotten items. In healthcare, Daali describes one agent scanning records, another reading lab results, and a third recommending treatment, so the group produces a fuller diagnosis than any one could.
Read alongside the world-model article, this piece sets a boundary on its own claims.
The employee framing assumes agents already hold memory and plan over time, which the other article describes as unsolved in current systems.
Daali does not withdraw the analogy, but the two pieces together frame memory and planning as claims to prove rather than settled features.
Language models predict words; they do not remember, plan, or understand the physical world
Daali's second article argues that large language models are built to predict tokens, not to reason, remember, or interact with the real world.
Read more
It identifies four specific failures of current agent systems: memory that does not survive across sessions and tasks, poor generalization when the environment changes, loss of coherence over long sequences, and no sense of space, movement, or consequence.
Daali concludes that more data, tokens, and GPUs cannot close the gap because it is architectural.
They favor world modeling, embodiment, and multimodal interaction: systems that estimate missing information about the state of the world and predict plausible future states.
They point to Yann LeCun's Joint Embedding Predictive Architecture (JEPA), which predicts abstract representations of future states in a compressed latent space rather than raw pixels or tokens, combining vision, action, and other modalities in one shared internal space.
Daali also cites LeCun's Advanced Machine Intelligence framing and developmental timeline.
Their public timeline is that fully embodied intelligent agents within five years is overly ambitious, while ten years is plausible for scaled, domain-specific deployment, arriving first in bounded environments where the range of inputs and actions is limited.
The central comparison is quantitative.
An LLM trained on twenty to thirty trillion tokens, roughly one hundred trillion bytes, has processed about the same raw volume as a four-year-old's visual experience of the physical world.
Daali argues the child's learning is richer because it is embodied and reinforced by consequence.
They expect world models to matter first in scientific discovery and medical research, with agents simulating molecules and proteins or modeling disease progression and treatment outcomes; the first places for deployment are labs, factories, and hospitals.
Attribution matters here.
JEPA, the term Advanced Machine Intelligence, and the infant concept-acquisition timeline belong to Yann LeCun's work, which Daali cites and explains rather than originates; the claim that the child's learning is richer is their interpretation.
The article is an explainer that assembles other researchers' arguments into a case for embodied AI.
Manufacturing AI should elevate operators and engineers, not replace them
In a short LinkedIn post, Daali frames the current moment as a question of human purpose: a century was spent teaching humans to think like machines—precise, repeatable, and efficient—and now the machines have learned.
Read more
They argue that the future factory is not about replacing operators and engineers but elevating them, and that the conversation starts with human-centered AI in manufacturing. The priority is deployment that does not erode trust or expertise, aimed at people leading in engineering, operations, or manufacturing and the advisors who work with them.
The post promotes a talk by Sangati Seth on Wednesday, February 25 at 2 PM CT about where AI is creating real value in manufacturing today and how to deploy it without eroding trust or expertise, with a registration link to a Luma event page. The framing question about what humans are for, and the claim that AI should elevate rather than replace operators, are Daali's own words; the talk's content and the point about deploying without eroding trust are credited to Sangati Seth as the guest speaker.
This is the clearest statement of the vocabulary Amy Daali uses elsewhere: what separates an agent from a chatbot or copilot, and why they argue agents should be treated as employees rather than software tools.
Read more
Newcomers should start here; the definitions set up the arguments in later pieces.
Daali opens their Substack by rejecting the common view of AI agents as advanced tools and proposing they be treated as digital employees.
They distinguish three categories: chatbots handle conversation and information retrieval with limited decision-making; copilots suggest and collaborate without acting independently; agents plan, execute, adapt and decide on their own.
They identify four components of an agent—tool use, planning, reflection and memory—and illustrate each with plain examples: a vacation planner that draws on flight and lodging tools, an event organizer that sequences venue, invitations and catering, a writing assistant that reviews and edits its own draft, and a shopping assistant that remembers previous orders and anticipates needs.
