If doing X helped, do more of it. If doing X wasn’t helpful, stop. (Indeed, there might be a period of time in your life, where you have “protected time” for this kind of self-experimentation - such as when you are a student - if so, take advantage of it!) By trying out different strategies or practices - meditating daily, reading a book a week, writing 10 pages every morning - and then documenting and analyzing the outcomes, you can, over time, learn how your brain works. Over time, you begin to know yourself. You might eventually form intuitions about your mind, which help you choose paths that are better for you, with less struggle.

1. Treating your life as an experiment

Each of us (Ed, Claire, Nina) has benefited from such self-experimentation. We’ll now give a few examples. Through such experimentation and analysis, one of us (Ed) learned lots of tricks about his mind. (As in past essays, we’ll switch into singular first-person voice, for these personal narratives!) I learned that I needed to put my tasks on a calendar, rather than on a todo list. A calendar forced me to schedule tasks to begin at specific times, and to block off enough time to finish each task. In contrast, a todo list could grow endlessly, without enough time being reserved to complete any task. Forcing myself to assign tasks to specific times, curiously, led me to develop a productive use for procrastination. If I couldn’t complete a task in the time allotted, I would need to find a new time to finish it (or perhaps I would reschedule the task entirely, before beginning it). I decided that if I procrastinated a task 3 times, then I would analyze why this task needed procrastination: perhaps it needed to be broken down into parts, or increased in priority and given more time, or perhaps canceled altogether. Most usefully, quite often I would realize that I was missing a preparatory step: I was trying to do something before I was ready, and needed to do some other task first. In short, I turned procrastination into a way to analyze tasks for hidden assumptions. (If your goal is to climb an icy mountain, but nobody’s invented the ice axe yet - perhaps your most urgent goal should be to invent the ice axe.)

As a second example, I had trouble sustaining my attention to read scientific textbooks in college (and at one point, I was even diagnosed with ADHD). I tried different strategies - highlighting text, reading text out loud, studying with friends - until I discovered that if I copied books by hand (summarizing when possible, of course), tagging each thing that I read with the kind of knowledge it was (e.g., labeling a sentence with “@” for example, “Δ” for definition, “A” for assumption, “R” for result, etc. - a strategy I called “semantic tagging”), I could force myself to pay attention to the knowledge. Soon, I began going beyond mere attention and started drawing arrows between concepts as I realized what kind of knowledge each sentence was, linking them into emergent networks of insight. In this way, I made my way through many entire textbooks. Also, summarizing helped me remember content better (and I also got, in the end, a useful reference sheet). I could go on - such experiments led me, over the years, to sleep early and wake early, to have a “people advisory board” to give me perspective on difficult situations, to tune my caffeine intake, and more.

Years ago, another of us (Nina) spent time experimenting with ways to become more productive. Because of this period of my life, I regularly get asked for tips on efficiency, but I am hesitant to answer, since different people might be unproductive for different personal reasons. And, different such reasons might benefit from different productivity tools. At the beginning of my journey, I tried conventional tools like Apple Notes and Google Calendar, but they didn’t address the personal reasons I was being unproductive. Indeed, Notes felt ‘messy’ - I couldn’t find a good way to visualize the priority of my tasks. To solve this, I tried bolding text to denote task importance, but that felt too binary. Then, I switched to Google Calendar, and while I could color-code tasks by type and priority, I found that scheduling tasks for particular times was too rigid for my hectic and dynamic schedule (curiously, the opposite of Ed’s finding, driving home the point of this article!). Finally, I switched to a Notion page, with color-coded blocks representing tasks that I could shift around from day to day. This has been my trusty setup for six years now.

That all being said, I notice that many people asking for productivity advice already know what they have to do, and just do not want to do it! No shame here; I’ve definitely felt this way in periods of my life before. It is important to note that the fix here is not simply task tracking, which, for some people, doesn’t help them attain the motivation to do the work. In this situation, reframing your mentality around work might be a better bet. How can you do that? Back in freshman year, when I felt this way, I ran self-experiments to understand what was at the root of this issue. I would permute a variable about the work - for example, choosing assignments that vary in the amount of writing I committed to do - and see how I felt. I noticed that when I would tell myself that I would not feel guilty if I did not do the work, I felt a lot better about doing the work. For you, though, it may be different. Improvement requires honest self-awareness and a willingness to experiment long enough to figure out what the real problem is.

