# Personal assistants 101

Canonical: https://mudpie.ai/101/personal-assistants/
Updated: 2026-09-20
Curator: Ali Abouelatta (https://mudpie.ai/authors/ali-abouelatta/)

The half-century path from personal computing and calm technology to assistants with memory, tools, initiative and the ability to act across software.

**TL;DR:** A personal agent is more than chat: it needs a model of the person, durable context, tools, judgment about interruption and a trust boundary. The old visions are useful because they expose how much of that product problem is still unsolved.

## Read the original visions

Before today's models, the best computing thinkers had already described personal media, conversational guides and technology that recedes into daily life.

### [A Personal Computer for Children of All Ages](https://mprove.de/diplom/gui/kay72.html)

Authority: [Alan Kay](https://mudpie.ai/authors/alan-kay/)

**TL;DR:** Kay's Dynabook is a personal, portable medium for learning and creation—not a smaller terminal for consuming somebody else's software.

**What you'll read:** The 1972 design vision for a device, interface and programming environment that adapts to a person's ideas and development.

### [The Computer for the 21st Century](https://graphics.stanford.edu/courses/cs428-03-spring/Papers/readings/General/Weiser_SciAm91.htm)

Authority: [Mark Weiser](https://mudpie.ai/authors/mark-weiser/)

**TL;DR:** The most profound technologies disappear into ordinary activity; an assistant succeeds when it reduces attention demands instead of becoming another destination.

**What you'll read:** Weiser's ubiquitous-computing thesis, tabs, pads, boards and the social design principle now called calm technology.

### [Steve Jobs at the 1983 Aspen Design Conference](https://putsomethingback.stevejobsarchive.com/international-design-conference-in-aspen)

Authority: [Steve Jobs](https://mudpie.ai/authors/steve-jobs/)

**TL;DR:** Jobs imagined computers becoming constant companions and intellectual guides, with design deciding whether that scale of adoption improved daily life.

**What you'll read:** The original talk and artifacts around a prescient vision of portable computers, networked communication and a future digital Aristotle.

### [The Knowledge Navigator concept](https://archive.org/details/knowledge-navigator)

Authority: [Apple](https://mudpie.ai/authors/apple/)

**TL;DR:** Apple's concept video makes the assistant legible as memory, conversation, research, communication and action woven into a person's workday.

**What you'll read:** A historical product film whose successes and awkward assumptions are still useful design prompts for voice, context and proactive assistance.

### [AI agents will change how we use computers](https://www.gatesnotes.com/AI-agents)

Authority: [Bill Gates](https://mudpie.ai/authors/bill-gates/)

**TL;DR:** Agents could learn a person's preferences and coordinate services across applications, shifting software from menus and apps to delegated intent.

**What you'll read:** A modern platform thesis for personal, work, health and education agents, alongside questions about privacy, standards and market structure.

## Design the working relationship

The interface problem is delegation: what the assistant knows, when it acts, how it explains itself and how the person corrects it.

### [On-boarding your AI intern](https://www.oneusefulthing.org/p/on-boarding-your-ai-intern)

Authority: [Ethan Mollick](https://mudpie.ai/authors/ethan-mollick/)

**TL;DR:** Treat a general AI like a talented but context-poor intern: give goals, examples, constraints and feedback instead of expecting one perfect prompt.

**What you'll read:** A practical management metaphor for delegation, iteration and calibrating work that is impressive but uneven.

### [Centaurs and cyborgs on the jagged frontier](https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the-jagged)

Authority: [Ethan Mollick](https://mudpie.ai/authors/ethan-mollick/)

**TL;DR:** People work with AI either by dividing tasks or intertwining their effort, and both approaches fail when they misjudge the model's jagged capability boundary.

**What you'll read:** Evidence and operating patterns for deciding when to delegate, verify, collaborate closely or keep the task human.

### [Agency and agents](https://www.oneusefulthing.org/p/agency-and-agents)

Authority: [Ethan Mollick](https://mudpie.ai/authors/ethan-mollick/)

**TL;DR:** Giving models tools and room to plan produces useful initiative but also longer chains of compounding error and harder supervision.

**What you'll read:** Hands-on examples of agent behavior and a clear account of the gap between impressive autonomy and dependable work.

### [Intro to Large Language Models: the LLM OS](https://www.youtube.com/watch?v=zjkBMFhNj_g)

Authority: [Andrej Karpathy](https://mudpie.ai/authors/andrej-karpathy/)

**TL;DR:** Karpathy's LLM OS frame treats the model as a CPU surrounded by memory, tools, retrieval, multimodal I/O and orchestration.

**What you'll read:** A technical but accessible architecture lecture that turns 'assistant' into concrete subsystems and security questions.

### [The coming wave](https://www.the-coming-wave.com/)

Authority: [Mustafa Suleyman](https://mudpie.ai/authors/mustafa-suleyman/)

**TL;DR:** Powerful general systems spread because they are useful, while containment, accountability and social adaptation lag behind capability.

**What you'll read:** Suleyman's case for the technological wave, the containment problem and why personal agents cannot be designed as a capability race alone.

## Study today's product surfaces

These are the current attempts at memory, scheduled initiative, browsing and computer use. Read them as product contracts, not promises of general autonomy.

### [Memory and new controls for ChatGPT](https://openai.com/index/memory-and-new-controls-for-chatgpt/)

Authority: [OpenAI](https://mudpie.ai/authors/openai/)

**TL;DR:** Useful personalization requires persistent context plus controls to inspect, remove, disable and separate what the assistant remembers.

**What you'll read:** OpenAI's memory model, saved memories, chat-history reference and the product controls that make persistence understandable.

### [Scheduled tasks in ChatGPT](https://help.openai.com/en/articles/10291617-scheduled-tasks-in-chatgpt)

Authority: [OpenAI](https://mudpie.ai/authors/openai/)

**TL;DR:** A personal assistant becomes proactive when work can run later, but recurring execution needs visible schedules, notifications and easy cancellation.

**What you'll read:** The current task surface, supported schedules, plan limits, management controls and the boundary between a reminder and an autonomous workflow.

### [Introducing Operator](https://openai.com/index/introducing-operator/)

Authority: [OpenAI](https://mudpie.ai/authors/openai/)

**TL;DR:** Computer use lets an assistant act through existing websites, but confirmations, takeover and task boundaries remain central to safe completion.

**What you'll read:** OpenAI's original browser-agent product, the Computer-Using Agent model, early task classes and the safety design around consequential actions.

### [Project Mariner](https://deepmind.google/models/project-mariner/)

Authority: [Google](https://mudpie.ai/authors/google/)

**TL;DR:** Google's browser agent research explores planning and acting across web interfaces while keeping a person able to observe and redirect the work.

**What you'll read:** A primary description of the research prototype, multimodal browser understanding, parallel tasks and interaction-safety constraints.

### [Apple Intelligence](https://www.apple.com/apple-intelligence/)

Authority: [Apple](https://mudpie.ai/authors/apple/)

**TL;DR:** Apple's assistant strategy combines personal context, on-device processing, private cloud compute and cross-app actions under an explicit privacy story.

**What you'll read:** The shipped product framing for writing, Siri, visual intelligence, app intents and the system architecture used to protect personal data.