“If Kathy calls, tell her I’ll be there at two.”
Of everything in Apple’s 1987 Knowledge Navigator video - the tablet, the video calls, the bow-tied agent - that throwaway line is the one nobody has delivered. Siri AI and ChatGPT both understand the sentence perfectly. Neither can do it properly.
This post started with a question. SmartFriend™ Peter Marks asked it back in January on marxy.org: Apple’s Knowledge Navigator - are we there yet? His verdict was close, with long-term memory and autonomous multi-step tasks still missing. “Not quite there, but it is within reach.”
To see how far ahead of its time this was, look at what Apple was actually selling. A well-equipped Macintosh SE, launched in March 1987, had a Motorola 68000 processor running at about 8 MHz, 1 MB of RAM and a 20 MB hard disk. Its built-in nine-inch (23 cm) monochrome screen managed 512 × 342 pixels. Plenty of Mac owners were still living on floppy disks, and MultiFinder, the first taste of multitasking on the Mac, could juggle only two applications at once.
Against that backdrop, Apple imagined a portable, networked computer with a large touchscreen, video conferencing, natural spoken conversation, instant access to academic research and an agent that managed communications and collaborated on scientific simulations. The gap between what Apple could build and what it imagined was breathtaking.
Now flip the telescope. The iPhone 18 Pro has 12 GB of RAM and starts at 256 GB of storage. That is roughly 12,000 times the memory and 12,800 times the storage of that SE. Order the 2 TB model and the storage multiple passes 100,000. The hardware gap Apple dreamed across has closed. What’s left is the software.
Nine months after Peter’s question, Apple’s rebuilt Siri AI has shipped in beta and OpenAI has launched Dots. I wanted a number, not a feeling. So I went back to the film.
▶ Watch on YouTube: Apple’s Knowledge Navigator (1987), high resolution
What Apple imagined in 1987
A university professor sits at a folding tablet and simply talks. His agent summarises his messages, remembers his father’s birthday, finds a paper by a researcher whose name he half recalls and compares that researcher’s predictions with what actually happened. It rings a colleague, leaves a message, then interrupts to say she’s calling back and patches her through on video. The two of them merge their data and run a deforestation simulation together.
Meanwhile the agent keeps working. It reminds him about the birthday cake, prints an article and takes instructions for the calls he’ll miss while he’s at lunch.
No apps. No menus. No command syntax. The professor says what he wants and the computer takes responsibility for it. That was the radical idea, and it’s why Ross Rubin’s 2023 Fast Company piece rated the hardware as surpassed while judging the experience as “we’re getting there”. ChatGPT-style tools, he noted, match parts of the functionality but not the anticipation of user needs.
How I scored it
I broke the film into 35 distinct interactions between the professor and his agent, grouped them into five categories and scored each from zero to five. Then I weighted the categories by how much of the film they occupy.
My first pass compared Siri AI on an iPhone (iOS 27) with ChatGPT (Voice, Work and connected apps). Then I realised I had set Apple a harder test than the film did. The Knowledge Navigator wasn’t a smartphone assistant. It was a complete personal computing environment: a research workstation, a communications device and a conversational agent in one. So I added a third column, Siri AI on macOS 27 working with an iPhone through Apple’s Continuity features.
| Capability | Weight | Siri AI on iPhone | Siri AI on Mac + iPhone | ChatGPT |
|---|---|---|---|---|
| Personal briefing and scheduling | 20% | 84% | 86% | 88% |
| Research and literature analysis | 25% | 66% | 75% | 71% |
| Geographic research and data retrieval | 15% | 40% | 58% | 53% |
| Live collaboration and simulation | 25% | 53% | 67% | 36% |
| Background assistance and delegation | 15% | 56% | 64% | 29% |
| Overall | 100% | 61% | 71% | 57% |
On an iPhone, Siri edges ChatGPT 61% to 57%. Add a Mac and it jumps to 71%. These are capability estimates, not test results, so don’t read much into small margins. The 71% assumes a configured Mac and iPhone with supported apps and Shortcuts. Insist that every task be done with one spoken instruction and no manual help, and I’d put it nearer 65%. The pattern underneath is the interesting part.
