On this episode of Learning from Machine Learning, Dan Bricklin, co-creator of VisiCalc, the first electronic spreadsheet and the first killer app of the PC era, shares five decades of perspective on computing revolutions, from mainframes to AI. His framework for evaluating transformative technology remains remarkably relevant today. His most important insight: we never fully know the impact of what we build, so enjoy the journey, and build with care and intention, because the moments that matter most are rarely the ones you plan for.
Takeaways
The 100x rule for killer apps: Transformative technology must be at least 100 times better than existing solutions, not incrementally better
Two-week payback principle: Technologies that pay for themselves in two weeks or less drive rapid adoption
We accept worse quality for new capabilities: People tolerated fuzzy fax quality and “can you hear me now?” cell phones because the new capability was worth the trade-off
Fertile platforms enable innovation: Once hardware is ubiquitous, you don’t need killer apps—just useful apps building on established infrastructure
Enjoy the journey: The high points in your career may not be the ones you expect; be present for them when they happen
Summary
Technology adoption follows predictable patterns across eras. Dan’s career spans every major computing platform shift, from mainframes requiring special rooms with raised floors and air conditioning, to personal computers small enough to fit in your pocket. Through this journey, he’s identified consistent patterns in what makes technology transformative versus what becomes forgotten.
VisiCalc succeeded by beating paper and calculator on the first use. The spreadsheet had to be faster than manual calculation immediately, not after a learning curve. Dan optimized for minimum keystrokes because he came from typesetting where you’re “paid by the keystroke.” This obsessive focus on speed meant the offering could be competitive against pen and paper from day one.
Memory constraints drove brilliant design decisions. Working within 32KB of total system memory (including the operating system and screen buffer), Dan and Bob Frankston had to make fundamental choices: How many significant digits? They imagined calculating the U.S. federal budget in dollars, and sure enough, VisiCalc was later used for exactly that purpose.
The best interface is still being discovered for each new capability. Just as the mouse proved vastly superior to arrow keys for precise positioning, and the laser printer revolutionized document quality, we’re still learning what interfaces work best for AI. Dan notes that even established technologies continue evolving - we moved away from pens for touch screens, then added the Apple Pencil back when certain use cases demanded it.
Understanding limitations matters as much as understanding capabilities. With AI, Dan observes the same pattern he saw with early computing: people either dismiss it entirely (”computers can do anything already”) or overestimate it (”it must be all-knowing because it’s trained on the web”). The reality requires nuanced understanding of what specific applications benefit from current capabilities.
Why it matters
Whether you’re building AI applications, evaluating new technologies, or simply trying to understand where we are in the current hype cycle, Dan’s framework offers a grounded perspective. His questions cut through both hype and dismissiveness to reveal genuine transformative potential:
What is this 100x better at?
What does that enable that was previously impossible?
What are we willing to trade off to get it?
Career Journey and Background
Early Computing Access (1960s-1970s)
Dan’s path into computing began unusually early. At 15, his cousin brought home a manual for QuickTran (a version of Fortran), and Dan taught himself programming:
“One of the first programs that I wrote [involved what] we had learned in school patterns of English... So I put in all these types of sentences and I put in a list of nouns and verbs and it would make up sentences that were technically correct.”
This early experience shaped his understanding of how programs could follow patterns and rules. He worked creatively to find computing access, convincing the Philadelphia Bureau of Public Education to let him use their IBM 1401 computers as “the only one who would go down” there. A summer course at the University of Pennsylvania led to an afternoon job at Wharton Computational Services while still in high school.
Word Processing at Digital Equipment Corporation
After MIT, Dan joined Digital Equipment Corporation working on computerized typesetting and eventually their first word processor—the DECmate WPS-8:
“Document oriented is what we’re familiar with today. Ours used just one long thing of text and it broke it up into pages and cared about layout horizontally with tabs and decimal tabs that line up just right.”
This work taught him about scrolling, locked panes, and synchronized windows which were all concepts he’d later apply to VisiCalc. More importantly, it taught him to optimize for minimum keystrokes, since typesetting workers were paid per keystroke.