Daali then describes agents working in networks, with task specialization, coordinated problem-solving, autonomous decision-making and feedback loops, using hospital diagnosis and logistics routing as examples.
The piece closes with open questions about managing an AI workforce, what work to delegate, safety and ethics, and how to measure agent performance, plus an invitation to join their community.
For brands planning a campaign
A team selling persistent-memory agents can use this post’s framing, but should address its gap: memory appears as one of four components, illustrated by a shopping assistant that remembers preferences, with no discussion of session limits or architecture. Show how the product delivers the continuity Daali describes rather than assuming readers already believe agents have it.
This is the original piece that speaks most directly to whether current AI can retain memory and act in the physical world.
Read more
It explains what Daali thinks language models cannot do, what should replace them and how long they expect the shift to take.
Daali argues that large language models are built to predict tokens, not to reason, remember or interact with the real world, and lists four failures of current agent systems: persistent memory that does not survive across sessions, poor generalization when the environment changes, loss of coherence in long sequences, and no physical grounding or sense of consequence.
They compare the scale of language training—roughly one hundred trillion bytes at twenty to thirty trillion tokens—with the visual data a four-year-old has processed, concluding that the volume is similar but the child’s learning is richer because it is embodied and reinforced by feedback.
Daali points to Yann LeCun’s Joint Embedding Predictive Architecture and the Advanced Machine Intelligence framing as the direction they favor: systems that predict abstract representations of future states in a latent space while using vision, action and other modalities together.
They expect early applications in scientific discovery and medical research, with molecules, proteins and treatment planning as examples.
They say fully embodied intelligent agents within five years is overly ambitious, while ten years is plausible for scaled, domain-specific deployment, most likely first in bounded settings such as labs, factories and hospitals.
The post ends with a question to readers and a pointer to their community.
For brands planning a campaign
For a robotics product campaign, this is the decisive piece.
Daali’s timeline—ten years for scaled deployment, with earlier progress in bounded environments—argues for pitching a product as domain-specific and suited to one structured setting, not as general-purpose autonomy.
Their view that intelligence is grounded in experience, not just data volume, gives vendors an argument for embodied hardware plus memory over a larger language model alone.
This short post shows how Daali uses their feed beyond their own publications: they amplify a practitioner talk and take a position on what factory AI should and should not replace.
Read more
It explains what “human-centered AI in manufacturing” means in their own words.
Daali writes that a century was spent teaching humans to think like machines—precisely, repeatedly and efficiently—and that now machines have learned, leaving the question of what humans are for as the most interesting one in manufacturing AI.
They say the future factory is about elevating operators and engineers, not replacing them, and position human-centered AI in manufacturing as the starting point for that conversation.
Most of the post promotes a talk by Sangati Seth on Wednesday, February 25 at 2 PM CT about where AI creates real value in manufacturing and how to deploy it without eroding trust or expertise.
Daali names the intended audience: people leading in engineering, operations or manufacturing, and advisors to those areas.
The post includes a registration link to a Luma event page.
For brands planning a campaign
Any pitch that suggests replacing plant operators or engineers runs against Daali’s stated position.
A manufacturing or robotics campaign aimed at them should lead with augmentation, preserved operator expertise and trust, and be ready to explain how trust is maintained as autonomy increases.
The post also shows Daali will feature a sponsor or event when it fits their theme, making a practitioner session a more realistic channel for a vendor than a product advertisement.
Recurring themes and development across the corpus3
AI agents as digital employees1
The limits of LLMs and the case for world models1
Human-centered AI in manufacturing1
AI agents as digital employees
In 'AI Agents: More Than Just Tools', Daali argues that AI agents are more than advanced tools: they are autonomous digital employees.
Read more
The post contrasts chatbots (reactive, with limited decision-making), copilots (suggestive and collaborative), and agents (proactive, capable of independent decisions, learning, and action).
