(2019-10-04) Nielsen Matuschak How Can We Develop Transformative Tools For Thought

Andy Matuschak and Michael Nielsen: How can we develop transformative tools for thought (thinking tools)? Part of the origin myth of modern computing is the story of a golden age in the 1960s and 1970s. In this story, visionary pioneers pursued a dream in which computers enabled powerful tools for thought, that is, tools to augment human intelligence E.g., Douglas Engelbart, Augmenting Human Intellect... retrospectively it's difficult not to be disappointed, to feel that computers have not yet been nearly as transformative as far older tools for thought, such as language and writing. We believe now is a good time to work hard on this vision again. In this essay we sketch out a set of ideas we believe can be used to help develop transformative new tools for thought

In the first part of the essay we describe an experimental prototype system that we've built, a kind of mnemonic medium intended to augment human memory. This is a snapshot of an ongoing project, detailing both encouraging progress as well as many challenges and opportunities. In the second part of the essay, we broaden the focus. We sketch several other prototype systems. And we address the question: why is it that the technology industry has made comparatively little effort developing this vision of transformative tools for thought?

we mentioned some visionaries of the past

veneration can veer into an unhealthy reverence

Yes, those pioneers did amazing things

But they also made mistakes, and were ignorant of powerful ideas that are available today.

A word on nomenclature: the term “tools for thought” rolls off neither the tongue nor the keyboard

Alan Kay, User Interface: A Personal View (1989), among other places. that a more powerful aim is to develop a new medium for thought. (Media Inventor)

Such a medium creates a powerful immersive context, a context in which the user can have new kinds of thought, thoughts that were formerly impossible for them. Speaking loosely, the range of expressive thoughts possible in such a medium is an emergent property of the elementary objects and actions in that medium. If those are well chosen, the medium expands the possible range of human thought.

With that said, the term “tools for thought” has been widely used since Iverson's 1950s and 1960s work

And so we shall use “tools for thought” as our catch all phrase

such changes have happened multiple times during human history: the development of language, of writing, and our other most powerful tools for thought. And, for better and worse, computers really have affected the thought patterns of our civilization over the past 60 years, and those changes seem like just the beginning

Part I: Memory systems

Introducing the mnemonic medium

Few subjects are more widely regarded as difficult than quantum computing and quantum mechanics

What makes these subjects difficult? In fact, individually many of the underlying ideas are not too complicated for people with a technical background. But the ideas come in an overwhelming number, a tsunami of unfamiliar concepts and notation.

As an experiment, we have developed a website, Quantum Country

Quantum Country is a prototype for a new type of mnemonic medium. Aspirationally, the mnemonic medium makes it almost effortless for users to remember what they read

What makes it plausible is that cognitive scientists know a considerable amount

Unfortunately, those steps are poorly supported by existing media.For more on this argument, see Andy Matuschak, Why books don’t work (2019). Is it possible to design a new medium which much more actively supports memorization? (2019-06-01-MatuschakWhyBooksDontWork)

Of course, on its own this wouldn't make it trivial to learn subjects such as quantum mechanics and quantum computing – learning those subjects is about much more than memory. But it would help in addressing one core difficulty: the overwhelming number of new concepts and notation.

In fact, there are many ways of redesigning the essay medium to do that

is it possible to 2x what people remember? 10x? And would that make any long-term difference to their effectiveness?

Embedded within the text of the essay are 112 questions about that text.

quizzed as they read on whether they remember the answers to those questions.

Note that this interaction occurs within the text of the essay itself

Of course, for long-term memory it's not enough for users to be tested just once on their recall. Instead, a few days after first reading the essay, the user receives an email asking them to sign into a review session

The literature on this effect is vast. A useful entrée is: Gwern Branwen, Spaced Repetition for Efficient Learning. (SRS)

The early impact of the prototype mnemonic medium

Plotted below is the demonstrated retention of answers for each user, versus the number of times each question in the mnemonic essay has been reviewed:

by the sixth review that rises to an average of 54 days of demonstrated retention. That typically takes about 95 minutes of total review time to achieve. Given that the essay takes about 4 or so hours to read, this suggests that a less than 50% overhead in time commitment can provide many months or years of retention for almost all the important details in the essay.