For another of us (Claire), self-experimentation has been useful for choosing a career path. Many people offered me advice on the ‘right’ thing to work on - what sort of job would be best, or how to gain optionality. Much of it sounded cliché. So rather than accepting or rejecting it, I tested it. I took a gap semester and deliberately set up contrasts: I worked as a machine learning engineer at a neurotech research startup (a focused research organization (FRO) in SF), then pivoted to a systems internship at Apple - a completely different field, scale, and culture. I ran shorter-term experiments too - rotating through wet lab, hardware, and computational work to see which held my attention. Through analysis and a lot of writing, I was able to pinpoint specifics: I loved research engineering but not pure engineering, and I found that FROs delivered on their promise of doing scalable science faster. Perhaps most unexpectedly, the gap taught me that confidence doesn’t come from accumulating personal proof, but rather, for me, it came from watching other people who, when faced with a hard problem, simply decided they’d figure it out. That reframe changed me more than any single job did. It turns out the cliché advice (”explore broadly,” “try things,” “gain optionality”) wasn’t wrong, but it was useless to me until I had the experiential ground truth to parse it.

As a second example - on a more personal level, a high school friend pointed out to me that I relied on changes in environment to fix internal problems: choosing colleges, cities, and jobs as though the right place would make me the right person. She was ultimately right, but it took a few rounds of experimentation - and experiment failure - to appreciate her insight: moving away from MIT to SF didn’t fix the things I wanted fixed, and then coming back to MIT didn’t either. What worked was changing behavior, consciously, within an environment: joining a dorm with a stronger community helped me overcome loneliness more than moving to a different city, starting a radio show helped me spend more time on what I loved than just surrounding myself with music people and hoping for growth, and structuring weekly rituals with friends helped more than endless scheduled hangouts. Such insights might seem banal, but I couldn’t have received them as advice. I had to fail first. (As the old saying goes, six months in the lab can save you an afternoon in the library.)

Of course, self-experimentation works best when you can rapidly see the outcome of a choice - feedback must arrive fast enough for you to choose your next action appropriately, given what you learned from the outcome of your last choice. But sometimes a clear outcome is quite delayed, relative to the time of making a decision. One might, in such a situation, pursue a heuristic such as regret minimization: will I regret the negative outcome of trying, vs. not taking the risk at all? But sometimes the stakes are big, and the outcome is delayed for years or even decades. Should you quit your job and start a company, or stay in your current job? Should you do a PhD, or go into industry? Should you pursue idea X, or idea Y, for the next 5 years? In such situations, where the outcomes might not become clear for years, you can’t rely only on your own future experiences - you might need to rely on the experiences and thoughts of others. You might need… advice.

As we noted above, though, the advice-giver might tell the advice-asker exactly the wrong thing. For every entrepreneur who needed encouragement amidst their startup’s struggles, another would have been better off moving on from the dire reality of their situation. For every dropout who quit school, another might have benefited from more educational structure. The advice-giver almost has too much responsibility: to give good advice, they must have an almost otherworldly knowledge of their listener - indeed, more knowledge about their listener than the listener themselves might have. So, the advice-asker has a hidden burden to bear and some tacit tasks to pursue.

First of all, the advice-asker must choose who to ask for advice from. The advice someone gives will be biased by their own experiences and actions - many of which will not apply to the asker. For example, the advice-giver may have grown up or made their choices or achieved their goals in an earlier time, and now the world is different. What was a good decision in 1995 or 2005 might be a bad one today. Second, the advice-asker must choose the right question to ask. Of course, quite often, the advice-asker doesn’t know enough to choose well - that’s why they are asking for advice in the first place. Asking for advice with too narrow a scope, for example, could backfire. Imagine that an advice-asker asks an advice-giver whether one should join company A or company B. But perhaps, the question they really should be asking is whether they should join a company in the first place, vs. starting their own.