The VisiCalc Origin Story
At Harvard Business School, Dan found himself repeatedly recalculating spreadsheets for case studies. The process was tedious: change one assumption, recalculate everything downstream with a calculator, erase and rewrite the paper. He envisioned something better:
“Here was a thing that did things that they would have to do by hand... but it could do it so much faster and opened up all sorts of worlds.”
Working with Bob Frankston, who wrote most of the code, Dan brought VisiCalc to market in 1979 on the Apple II. Steve Jobs later said VisiCalc “propelled the Apple II to the success it achieved.”
VisiCalc: Designing the First Killer App
The Memory Challenge
The technical constraints were severe. The original Apple II had just 32KB total memory:
“That was the program, the operating system, the screen buffer of what was going to be on the screen, and your data.”
Most users had 48KB, but Dan and Bob had to make the software work within those limits. There were special memory expansion cards that gave another 16KB, and “people would spend hundreds of dollars on this card and throw away the software that came with it” just to get larger spreadsheets.
Critical Design Decisions
Dynamic screen rendering: Memory couldn’t hold the entire display. The system had to generate what you see as you scrolled, fast enough to keep up with the repeat key.
Fixed-size storage: Everything had to be “quickly garbage collected using fixed size” blocks to ensure spreadsheets could be saved and reloaded reliably.
Maximum number size: They used the number of significant digits needed to calculate the U.S. federal budget in dollars before switching to scientific notation.
The Interface Innovation
VisiCalc introduced concepts now taken for granted:
Scrolling with locked titles (now called “frozen panes”)
Two synchronized windows to see different parts of the spreadsheet
Memory-efficient help system: Users typed a slash and saw available command characters
Formula bar and cell references enabling cascading calculations
The slash commands (like /IR for Insert Row) were so intuitive they persisted in Excel for years.
What They Left Out
Dan is candid about limitations they accepted:
“There were certain things we left out because they were hard to program or we weren’t willing to make certain compromises.”
These included:
Thousand separators in numbers
Variable column widths
On-screen graphics
Comprehensive help system
Lotus 1-2-3 later succeeded partly by adding these features on more powerful hardware (the IBM PC). Excel went further with graphical grids, fonts, and eventually bundling with Office.
The Framework for Transformative Technology
The 100x Better Rule
Dan’s central insight: transformative technology must be dramatically better, not incrementally better:
“At least 100 times better than what was before... Every time there is some new capability from a hardware viewpoint or hardware and software combination that is at least 100 times better than what was before.”
Examples:
Mouse vs. arrow keys for precise positioning
Laser printer vs. typewriter at 300 DPI with pictures
Cell phone enabling calls from anywhere vs. finding a payphone
Email delivering messages in seconds vs. waiting days for mail
The Trade-Off Principle
People accept significant quality downgrades when new capabilities matter more:
“We worked really hard in word processing to be letter perfect... Then what do we do? We said, oh, well, stick it in the fax machine that turns it down to 150 dots per inch... barely readable thing on paper that was kind of wet.”
Early cell phones were enormous with terrible audio quality (”Can you hear me now?”), but the portability justified the compromise. MP3s reduced audio quality compared to records, but carrying hundreds of songs in your pocket made it worthwhile.
Clay Christensen’s “The Innovator’s Dilemma” describes this pattern: disruptive technologies start worse in traditional metrics but excel in new dimensions customers value.
The Two-Week Payback Test
Killer apps demonstrate immediate, obvious value:
“It pays for itself the first time you use it in a couple weeks. No brainer. I mean, it’s obvious. None of this one year payback, two year payback.”
VisiCalc example: Time-sharing systems for financial forecasting cost $6,000/month. For that same amount, you could buy an Apple II, monitor, VisiCalc, and printer—and own it forever.
Desktop publishing example: Hiring a typesetter and printer for a newsletter was expensive and slow. A laser printer with PageMaker paid for itself on the first use.
Cell phone example: When Dan’s mother had a flat tire in the middle of nowhere, he immediately bought her a cell phone. The value was proven in one emergency.
Fertile Platforms and. Killer Apps
As platforms mature, the bar changes:
“Once you have a fertile environment, meaning it’s a thing where it’s a good place to do different types of applications... you can do things for that.”