They break down key components: tool use (for example, using Google Flights and Airbnb for a vacation), planning (organizing an event step-by-step), reflection (an AI writing assistant reviewing its own draft), and memory (a grocery assistant remembering preferences).
They also describe how agents collaborate in networks through task specialization, coordinated problem-solving, autonomous decision-making, and adaptive learning.
Definition-and-comparison explainer with concrete examples: vacation planning, event organizing, customer service, healthcare, and logistics.
For brands planning a campaign
For a robotics startup building embodied AI agents with persistent memory, this piece offers accessible framing for non-technical stakeholders.
Daali's agent-as-employee analogy can help explain delegation, governance, and performance management.
Their emphasis on memory directly supports a campaign focused on persistent memory.
In 'World Models: From Predicting Words…', Daali argues that current large language models are prediction engines, not reasoners.
Read more
They lack persistent memory, generalization to new environments, long-term planning, and physical grounding.
They cite Yann LeCun's Joint Embedding Predictive Architecture (JEPA) as a promising alternative that learns abstract representations of future states rather than raw tokens.
To illustrate the difference between data volume and embodied experience, they compare LLMs trained on 20–30 trillion tokens (roughly 10^14 bytes) with a four-year-old child who has processed a similar amount of raw visual data through interaction with the physical world.
They predict that fully embodied, intelligent agents may take 10 years to reach scaled, domain-specific deployment, starting in structured environments like labs, factories, and hospitals.
Technical explainer with an expert citation, quantitative analogy, and comparison of developmental timelines.
For brands planning a campaign
A robotics startup building embodied AI agents can use Daali's arguments to explain why world models and persistent memory matter.
The contrasts between prediction and reasoning, and tokens and experience, give campaign content a ready-made conceptual framework.
The 10-year horizon and focus on structured environments set realistic deployment expectations.
In a LinkedIn post, Daali asks about human purpose in an age of intelligent machines: 'We've spent 100 years teaching humans to think like machines: precise, repeatable, and efficient.
Read more
Now the machines have learned.
So what are humans actually for?' They promote a talk on human-centered AI in manufacturing, emphasizing the elevation of operators and engineers and the deployment of AI without eroding trust or expertise.
The post highlights where AI creates real value in manufacturing today.
A LinkedIn event promotion that shows Daali's framing and curatorial role, and highlights a talk on human-centered AI in manufacturing.
For brands planning a campaign
Brands in manufacturing or industrial AI can consider Daali as a speaker, moderator, or thought partner who connects AI capabilities to workforce concerns. The framing around augmentation, trust, and engineering leadership can support campaigns aimed at technical leaders and operators.
Recurring themes and development across the corpus
Across the three sources, Daali returns to a few core ideas: AI should augment rather than replace humans; current AI lacks embodied understanding; and new architectures (agents, world models) are needed for real-world tasks.
Read more
The two Substack posts move from an introductory explanation of AI agents (described as the first Substack post) to a technical dive into world models.
The LinkedIn post applies these themes to manufacturing.
They use analogies to human roles and learning—employees, children, and operators—to make technical concepts accessible.
Consistent use of human analogies—digital employees, child development, and operators—to explain AI concepts.
For brands planning a campaign
For a robotics startup focused on embodied AI agents and persistent memory, Daali's emphasis on agent fundamentals and advocacy for world models are directly relevant.
They can explain why memory and physical grounding matter and connect those ideas to practical deployment.
Their use of analogies suits campaigns that need to bridge technical and non-technical audiences.
Amy Daali offers a clear, analogy-driven entry point to the ideas behind agentic AI.
Read more
Two AI Minds Substack essays do the explanatory work: “AI Agents: More Than Just Tools” distinguishes agents from chatbots and copilots and identifies tool use, planning, reflection, and memory as the parts that make an agent work; “World Models: From Predicting Words” argues that predicting tokens is not the same as understanding a world and points to Yann LeCun’s JEPA as a different direction.