By asking the wrong question, the advice-asker would artificially limit the scope of answers that they could receive, or even send the advice-giver down the wrong track entirely. Perhaps a good advice-giver will recognize this and invoke a higher-level framing in their response. But there are lots of potential advice-givers with important lived experience, or stunning accomplishments, who might not know how to project their value into their listener’s reference frame, or possess the empathy to figure things out on the fly. So in the end, the advice-asker needs a good strategy in order to make the most of the advice they receive. Ideally, the advice-asker doesn’t ask a question that is impossible to answer, and then simply follows the response without questioning.

2. When you hear advice, simulate experiments on your life

We started this essay by talking about experimenting on your own life. Listening to advice might best be followed by simulating an experiment on your life. This could take many forms, which are thankfully learnable and teachable. For example, if you are wondering which path to take and want to seek advice, you might start by making a tiling tree of all the possible paths you could take and then ask people for feedback on each path. Or, if you ask various advice-givers for advice on a question, and you get several different answers, you might create a tiling tree to organize those answers, and then use the tiling tree method to extend the tree to generate alternative answers that perhaps no advice-giver would think to give.

A simplified strategy would be to ask someone for advice, and then visualize, in as much detail as possible, what would happen if you did the opposite. Don’t just visualize the consequences, but also visualize the emotions associated with those outcomes. Depending on the topic at hand, you might need to think many years out. If, after you do that visualization, you realize that the opposite path looks better than the path you were advised to follow, you just might decide to take the opposite path! (By the way, this is a generally useful strategy for problem solving: consider how people are trying to solve a problem, and then visualize what would happen if you pursued the most opposite possible idea. You can read about an example of this strategy working in the story of inventing expansion microscopy1.)

In our own lives, this way of thinking has proven useful. We’ll again give some examples, switching to the first person singular voice. For one of us (Ed): when I started my lab at MIT, I recall some senior scientists telling me that I shouldn’t collaborate too much with other scientists when doing my research. After all, to be awarded tenure at MIT, I’d need to develop an identity distinct from those of other scientists. And, things were already off to a bad start - I was rejected from most of the places I applied to for faculty jobs, including all the original departments at MIT that I had applied to2. But then I considered the opposite situation: if I were building tools and collaborating with many people to generate them and to apply them to problems, my group could become a hub within a huge network, generating lots of impact. I tried to visualize what this would look like - 1, 5, 10 years out. My group ended up collaborating with other groups on everything from molecular tools for controlling the brain with light to new strategies to noninvasively help treat Alzheimer’s to nanofabricated neural probes and beyond. We sent technologies out to other research groups tens of thousands of times over the years (working with partners like Addgene), sharing as freely as possible3. The strategy worked - we became a hub in a vast network that could generate innovations sustainably and usefully. (And, I did get tenure at MIT.)

For another of us (Claire), I kept hearing versions of ‘be realistic’ - about what I could change, about what was within my control, about what someone my age should expect. The advice came from a kind place, but I noticed it mostly functioned as permission to not try. When I considered the opposite, that I actually could change most things about my situation if I was willing to act and make big changes around me, things started moving. For instance, everyone told me to stop overcommitting and learn to say no, but the real fix was the opposite: instead, I leaned into committing hard to specific things I actually cared about, like a radio show, a demanding fabrication class, and weekly dinners I hosted - and, naturally, I stopped spreading myself thin across obligations I didn’t care about. The problem was never volume; it was direction. Similarly, I worried my interest in neurotech was just momentum from high school, and the standard advice was to stick with what I knew. Instead, I deliberately pivoted to jobs in completely different fields and cultures, like computer systems or robotics, which ended up confirming I actually did love neurotech, while also teaching me what I didn’t enjoy (e.g., pure software engineering). I needed the opposite path to validate my original one.

For another of us (Nina) - I get told often that I should temper how I present my goals. For context, a major life goal of mine is to cure Alzheimer’s disease. Friends and mentors have often told me that when applying for fellowships or speaking to other scientists, I should take more effort to make my goals sound reasonable. And on one hand, I get it: overly ambitious goals can make you sound unrealistic. But I have never had so much conviction about anything in my life. And I’ve come to realize that in interviews and conversations, that belief actually excites the people whom I’m talking to. I don’t completely ignore the advice, though; I recognize that my mentors’ concerns come from a place of worry that people will think I don’t understand the sheer complexity of biomedical research when I state something so ambitious. This is why I always follow up my excitement with context, for example, sharing my thoughts on experiments, funding policies, and changes in the field as a whole that I believe are needed to get there.