VisiCalc needed to be a killer app because personal computers weren’t yet ubiquitous
Lotus Notes succeeded because companies already had Windows PCs on desks
Netscape didn’t need to justify hardware purchases - computers were already connected to LANs
Facebook launched to students who already had computers and internet access
With AI, we’re seeing both patterns: some applications justify new GPU infrastructure, while others build on existing cloud platforms and devices.
Applying the Framework to AI
What AI is Actually Good For
Dan approaches AI with the same pattern-recognition lens he’s applied to every computing shift:
“Whenever there’s a new capability, look at what is it good for and better than anything before by not a little bit, but by a lot, a hundred times better type of thing.”
Machine learning for computer vision proved to be better at face detection for camera autofocus which is something we now take for granted. It replaced engineers manually programming detection algorithms with systems that learned from examples.
Large language models excel at taking inputs and producing outputs in natural language, making certain tasks dramatically easier. But understanding limitations is crucial.
The Hype vs. Reality Problem
Dan identifies three types of responses to new technology He recalls the three personas when he showed people VisiCalc:
Computer people: “What’s so special here? I can write a program that does that”
Normal people: “Computers can do anything. What’s special about that?”
People who actually do the work: “Here’s my credit card. Give me that.”
With AI, similar patterns emerge:
“People look at it and they say, they don’t really understand its limitations or what’s what, but it looks like it’s real smart, like a person who can do things I can’t.”
The danger: assuming confidence equals correctness, or that training on “the whole web” means comprehensive knowledge.
Current Limitations
Training data bias: Medical literature often reflects studies on populations researchers could easily access, not representative samples. AI trained on this data inherits these limitations.
The streetlight effect: Dan shares the joke about the drunk looking for keys under the streetlight not because that’s where he lost them, but because “the light’s better.” We measure and optimize for what’s easy to measure, not necessarily what matters most.
Interface uncertainty: Just as we’re still discovering the best ways to interact with AI:
“Using the keyboard is better than reading into voice input for a spreadsheet with numbers and words and all that.”
Voice works great for some applications, fails for others requiring precision.
What We’re Still Learning
Prompt engineering skills are ephemeral: Mastering GPT-3 prompting doesn’t guarantee success with GPT-4 or future models. The landscape keeps shifting.
Error rates matter differently by application:
“If you have 0.0001 mistakes, you’re still gonna have some people who are in the wrong seats” in a stadium with thousands of people.
For getting “to the ballpark” vs. finding your specific seat, different error rates are acceptable.
Feedback mechanisms are still primitive: Dan asks: “How do I present things to you so you know what to check?” When an LLM generates 10,000 words, how do you verify accuracy efficiently?
Key Insights and Highlights
On User Interface Evolution
Dan’s career demonstrates that interface design is never settled:
Arrow keys vs. mouse: Both have roles. Spreadsheet power users often prefer keyboard navigation for speed, while the mouse excels at precise positioning.
Touch screens vs. pens: The iPad launched without stylus support, but the Apple Pencil became essential for certain applications. Different tools for different contexts.
Voice vs. keyboard: “You can tell by anybody’s texting who texts you with voice” whether they’re using voice recognition. For life-and-death medical applications, more reliable input methods may be necessary.
Physical form factors: The HP-35 calculator succeeded partly because it fit in a shirt pocket—Hewlett or Packard literally measured his pocket. Similarly, cell phones needed to fit in purses and pockets to achieve mass adoption.
On Building General-Purpose Tools
VisiCalc succeeded as a general-purpose tool rather than a specialized application:
“A word processor lets you write anything, you know, that’s a general purpose tool. Well spreadsheet was a general purpose tool for numbers like a word processor was a general purpose tool for paragraphs.”
This generality enabled uses Dan never imagined:
People used spreadsheets only for text, keeping lists without any calculations
The flexible layout (not rigid columns/rows) let users design custom forms
Within the first year, applications far exceeded what Dan had learned about in business school
The lesson: general-purpose tools enable creativity and unexpected applications.
On Technology Adoption Cycles
Dan has witnessed multiple boom-and-bust cycles:
Word processing in 1979: “People were not adopting computers at the rate we expected. Word processing was not taking off at that point.”
VisiCalc’s impact: It wasn’t until inexpensive portable computers (Osborne, Kaypro) that journalists could afford their own machines for writing. PCs became popular because “people involved in money liked the spreadsheet.”