ChatGPT is the better researcher on a phone
The research sequence is where language models shine. The professor remembers a paper by “Dr Flemson”, and the agent works out he means John Fleming, finds the publication, pulls out his numerical predictions and checks them against later observations. No search syntax required. Fuzzy recollection in, answer out.
ChatGPT leads here on extracting figures, summarising arguments and doing the arithmetic. Siri has improved enormously at understanding documents and on-screen content, but ChatGPT has the richer environment for analysis.
Both stumble on the same two things on a phone. Finding papers the professor hasn’t read needs a complete reading history, which neither assistant has. And academic access is fragmented behind paywalls and repositories that general assistants can’t reach.
The map sequence is worse for both. Filtering a university network, selecting Brazil and pulling 30 years of monthly observations is less an AI problem than a plumbing problem. It needs the right geospatial data and an application that exposes it. The film assumes everything is already connected.
Siri wins because Apple owns the phone
Then Jill calls back.
“Contact Jill” is a one-word instruction on an iPhone. Siri can place the call, announce the return call, answer it and connect the video. ChatGPT can draft Jill a message and not much else, because telephony simply isn’t part of its world. That single architectural fact drives most of Siri’s lead in the collaboration and delegation categories.
Apple also has a structural edge in how it’s building the assistant. App Intents lets developers describe their data and actions in a form Siri can discover, including tying what’s on screen to a real object so “send this to Jill” means something. Apple’s own launch example is a good illustration: Siri finds a message where someone mentions baking a family recipe, pulls the instructions from an email that relative sent and adds the ingredients to Reminders. That’s cross-application reasoning of exactly the kind the professor takes for granted.
I wrote about this architecture when Apple put the user at the centre of its AI. The assistant starts from your personal context rather than treating each request as a cold question.
The limits are real too. Siri AI is a beta, and it can only drive what third-party developers choose to expose. ChatGPT is more flexible across platforms but pays for it in permissions, interfaces and formats that differ app by app. Neither has solved universal application control.
Add a Mac and Siri jumps to 71%
Everything above treats Siri as a phone assistant. Give it a desktop and the picture changes.
Siri AI runs on macOS 27 (any Mac with an M1 or later, in English and still in beta). It sits inside Spotlight, and Apple’s own Mac documentation reads like a Knowledge Navigator script. Select two documents in Finder, ask Siri “Which quarterly report grew faster?” and it compares them. Open a diagram in Notes and it will explain it. That is the professor retrieving and interpreting the literature.
It also fixes the film’s habit of saying “And what happened?” or “Let’s put these together” with no further explanation. On the Mac you can select a window or an area of the screen and ask about what’s in it. The limit is that understanding a graph doesn’t mean Siri can change the spreadsheet or scientific model behind it. The application has to expose that.
Then Continuity does the telephone work. A Mac signed in to the same Apple Account as your iPhone can make and receive iPhone calls through the Phone app or FaceTime, and Apple lists “Call Ashley Rico” as a Siri command on the Mac. Replay the Jill sequence. The professor asks Siri to call Jill. She rings back on his iPhone number, he takes it on the Mac and shares his screen over FaceTime with his research still open. That is a close copy of the film’s communications.
The Mac also brings the desktop tools that Siri can orchestrate. Siri can run any shortcut by name and Apple says you can now create complex automations by describing them. “Copy the last 30 years at this location at one month intervals” becomes a workflow, provided there is a dataset and an application to run it against. Siri orchestrates an existing capability. It doesn’t invent one.