Lesson learned: the opposite of good advice, could be better advice. For any advice you hear, consider the opposite of it – the anti-advice, if you will. Of course, don’t just blindly do the opposite: the opposite of good advice could also be terrible advice. Instead, visualize - that is, simulate in your mind - the outcome of following the advice that you hear, as well as the outcome of following the opposite of that advice. If you have a mentor and are getting advice from them over an extended period of time, you might repeatedly apply this method in order to generate many instances of anti-advice over the years. (If you have a mentor whose anti-advice that you generate from their advice is repeatedly better than the advice they give you, you might consider them to be an anti-mentor. Pro tip: don’t tell them they’re your anti-mentor.)

3. Solving a bigger problem can be easier than solving a smaller one

One corollary of this way of thinking is, “don’t take requests.” As the old saying goes, if Henry Ford asked customers what they wanted, they would all say they wanted a faster horse. His customers just didn’t have the concept of an automobile in their minds. Steve Jobs famously didn’t like market research: how would his customers know they wanted a phone with a touchscreen when they had never seen one that worked? You can consider this as a form of frame reshaping: someone providing a request to you will implicitly control the range of variables to be considered (e.g., a request for a faster horse limits the frame to that of transportation animals); you can reshape the frame (in the horse example, by broadening the frame to include all forms of transportation, including mechanical possibilities). Of course, it’s great to learn what people want - but don’t blindly implement what you learn; take a step back and think about the bigger context of the request: what does the request mean amidst the state of the world, and when augmented by imagined possibility?

Curiously, solving a bigger problem is sometimes easier than solving a smaller problem: the reason is, when you take a step back from a narrowly scoped problem, and consider a bigger problem space, you might also gain access to a bigger solution space - and sometimes, that bigger solution space will enable a far more efficacious, low-risk, or practical outcome than the original narrow solution space would have permitted. To continue the horse/car example: broadening the space of transportation methods beyond just animals, to include machinery innovations, makes the problem far easier to solve. Building a car turned out to be quite practical, whereas breeding a faster horse would have been extremely difficult, given the science of the time. Of course, if what you end up doing doesn’t cleanly fit into people’s expectations, you might need to work on it alone (or with trusted, inner-circle colleagues) for a while. It might take time until what you’re doing has matured to the point where it connects with people’s needs. Hence the need for “skunk works,” where people work in relative isolation, protected from too much external critique, in the short term. Perhaps this is why “involuntary collaboration”4 is such a frequent pattern in revolutionary innovations - it allows each step to be done in isolation, with information sharing helping the whole to emerge over time, as far more than the sum of the parts.

Another nice corollary of this way of thinking is that you can learn to be your own mentor. You can propose some advice to yourself, and then visualize the outcome of following it - and then repeat the exercise for the opposite advice. It’s as if you have your own mentor and anti-mentor constantly offering you help at all times. (And, as you progress, you might start generating advice tiling trees, as a systematic way of generating all the possible kinds of advice that you could give yourself.) Of course, you can be your own meta-mentor, as well - finding a set of mentors for yourself, and then assembling all that you learn into your own framework. As you walk around and see potential solutions, actions, and possibilities, you can constantly think, “What if I did the opposite of that?” It can become a reflex that yields on-the-fly creativity, to visualize the opposite of what you see before you. It becomes a fun game you can play to help you question everything, generating innovations as you go throughout your day.

Generating anti-advice can be difficult, in the beginning. Indeed, it can be hard to take advice in general: a common saying is that if a piece of advice is hard to hear, that means you should take it. But perhaps advice is hard to hear because it’s bad advice, or perhaps there’s something in your internal state that you need to process before you’re ready to take any advice at all. Sometimes you need to have the advice broken down into steps - what you hear at first may present too big a leap. Comfort is not always a bad thing - it can help you realize what you need to do to get where you want to go. By considering advice and its opposite, you can optimize for the big picture.