It lists specific failures — memory that does not survive sessions, weak generalization, lost coherence over long sequences, and no physical grounding — rather than simply asserting them.
Read the essays together: the second puts limits on the first, making an “employee” that remembers and plans a claim to prove, not a shipped feature.
A shorter LinkedIn post on human-centered AI in manufacturing shows how Daali brings the same argument to engineering and operations audiences, framing AI around elevating operators rather than replacing them.
Useful next stops are creators working on adjacent topics: Santi Adavani on digital twins and factory simulation, Manlio Carrelli on which AI companies are gaining momentum, and Terry Miles on applying language models to messy procurement data.
Both write short, scannable posts that sort a fast-moving AI field into named categories readers can act on.
Read more
Carrelli’s list of eight companies with their momentum scores and breakdown of brain-computer interface progress into surgical, manufacturing, and neural-decoding tracks use the same structure as Daali’s distinctions between chatbots, copilots, and agents and list of four agent components.
Carrelli focuses on commercial momentum and market signals — who is gaining, and what funding and customers show — while Daali explains how the underlying systems work and why current architectures fall short. Follow Carrelli for company and investment patterns, and Daali for the mechanisms.
Adavani writes for the same industrial audience as Daali’s manufacturing post.
Read more
Both explain industrial AI through numbered components — Adavani covers digital twin data sources and use cases, while Daali covers agent capabilities — and keep the writing accessible rather than academic.
Adavani stays on the factory floor: digital twins, production scheduling, bottleneck identification, and staged levels of simulation maturity. Daali moves between agent architecture and the research case for world models, treating manufacturing as one application of a broader argument about human-centered deployment.
Memory is a shared subject: Daali names it as a core agent component and a failure that keeps agents shallow, while Miles separates short-term, semantic, procedural, and episodic memory in the context-engineering work he recommends.
Read more
Both translate technical AI concepts for practitioners rather than researchers.
Miles anchors everything in procurement — contracts, invoices, purchase orders, and small language models installed off the public internet — and often points readers to others’ explanations, such as Paul Iusztin’s Decoding ML. Daali builds her own analogies, including digital employees and child development, and focuses on agents and world models.
Reach is concentrated on LinkedIn, where the profile shows 13,988 followers as of October 7, 2026.
Read more
One post addresses engineering, operations, and manufacturing leaders, along with the consultants who advise them.
The AI Minds Substack carries longer technical writing: one article distinguishes AI agents, chatbots, and copilots; another argues that language models need world models, persistent memory, and physical grounding.
For a robotics campaign on embodied agents with persistent memory, the case rests on topic match and the professional audience the posts address, not on a profiled reader base.
Before committing, request recent LinkedIn engagement on comparable AI-architecture posts (impressions, reactions, and who engaged), any Substack subscriber or open-rate figures, and, if seniority matters, an audience survey or similar documentation.
Amy Daali's posts describe no sponsorship formats, partner requirements, or published inquiry route.
Read more
Their LinkedIn post promotes a manufacturing AI talk by another speaker and provides an event registration link; it is not a sponsorship offer.
The two AI Minds articles discuss AI agents and world models and close with a Substack subscription and a community invitation, not a partnership process.
Brands considering collaboration should ask directly about partnership intake, rates, and operational requirements.
The robotics campaign is a strong match because it aligns with Daali’s argument, not just a topic label.
Read more
The campaign needs someone who can speak to building embodied AI agents with persistent memory.
In “World Models: From Predicting Words to Understanding the World,” they list persistent memory as the first of four failures in current agent systems: agents forget across sessions, tasks, and time; do not generalize when the environment changes; lose coherence over long sequences; and lack a sense of space, movement, or consequence in the real world.
They favor systems that support long-term memory, hierarchical planning, and sensorimotor learning, and highlight Yann LeCun’s Joint Embedding Predictive Architecture, which predicts abstract representations of future states in a latent space using vision, action, and other modalities together instead of raw tokens.