Six months after release of the prototype, 195 users had demonstrated one full month of retention on at least 80% of cards in the essay

the mnemonic medium has many surprising properties. It turns out that flashcards are dramatically under-appreciated, and it's possible to go much, much further in developing the mnemonic medium than is a priori obvious.

we also need to step back and think more skeptically about questions such as: is this medium really working?

How important a role does memory play in cognition, anyway?

Expanding the scope of memory systems: what types of understanding can they be used for?

In modern times, many memory systems have been developed. Among the better known are Anki, SuperMemo, Quizlet, Duolingo, and Memrise.

SRM systems are most widely used in language learning

What about memory systems for uses beyond language? Quizlet is popular

there is a thriving population of medical students using Anki..

Other people have also developed ways of using memory systems for abstract, conceptual knowledge. Perhaps most prominently, the creator of the SuperMemo system, Piotr Wozniak, has written extensively about the many ingenious ways he uses memory systems

employing those strategies requires considerable skill

By contrast, in Quantum Country an expert writes the cards, an expert who is skilled not only in the subject matter of the essay, but also in strategies which can be used to encode abstract, conceptual knowledge.

Improving the mnemonic medium: making better cards

cards are fundamental building blocks of the mnemonic medium, and card-writing is better thought of as an open-ended skill.

answering the question “how to write good cards?” requires thinking hard about your theory of knowledge and how to represent it, and your theory of learning.

All that said, let's make a few concrete observations about good card-writing.

Note that these are just three of many more principles – a more detailed discussion of good principles of card construction may be found in Augmenting Long-term Memory.

Most questions and answers should be atomic

Make sure the early questions in a mnemonic essay are trivial: it helps many users realize they aren't paying enough attention as they read

Avoid orphan cards

for the mnemonic medium to work effectively, spaced repetition must be deployed in concert with many other ideas. The three ideas we just described – atomicity of questions and answers, making early questions trivial, avoiding orphan cards – are just three of dozens of important ideas used in the mnemonic medium

One of us has previously asserted (Michael Nielsen, Augmenting Long-Term Memory (2018)) that in spaced-repetition memory systems, users need to make their own cards. (2018-07-31-NielsenAugmentingLongTermMemory)

The reason seems to be that making the cards is itself an important act of understanding

Quantum Country violates this principle, since users are not making the cards. This violation was a major concern when we began working on Quantum Country. However, preliminary user feedback suggests it has worked out adequately.

Here's three questions suggesting experiments in this vein:

How can we ensure users don't just learn surface features of questions?

How to best help users when they forget the answer to a question?

How to encode stories in the mnemonic medium?

In fact, there are ideas about memory very different from spaced repetition, but of comparable power. One such idea is elaborative encoding. Roughly speaking, this is the idea that the richer the associations we have to a concept, the better we will remember it. As a consequence, we can improve our memory by enriching that network of associations. (Associative)

Here's three more suggestions which build on elaborative encoding:

Provide questions and answers in multiple forms

pictures and words together are often recalled substantially better than words alone.

Vary the context

studying material in two different places, instead of twice in the same place, provided a 40% improvement in later recall.

How do the cards interact with one another? What is the ideal network structure of knowledge?

Two cheers for mnemonic techniques

An enjoyable extended introduction to such techniques may be found in Joshua Foer's book “Moonwalking with Einstein” (2012)..

techniques such as the method of loci

memory palace

even novices are often shocked by how well such techniques work, with just a small amount of practice.

We're enthusiastic about such mnemonic techniques. But it's important to understand their limitations, and not be bedazzled by the impressiveness of someone who can rapidly memorize a deck of cards.

One caution concerns the range of what can be memorized using mnemonic techniques

Furthermore, the mnemonic techniques tend to be much better suited for concrete objects than abstract conceptual knowledge

A second caution relates to elaborative encoding.

when an expert learns new information in their field, they don't make up artificial connections to their memory palace. Instead, they find meaningful connections to what they already know.