This essay was generated as the result of many years of observation and 20 years of operation of the Synthetic Neurobiology group5. But as we wrote it, we encountered synergistic documents, both in our own past and by others. Claire wrote a less scientific-research-focused post in the past, which we’ve summarized when applicable, at points throughout this article6. There are additional notes at that link, on giving/getting advice in less career-based situations. We also came across an article by Slate Star Codex, which presented a related concept, to ‘reverse any advice you hear’7.

In any case, when asked to give advice, one of us (Ed) usually offers a few pieces of advice (with copious caveats, as you might guess). But then I end by saying, ‘whenever you hear advice, consider doing the opposite.’ I then add, paradoxically and half jokingly, ‘including that advice, which I just gave you.’ Sometimes it lands.

A student at a lunch I hosted at a university I was lecturing at, emailed me a few months later. At the time, she’d been considering starting a company, but faculty at her university advised against it. She converted their advice to anti-advice, founded the company, and became its chief science officer. It ended up bringing forth life-saving ideas to the world. She’s now working on her second startup, which aims to bring forth therapeutic technologies in a second problem space. Perhaps you will also save lives or change the world by following anti-advice or finding anti-mentors.

And perhaps through struggling with such processes, you will also learn to give better advice to others down the road.

Thinking of every possible solution for a given problem can be a tricky task. How can we possibly catalogue all of the solutions that could be done - and then, to choose the optimal path, perhaps one that has never been thought of before?

We described the Tiling Tree Method1, which enables you to do just this, in an earlier article. In summary, this method requires one to perform iterative splits on a solution space, dividing a set of possible solutions into increasingly smaller subsets. The resulting diagram looks like a tree. The subsets carved from a parent set should “tile” the space of possibilities - that is, one subset should not overlap with another subset, and the subsets, taken collectively, should not leave out any possibility. Ultimately, you might arrive at individual ideas, the “leaves” of the tree, which then could be evaluated for their impact, uniqueness, and feasibility - or whatever criteria are at hand. (See the aforementioned article for examples, and general guidelines.)

This strategy may appear simple to execute, yet to do it well is quite nuanced, since there are many ways to do it wrong (though practice, as with any skill, is extremely helpful). We (Claire and Nina) are newer to the art than Ed, and for the purposes of this blog and our own learning, decided to create a tree to help us think of novel ways to deliver genes into the body - a big need in basic biology research, and also in therapeutic treatment of diseases. Delivery vehicles for genetic material could be very useful for gene therapy, for example. We wanted to interact, as we generated the trees, as researchers might while working at the bench in the lab, getting feedback from Ed along the way.

We went through multiple iterations and learned a lot each time. Importantly, the changes we made over the iterations, addressed mistakes that people commonly make when generating such trees. Thus, our example may help you avoid pitfalls, and make trees that truly tile—helpful for generating novel ideas.

Our first tiling tree

This was the first tiling tree we generated for the problem of gene delivery, and while it may seem passable, it actually has a few flaws that may prevent us from coming up with novel ideas.

Please note - the example we chose here may seem a bit technical. We will try to emphasize the logic that we applied for each decision, so don’t over worry about the technical details - although, if you want to learn more, there are lots of references you can identify (e.g., by searching Google Scholar, and so forth).

For example, our “delivery vehicle only” branch was split into vesicles, nanocarriers, and “other materials”. This was a retroactive split—we started by listing “vesicles” and “nanocarriers,” categories that we found already to exist in the literature. When you split by known solution types, the resulting tree might only produce variations within those categories. The structure itself could thus prevent you from finding anything that exists outside the taxonomy you’ve imported.

In addition, “other materials” is unsatisfying because it is a catch-all for many, potentially creative, ideas. But by not splitting out those “other materials,” we’ve converted a potential sea of opportunity into a dead end. It’s not just that the term is vague (though it is). It’s that vesicles and nanocarriers were interesting subsets, but there could be equally interesting alternatives that we have explicitly chosen not to identify. We simply left a catch-all bucket for whatever didn’t fit into the pre-existing categories - but there’s nothing to do next. You can’t explore systematically when the set you need to split into subsets is simply called “everything else.” What could the next split possibly be?