AI winters: Multiple periods where AI funding dried up and investors avoided anything AI-related. Then breakthroughs like AlphaGo reignite interest.
The pattern: technology enables new capabilities, early adopters prove value, costs drop, mainstream adoption follows, enabling the next wave of innovation.
On Design Constraints Driving Innovation
Some of VisiCalc’s best features emerged from limitations:
Help system: No memory for comprehensive help led to showing command letters on-screen—which doubled as the actual characters users typed for commands.
Scrolling: The small screen (40 characters wide, 25 rows) forced innovation in locked panes and synchronized windows—features that remain standard today.
Speed optimization: Needing to beat paper-and-calculator forced ruthless efficiency, creating a tool professionals immediately preferred.
Constraints force creative solutions that often prove superior to unconstrained alternatives.
Broader Perspectives
The Hammer and Nail Problem
When people get a powerful new tool, they initially misapply it:
“The first things we did were look like ransom notes. These are our status reports where we could have just made it look like a typewriter, but no, we wanted to do all that.”
With laser printers and fonts, everyone made garish documents mixing typefaces wildly. Early web design suffered similar excess. Now we’ve learned restraint.
With AI: “I don’t know what it’s really good for, but I’ll use it for everything.” The experimentation phase is necessary but leads to misapplication before settling on appropriate uses.
On Hardware Enabling Software Innovation
Dan consistently emphasizes that hardware breakthroughs enable software innovation:
Bitmap displays enabling pixel-level control made graphical interfaces possible—essential for mouse-driven applications.
Neural engines in phones make on-device AI possible, enabling privacy-preserving applications and eliminating latency.
Laser printers at 300 DPI enabled desktop publishing, transforming how documents were created.
The question for AI: “What interfaces, input-output devices, do we need for the particular applications for which these things are useful for?”
On Humanoid Robots
Dan offers an interesting perspective on why humanoid form factors might succeed:
“So much of our society in the world around us is built around a person within a certain range with one or two hands... maybe those input-output devices, which is the humanoid robot, may be appropriate to fit in situations that we couldn’t automate without specialized machines.”
The insight: we’ve designed our physical world for human bodies. Rather than redesigning everything, humanoid robots can navigate and manipulate existing environments.
Advice and Life Lessons
Find the Right Partner
When asked about advice to his younger self, Dan immediately says: “Well, personal advice is yes, she’s the right one.”
He reflects that finding a life partner is critically important and something he got right—perhaps more important than any business decision.
Enjoy the Journey
Drawing from Steve Jobs’ response when Dan asked about Apple’s future direction:
“Steve said, it’s the journey... I had asked the same thing Gates at some point... that was the thing about enjoy the journey.”
Dan emphasizes being present for high points as they happen:
“When you have the really good time, enjoy that. Don’t feel bad that it could be... I’ve had some really special things happen in my life and been in unusual situations and no longer have that. So the more that hopefully I enjoy them at the time that I could enjoy them, that’s good.”
You Never Know Your Impact
One of Dan’s most moving examples:
“I hear from this mother who said, thank God, I really love your product. My daughter has cerebral palsy and now she can do her homework herself because she’s able to draw big on the homework and she doesn’t have to tell me what to do.”
He’d built a product allowing users to draw large on an iPad and have it shrink down—useful for marking up PDFs. He never imagined it enabling a child with cerebral palsy to do homework independently.
Other unexpected impacts:
People inspired by his TED talk who went on to build significant products
A couple who met because one was helping the other learn spreadsheets
Students who succeeded in careers because they learned VisiCalc
Do What You Love
“I’m still programming. I’m still writing... I’m still programming a lot by hand, you know, cause I like, it’s like people who like to paint. They’re not Picasso. They’re not Rembrandt, but they like to paint.”
You don’t have to be the best to find satisfaction in work. Like playing basketball without being LeBron, the activity itself can be rewarding.
On Paying Your Team
Dan made an unusual commitment when running companies:
“When I was paying people and sending stuff I said I want to do it that when you look back if we fail I want you not to feel bad about that you wasted your time here... I want to make sure that I paid you enough that it isn’t costing you to become poor.”
Not necessarily the best business practice, but it reflected his values about treating people well even when outcomes are uncertain.