We've had people go so far as to tell us that mnemonics make memory a solved problem. That is simply false. But with their limitations understood, they're a powerful tool

How important is memory, anyway?

group is skeptical or even repulsed. In caricature, they say: “Why should I care about memory? I want deeper kinds of understanding! Can't I just look stuff up on the internet? I want creativity! I want conceptual understanding! I want to know how to solve important problems! Only dull, detail-obsessed grinds focus on rote memory.”

one of us (MN) has often helped students learn technical subjects such as quantum mechanics. He noticed that people often think they're getting stuck on esoteric, complex issues. But, as suggested in the introduction to this essay, often what's really going on is that they're having a hard time with basic notation and terminology. It's difficult to understand quantum mechanics when you're unclear about every third word or piece of notation.

When people respond to the mnemonic medium with “why do you focus on all that boring memory stuff?”, they are missing the point. By largely automating away the problem of memory, the mnemonic medium makes it easier for people to spend more time focusing on other parts of learning, such as conceptual issues.

It is good to use what you're learning as part of your creative projects. Indeed, an ideal memory system might help that happen, prompting you as you work, rather than in an artificial card-based environment.

It's in this phase that memory systems shine. They can accelerate people through the awkward early stages of learning a subject. Ideally, they'll support and enable work on creative projects

An immense amount of research has been done on the relationship of memory to mastery.

one especially interesting series of studies was done in the 1970s by Herbert Simon and his collaborators. They studied chess players

learn to recognize somewhere between 25,000 and 100,000 patterns of chess pieces. These much more elaborate “chunks” (chunking) are combinations of pieces that the players perceive as a unity.

subsequent studies have found similar results in other areas of expertise

We've identified some ways in which criticisms of memory systems are mistaken or miss the point. But what about the ways in which those criticisms are insightful? What are the shortcomings of memory systems? In what ways should we be wary of them?

Memory systems don't make it easy to decide what to memorize

What's the real impact of the mnemonic medium on people's cognition? How does it change people's behavior?

How to invent Hindu-Arabic numerals?

what design process could take you from Roman numerals to Hindu-Arabic numerals?

From this discussion, we take away a warning and an aspiration.

The warning is this: conventional tech industry product practice will not produce deep enough subject matter insights to create transformative tools for thought.

That practice has been astoundingly successful at its purpose: creating great businesses. But it's also what Alan Kay has dubbed a pop culture, not a research culture

You need the insight-through-making loop to operate, whereby deep, original insights about the subject feed back to change and improve the system, and changes to the system result in deep, original insights about the subject.

Note that we are not making the common argument that making new tools can lead to new subject matter insights for the toolmaker, and vice versa. This is correct, but is much weaker than what we are saying. Rather: making new tools can lead to new subject matter insights for humanity as a whole (i.e., significant original research insights), and vice versa, and this would ideally be a rapidly-turning loop to develop the most transformative tools.

Part II: Exploring tools for thought more broadly

Mnemonic video

It's tempting to overlook or undervalue this kind of emotional connection to a subject. But it's the foundation of all effective learning and of all effective action. And it is much easier to create such an emotional connection using video than using text.

There's a flipside to this emotional connection, however. We've often heard people describe Sanderson's videos as about “teaching mathematics”. But in conversation he's told us he doesn't think more than a small fraction of viewers are taking away much detailed understanding of mathematics.

we plan to develop a mnemonic video form that provides both the emotional connection possible in video, and the mastery of details possible in the mnemonic medium.

Creating such a form is challenging. Many MOOC platforms have attempted something a little in this vein.

Why isn't there more work on tools for thought today?

Many pioneers of computing have been deeply disappointed in the limited use of computers as tools to improve human cognition.

Our experience is that many of today's technology leaders genuinely venerate Engelbart, Kay, and their colleagues. Many even feel that computers have huge potential as tools for improving human thinking. But they don't see how to build good businesses around developing new tools for thought. And without such business opportunities, work languishes.

Unfortunately for Adobe, such mediums are extremely expensive to develop, and it's difficult to prevent other companies from cheaply copying the ideas or developing near-equivalents. Consider, for example, the way the program Sketch has eaten into Adobe's market share

Put another way, many tools for thought are public goods. They often cost a lot to develop initially, but it's easy for others to duplicate and improve on them, free riding on the initial investment. While such duplication and improvement is good for our society as a whole, it's bad for the companies that make that initial investment.

the process-level explanation is a consequence of the public goods explanation: companies don't use the necessary processes because there's little value to them in doing so.