Two things I didn’t credit. iPhone Mirroring looks tempting, but Apple says the camera, microphone and calls aren’t available through it, so it isn’t what turns Siri into the Knowledge Navigator. And Siri on the Mac can hand a question to ChatGPT, such as summarising an open PDF. That’s a promising pairing, but it gives you ChatGPT’s answers, not its whole Work environment.
The biggest jumps are live collaboration (53% to 67%), geographic data (40% to 58%) and research (66% to 75%). Look at why. None of it needs a smarter language model. It comes from giving the assistant a more capable computer to work with.
The Kathy problem
Back to the line that opened this post.
The words are easy. Fulfilling them is not. The assistant has to remember a future obligation, recognise the moment Kathy rings, work out who she is, deliver the message and do all of it after the professor has walked out the door.
That’s the gap. Siri scores 2 out of 5 on it and ChatGPT 1, and a Mac doesn’t move the number. It can receive Kathy’s call and Siri can send her a message, but nothing intercepts her future call, remembers a delegated instruction and speaks for the professor. Continuity solves connectivity. It doesn’t solve autonomy.
Apple does have Call Screening and Live Voicemail, but they’re separate features, not one assistant taking responsibility for your calls. A chatbot, by design, waits to be asked.
This is why the autonomous agents matter more than another jump in benchmark scores. OpenAI’s Dots is the closest thing yet. A Dot has its own cloud computer, memory, the apps you connect and the ability to keep working between conversations, with rules for when it acts, asks or hands back to you. It’s rolling out to Pro users outside the EEA, Switzerland and the UK, and to Business Premium. Gemini scheduled tasks and Microsoft’s autonomous agents in Copilot Studio head the same way.
Note the qualification. Always-on doesn’t mean always listening. These agents run on a schedule or wake on an event. That’s enough for plenty of useful work, and still a long way short of the film’s bowtied assistant.
What’s left
Six gaps stand between us and the professor’s desk: an assistant that holds ongoing responsibilities rather than just past conversations; dependable access across applications; communications handled as one system; trustworthy scientific computing; a shared workspace where two people and an AI manipulate data together; and the judgement to know when to interrupt. That last one is subtler than it sounds. The agent breaks in when Jill rings back because it knows the call matters to what he’s doing right now.
The 29% that the best Apple configuration still misses is revealing. Most of it isn’t about understanding more sophisticated language. It’s about operating applications reliably, handling communications autonomously and taking ongoing responsibility for delegated work. Apple has spent years assembling the parts - Continuity, Handoff, FaceTime, iCloud, Spotlight, Shortcuts and App Intents - and Siri AI is the first conversational layer that ties them together.
My prediction, and it is a prediction: the next stretch of progress comes less from smarter models and more from reliable app access, memory that carries across days and carefully bounded autonomy. The best result probably combines both camps, a strong reasoning agent plugged into an operating system that exposes everything.
A word on how solid all this is. I identified the 35 interactions from the film and its transcript, used AI assistance for the analysis and checked the capabilities against product documentation. I did not run 35 tests. Siri is judged on an iPhone and then on a Mac and iPhone together, while ChatGPT gets credit for its wider ecosystem. I used ChatGPT 6 for the analysis so any bias should be in its favour. Running a representative subset of the tasks head to head would replace some of my judgement with evidence, and that’s the obvious next step, but it’s a big job - beyond the energy of this blogger.
In 1987 Apple imagined a computer that would understand what you wanted and get on with it. The understanding has arrived. The getting on with it is still waiting for Kathy to call.
Sources:
- Apple’s Knowledge Navigator - are we there yet? - Peter Marks, marxy.org
- Apple’s 1987 Knowledge Navigator video depicted a future that’s still a work in progress - Ross Rubin, Fast Company
- Siri AI, a profoundly more capable and personal assistant, is here - Apple Newsroom
- Apple Intelligence and Siri AI integration - Apple Developer
- Use Siri AI with on-screen awareness on Mac - Apple Support
- Make and receive phone calls on your Mac - Apple Support
- Getting started with your Dot - OpenAI Help Center