Their comparison of roughly one hundred trillion bytes in a 20 to 30 trillion token training run with the comparable volume a four-year-old takes in through physical interaction makes the case that more data alone does not solve physical reasoning—the objection a memory-and-embodiment company must answer for investors and technical buyers.
They also offer a realistic deployment frame: fully embodied intelligent agents within five years is overly ambitious; ten years is plausible for scaled, domain-specific deployment, first in bounded environments such as labs, factories, and hospitals.
An earlier Substack piece gives the campaign a vocabulary for non-specialists: the three-way split between chatbots, copilots, and agents, plus tool use, planning, reflection, and memory as components of an agent.
The LinkedIn post shows they argue for elevating operators and engineers rather than replacing them, keeping a manufacturing-adjacent pitch on the right side of the trust question.
Two points need attention in the plan.
The 2024 piece presents memory as one of four components, illustrated by a shopping assistant that remembers previous orders, but gives no session limits, retention windows, or architecture.
Campaign content should show how the product delivers continuity, and buyers should ask for evidence of Daali’s own client work, deployments, or technical results: the posts establish their framing and explanation, not a record of shipping embodied systems.
When to choose a specialist
A consumer campaign for a wellness app or electronics launch that needs broad mainstream reach, short-form video performance, and a documented record of paid sponsorships is a poor fit based on the available posts.
Read more
Amy Daali’s posts comprise one LinkedIn post promoting a manufacturing AI talk and two Substack essays about agent architecture and world models.
They show no short-form video, lifestyle or consumer product content, demonstration or unboxing format, or disclosed sponsorship or paid partnership.
The audience information is also limited: the LinkedIn profile shows 13,988 followers as of October 7, 2026, while no Substack subscriber figure, post-level impressions, reactions, views, or audience demographics are provided.
A professional-network follower count cannot establish mass-market reach.
Before treating a consumer campaign as viable, ask for recent post-level engagement data, Substack subscriber numbers, and any disclosed paid partnership history.
A worked campaign example
A worked angle for the robotics campaign: build on one argument Daali already makes and one question they already ask.
The campaign could be a short series titled “Memory that lasts the shift,” aimed at technical buyers and the operations leaders who sign off on pilots.
Deliverable one is a long-form explainer or recorded interview in which they walk through the four failures they name, starting with memory, and apply them to one bounded environment rather than autonomy in general.
A hospital floor or warehouse aisle fits their expectation of where early deployment will happen and brings the discussion to concrete questions: what the agent is expected to remember between shifts, what happens to that memory when the environment changes, and how a planner stays coherent over a long sequence of steps.
Deliverable two uses their four components—tool use, planning, reflection, and memory—as a scorecard buyers can apply to any vendor demo, including the client’s own.
It gives readers a practical way to test claims about continuity rather than simply accept them.
Deliverable three is an amplifier: a LinkedIn post in their established mode, pairing a sharp question with a specific invitation for the people they already address—leaders in engineering, operations, and manufacturing, and their advisors.
Across all three pieces, the series makes their argument: volume of data is not the same as grounded experience, and autonomy that cannot retain context is not autonomy a plant or a ward can rely on.
Two ground rules follow from their stated position.
Anything that reads as replacing operators or engineers works against the pitch, so augmentation, expertise, and preserved trust should lead every piece.
The timeline should stay at the stated level: scaled, domain-specific deployment in roughly ten years, starting in bounded settings.
A promise of general-purpose autonomy would contradict the posts the campaign is built on.
Before production, ask for post-level engagement on comparable architecture posts, Substack subscriber figures, and any documented partnership terms or inquiry route.
Choose Carrelli when the buyer wants to know whether the category is real, who is building in it, and how adjacent hardware is converging—not how a memory architecture should work.
Read more
One post reads eight AI companies through CB Insights' Mosaic success score, giving the buyer a usable language for early signals: developer adoption, named design partners, profitability reached by Series A.
It applies those signals to companies such as Firecrawl, Sola, Samaya AI and PlayerZero.