It's illuminating to contrast with video games. Game companies develop many genuinely new interface ideas.

there's a big difference between video game companies and companies such as Adobe. Many video games make most of their money from the first few months of sales.

By contrast, a company like Adobe builds their business around distribution and long-term lock in.

Another plausible solution to the public goods problem is patents, granting a temporary monopoly over use of an invention

the patent system is broken… What happened? A patent gold rush built by patent profiteers

Innovative companies can easily be attacked by patent trolls who have made broad and often rather vague claims in a huge portfolio of patents, none of which they've worked out in much detail. But when the innovative companies develop (at much greater cost) and ship a genuinely good new idea, others can often copy the essential core of that idea, while varying it enough to plausibly evade any patent

Is it possible to avoid the public goods problem altogether? Here's three classes of tools for thought which do:

Search engines such as Google are tools for thought. They avoid the public goods problem because their value is in their brand and in hard-to-duplicate and capital intensive backend elements

A service such as Twitter can be considered a tool for collective thought. While the interface is easily copied, the company is hard to duplicate, due to network effects

Novel hardware devices (e.g., for VR, or the Wii remote, or for new musical instruments)

many promising directions – including ideas such as the mnemonic medium and mnemonic videos – involve a substantial public goods element. Is it possible to solve the public goods problem in such cases? The two most promising approaches seem to us to be:

Philanthropic funding for research.

The model used by Adobe and similar companies, in which new tools for thought are a central part of the company's operations, but not the core of their competitive moat

Questioning our basic premises

What if the best tools for thought have already been discovered?

Really difficult problems – problems like inventing Hindu-Arabic numerals – aren't solved by good intentions and interest alone. A major thing missing is foundational ideas powerful enough to make progress

When small groups of motivated people do – as in pioneering labs such as PARC, SRI, and other DARPA-inspired early efforts, as well as modern labs such as Dynamicland – they make rapid progress. (R-and-D)

To us, that suggests scaling them up, becoming much more ambitious

Isn't this what the tech industry does? Isn't there a lot of ongoing progress on tools for thought?

while those may be valuable tools, they're certainly not “tools for thought”

Still, there are tech companies which really do develop tools for thought

Programmer tools are a case where the insight-through-making loop operates quite well.

seem likely that many of the most fundamental and powerful tools for thought do suffer the public goods problem. And that means tech companies focus elsewhere; it means many imaginative and ambitious people decide to focus elsewhere; it means we haven't developed the powerful practices needed to do work in the area, and a result the field is still in a pre-disciplinary stage.

Why not work on artificial general intelligence (AGI) or brain-computer interfaces (BCI) instead?

One of us wrote a book about artificial intelligence before deciding to focus primarily on tools for thought; it was not a decision made lightly, and it's one he revisits from time to time.

It seems plausible to us that work on tools for thought will be, over the next few decades, more important than work on AGI and BCI

Executable books

The computer scientist Peter Norvig has written an interactive essay discussing the distribution of wealth (wealth inequality) in society. Norvig's essay is a Jupyter notebook which expresses many of the ideas in running Python code.

That code sets up a population of agents

Part of what makes Norvig's essay beautiful is that with just a few lines of Python code Norvig is able to show some surprising results about wealth inequality.

Results like these will challenge the intuition of some users. But instead of those challenges being on the basis of easily-ignored abstract arguments, users can immediately engage with Norvig's model.

Norvig's essay is one of thousands (or perhaps even millions) of Jupyter notebooks that have been created.

Of course, systems like Jupyter go back decades. There are antecedents in Knuth's notion of literate programming, in Mathematica notebooks, in PARC's Learning Research Group, in the PLATO system (and, more broadly, computer-assisted instruction), to name but a few

We described Norvig's essay as an “interactive essay”. It's useful to have a more specific term, to distinguish it from other interactive forms, like the mnemonic medium. In this essay, we'll use the term “executable book”

Tools for thought must be developed while doing serious work. The aspiration to canonical content

Seymour Papert, one of the principal creators of the Logo programming language, had a remarkable aspiration for Logo.