Connecting to Today’s AI Moment
The Questions That Matter
Dan’s framework translates directly to evaluating AI applications:
What is this 100x better at? Not “what can it do?” but “what is it dramatically better at than alternatives?”
What does that capability enable? What becomes possible that was previously impossible or impractical?
What trade-offs are acceptable? What quality or accuracy issues will people tolerate for the new capability?
What’s the payback period? Does this justify itself immediately, or require multi-year ROI calculations?
Is the infrastructure ready? Are we at the killer app phase (need to justify hardware) or the fertile platform phase (infrastructure already exists)?
We’re Still in the Experimental Phase
Just as early word processor users made ransom-note documents and early spreadsheet users tried applying it to everything, we’re in the experimentation phase with AI:
“We’re learning. You know, remember with a lot of the visual stuff, you know, it uses hand motion. Well, is that really the best way or is it better to have a controller? We’ve now moved back to controllers.”
The best interfaces, applications, and use cases are still being discovered.
The Importance of Technical Understanding
Dan emphasizes learning to use tools properly:
“Those that learn to use the spreadsheet, those that learn to use the word processor, those that learn how to drive a car... Once you’ve learned how to drive a car and stuff like that, then you can do all sorts.”
With AI, this means:
Understanding what it’s actually trained on
Recognizing its limitations
Learning to verify outputs
Developing domain expertise to catch errors
“Unfortunately, they keep changing. So I become a prompt master. I’m really good at prompting for GPT-4 or 3... and then of course it’s changed.”
Stay Grounded in Reality
Dan’s perspective on the hype cycle comes from having seen many cycles:
“There are people who don’t know that, who just assume that it must be all knowing. And it’s better than me, so therefore it’s good.”
The antidote: talk to people who actually use the technology for real work. They’ll tell you what it’s genuinely good for, what its limitations are, and where it fails.
Like VisiCalc’s early users who “started shaking and said, here’s my credit card”—they understood exactly what problem was being solved and how much better the solution was.
References
VisiCalc
Companies and Products Mentioned
Digital Equipment Corporation (DEC)
DECmate Word Processor
Apple II
Lotus 1-2-3
Microsoft Excel
Google Docs
Key Figures
Bob Frankston - VisiCalc co-creator
Steve Jobs - Apple co-founder
Bill Gates - Microsoft co-founder
Clay Christensen - Author of “The Innovator’s Dilemma”
Universities
MIT
Harvard Business School
University of Pennsylvania
More from Learning from Machine Learning
Glossary of Key Terms
Killer App: An application so valuable that it justifies purchasing the hardware to run it. VisiCalc was the killer app for the Apple II.
Mainframe: Large, powerful computers requiring special rooms with raised floors, air conditioning, and restricted access. Dominant from the 1950s-1970s.
Time-sharing System: Computing systems where multiple users share access to a central computer via terminals, common before personal computers.
Bitmap Display: A screen where each pixel can be individually controlled, enabling graphical interfaces and precise positioning with a mouse.
Locked Panes/Frozen Panes: Spreadsheet feature where certain rows or columns remain visible while scrolling, invented in VisiCalc as “titles.”
Two-week Payback: Dan’s principle that transformative technologies must demonstrate value and pay for themselves within roughly two weeks of first use.
100x Better Rule: Dan’s framework that transformative technology must be at least 100 times better than existing solutions in at least one important dimension.
Fertile Platform: An established technology ecosystem where hardware and infrastructure are ubiquitous, enabling new applications without requiring users to purchase new hardware.
The Innovator’s Dilemma: Clay Christensen’s framework explaining how disruptive technologies initially appear inferior in traditional metrics but excel in new dimensions customers value.
General-Purpose Tool: Software designed for flexible, creative use rather than a single specific application. Spreadsheets and word processors are general-purpose tools.
Neural Engine: Specialized hardware in modern devices optimized for machine learning computations, enabling on-device AI processing.
Prompt Engineering: The practice of crafting inputs to large language models to elicit desired outputs.
LLM (Large Language Model): AI systems trained on vast amounts of text data, capable of generating human-like text responses.
Training Data Bias: Systematic limitations in AI systems resulting from biases, gaps, or unrepresentative samples in their training data.
The Streetlight Effect: A cognitive bias of looking for something where it’s easiest to look, rather than where it’s most likely to be.