The other examines a robotics startup's positioning against neural interfaces.
It treats brain-computer interfaces as having hit an inflection point, with funding tripled to $867M, then walks through three fronts—surgical precision, microscale manufacturing, and AI neural decoding through NVIDIA and Synchron's Chiral model—alongside the data gap: most decoding algorithms run on animal data and EEG rather than implant data.
For a robotics startup positioning against neural interfaces, or an investor seeking the commercial landscape, that supply-chain view does work an architecture essay does not.
Carrelli reports and interprets rather than arguing for a design position.
Neither post addresses persistent memory, planning, or how an agent retains context across sessions.
They show that embodied and neural-interface work is funded and shipping, but not why a particular memory approach would hold up.
The material sits within a commercial-intelligence frame, making it useful for a market read rather than a technical evaluation; buyers assessing an architecture should ask for technical evidence.
Choose Adavani when the pilot environment is a factory floor and the buyer needs the operating case rather than the agent design.
Read more
One post sets out what a live replica of a production line integrates—PLC data, MES systems, IoT sensors, inventory levels, demand forecasts, equipment performance and maintenance logs—and the decisions it improves: production scheduling across product lines, bottleneck identification that spreadsheets miss, layout testing before construction, predictive maintenance, and workforce planning.
It closes with a survey figure: 86% of manufacturing executives see practical applications for digital twins and 44% are already implementing them.
The second post adds a maturity ladder, starting at L1 descriptive twins built from asset models, physics-based simulation and virtual showrooms, and frames them as the foundation for later levels.
For a campaign about grounded physical systems in a bounded setting, this is a concrete argument with a credible deployment sequence.
A digital twin describes and simulates the environment; it is not agent memory.
Neither post covers how an agent retains context between shifts, generalizes when the setting changes, or keeps a plan coherent over a long sequence—the failures a campaign about agent memory is built around.
The material is shaped around manufacturing, making it a weaker fit for a hospital or warehouse campaign.
It also leans on vendor and analyst framing; buyers seeking proof of deployment should ask for first-hand results.
Choose Miles when the central question is memory itself—how to define it, distinguish its kinds, and curate what a model can hold—and the buyer can carry a framework into their own domain.
Read more
One post defines short-term memory as a function of personalization, not only time; splits long-term memory into semantic, procedural and episodic kinds; and uses an architecture diagram from a gaming simulation to show a case where personalization demands are high and long interactions crowd short-term memory.
The other gives a method for a first AI project that transfers: pick a narrow, valuable question that will force the model to reach every data store; run toward the complexity of legacy data rather than away from it; expect small language models to be sufficient for structured stores; install the model away from public internet access; and bring IT and security in from the start.
Together, the posts give a startup explaining memory management to a technical buyer a working vocabulary.
The memory framework is adapted from another writer's Substack, which Miles credits, and applied to procurement rather than robotics.
Neither post touches embodiment, sensorimotor learning, or physical grounding, so the buyer gets transferable scaffolding without domain evidence.
The procurement examples rely on contracts, invoices and purchase orders, which may not map onto a robot in a plant or a ward; buyers should ask for examples grounded in physical systems.
Amy Daali has published three pieces: two AI Minds Substack articles, one explaining AI agents against chatbots and copilots and one arguing that language models predict words without remembering, planning, or understanding the physical world, and a short LinkedIn post on human-centered AI in manufacturing.
The two articles describe her vocabulary and her stated limits on current agent systems; the post shows the same position in front of engineering and operations audiences.
Comparisons with other creators are based on six of their posts.
Her posts carry no publication dates and no employer, client, credential, or funding details, so this page describes the ideas and priorities Daali states in their own writing, not a verified professional record.
Her posts also contain no original technical research, transcripts, or deployment results of their own, and the articles build on other researchers' work, including Yann LeCun's Joint Embedding Predictive Architecture, which Daali explains rather than originates.