Papert wanted to create an immersive environment – a kind of “Mathland” – in which children could be immersed in mathematical ideas. In essence, children could learn differential geometry by going to Mathland.

But as far as we know, no professional differential geometer (or, more generally, mathematician) uses Logo seriously as a tool in their work. And upon reflection that seems troubling

How do the creators of Logo know that mastering Logo helps later with understanding real (forgive us!) differential geometry?

At the end of Norvig's economics essay is a short afterword explaining how he came to write the essay. Shortly before writing the essay he'd heard about the kinds of economic models discussed in the notebook, and he wanted to explore several questions about them. After talking it over with some colleagues they decided to each independently attack the problems, and to compare notes.

There's a lot of work on tools for thought that takes the form of toys, or “educational” environments. Tools for writing that aren't used by actual writers. Tools for mathematics that aren't used by actual mathematicians.

Often the creators of these toys have not ever done serious original work in the subjects for which they are supposedly building tools. How can they know what needs to be included?

There's a general principle here: good tools for thought arise mostly as a byproduct of doing original work on serious problems.

In serious mediums, there's a notion of canonical media.

For instance, Citizen Kane, The Godfather, and 2001 all expanded the range of film

YouTubers like Grant Sanderson have created canonical videos

And something like the Feynman Lectures on Physics does it for textbooks

there is no Citizen Kane of Jupyter notebooks

Aspiring to canonicity, one fun project would be to take the most recent IPCC climate assessment report (perhaps starting with a small part), and develop a version which is executable

One promising exploration in this direction is The Structure and Interpretation of Classical Mechanics, a beautiful executable book building up classical mechanics.

It uses a much more powerful underlying model than Jupyter, developing a new symbolic language as part of the book.

Stronger emotional connection through an inverted writing structure

Consider an author writing a popular book about quantum mechanics. Such an author is in a strong position: they can begin their book with astonishing phenomena such as black hole evaporation, quantum teleportation, and the role of quantum fluctuations in the early universe.

These are the kind of things which touch a chord inside many, perhaps most people. And so it's relatively easy to draw readers in, to get them engaged, and keep them connected.

By contrast, consider a typical technical book about quantum mechanics

One problem is that a person can spend years reading analogies about black hole evaporation, quantum teleportation, and so on. And at the end of all that reading they typically have… not much genuine understanding to show for it.

It's striking to contrast conventional technical books with the possibilities enabled by executable books

They could experiment with different parts of the quantum teleportation protocol, illustrating immediately the most striking ideas about it. The user wouldn't necessarily understand all that was going on. But they'd begin to internalize an accurate picture of the meaning of teleportation.

A similar argument has been made by Rachel Thomas, Providing a Good Education in Deep Learning (2016).

The masters of this are video game designers. See, for example, Dan Cook, Building a Princess Saving App (2008), and Jonathan Blow and Marc ten Bosch, Designing to Reveal the Nature of the Universe (2011).

Summary and Conclusion

Memory systems make memory into a choice, rather than an event left up to chance

Memory systems are in their infancy

What would a virtuoso use of the mnemonic medium look like?

Memory systems can be used to build genuine conceptual understanding, not just learn facts

Mnemonic techniques such as memory palaces are great, but not versatile enough to build genuine conceptual understanding:

Memory is far more important than people tend to think

The mnemonic medium is merely one prototype tool for thought

What practices would lead to tools for thought as transformative as Hindu-Arabic numerals? And in what ways does modern design practice and tech industry product practice fall short?

Tools for thought are (mostly) public goods, and as a result are undersupplied:

Take emotion seriously:

Tools for thought must be developed in tandem with deep, original creative work:

Let's return to the question that began the essay: how to build transformative tools for thought? Of course, we haven't even precisely defined what such transformative tools are! But they're the kind of tools where relatively low cost changes in practice produce transformative changes in outcome – non-linear returns and qualitative shifts in thinking.

Historically, most invention of tools for thought has been done bespoke, by inspired individuals and groups. But we believe that in the future there will be an established community that routinely does this kind of invention.