Compare that split to splitting by “number of kinds of molecule in the delivery vehicle”. This is a physical property that cuts across existing categories. A vesicle might contain one kind of component, or many, depending on its design. The mismatch between this split and the familiar-from-literature categories forces you to think about vesicles in new configurations. A single-component lipid-bound structure might behave extremely differently than a multi-component one - the latter might even approach the complexity of a simple living cell - such designs might not map cleanly onto “vesicle” at all. If your branches are named after solution subsets you could already Google, you could be just taxonomizing, not deconstructing in a fundamental way.

We also split by “forced entry” vs “natural process.” What does natural mean here? If you define it as “using endogenous cellular machinery,” then perhaps some viruses (despite being in nature) treat a cell naturally, and some treat the cell unnaturally. If you define it as “no exertion of physical force,” however, then you might consider a different set of viruses to use natural vs. unnatural processes. The word “natural” doesn’t map to a unique or specific physical property; it could mean different things in different contexts.

This vague categorization creates overlap between subsets split from a parent set. Consider endocytosis triggered by engineered ligands binding to receptors. The binding is artificial, but the internalization pathway is endogenous. Is this natural or forced? You can’t place a given solution without an arbitrary choice about which aspect matters more. Such terms are what we sometimes call “words that mean nothing” (WTMNs) - words that are not defined precisely enough to build conceptually on top of them. Of course, one can try to make a list of definitions of such words, and stick with them - that can greatly help, to have a “word choice list.” It can also be helpful to try to be more precise, reductionistic, or even quantitative, in one’s word choices.

Indeed, one fix, in the current example, is to use terms that reduce to atoms and bonds. Instead of forced vs natural, one might try “requires >X kJ/mol external energy input” where X might relate to some biophysically important threshold, or “disrupts lipid bilayer organization” (vs. not). These categories relate to measurements, not verbal interpretations. When you can’t classify a solution, it might be because your split is not a full, clean, mutually exclusive tiling, and that might be because you are using words that don’t have precise meanings.

The test: if replacing a word like “natural” with multiple other words that sound “natural”, but are mechanistically different, results in your solutions falling into different categories, then your original split may have been doing no work.

We tried again:

Second Tiling Tree

Our splits here, while more specific, still contained some vague phrasing and a considerable amount of overlap between branches, as well as splits that did not fully cover the space. For example, what does “based” mean in the context of “protein-based” or “lipid-based”? Can there be any protein in the lipid-based category? Is it a matter of percentages, and if so, is this an effective way to properly tile the space?

Tiling the space means that every possible solution must fall into exactly one category at each split. Like tiles on a floor—no gaps, no overlaps. When you split the delivery vehicles (e.g., those that are systemically delivered to the body, abbreviated “systemic” in the diagram) into “protein-based” and “lipid-based,” what about a vehicle that’s 50% protein and 50% lipid? It could fit both categories (overlap), or neither if you defined the categories as strictly containing one component (meaning many possibilities would be left off the table). The words do not have unique definitions. A proper tile might be “majority protein by mass” vs “not majority protein by mass”—now everything has exactly one home. The test: pick any random solution you can imagine. Can you find it by walking down your tree in exactly one path? If you hesitate at any branch, your tiles might be too fuzzy.

Additionally, look at the “protein-based” branch, which splits into “vector-mediated”, “cell-mediated”, and “field-assisted.” What do “mediated” and “assisted” actually mean, physically? And could something be both vector-mediated and field-assisted (e.g., an electric field could in principle get a virus across a barrier)? There is a bit of a category error here - the different subsets are of qualitatively different kinds. It’s like classifying vegetables as red, sweet, and crunchy - since the categories are along different dimensions, they inherently incur overlap.

A viral vector carrying a payload might fit the colloquial definition of vector-mediated. But the vector still has to interact with the cell membrane. Would that interaction count as cell-mediated? The three categories below “protein-based” are of different kinds (vector = payload carrier, cell = therapy recipient, field = physical mechanism) rather than being of the same kind (e.g., different types of protein-based carrier, or different ways of delivering a protein-based carrier). If we split “protein-based” (with the caveat, above, of its definitional ambiguity) into subsets of the same kind, then we reduce the risk of overlap. We are splitting along one consistent dimension.

Still, this tree generated zero new ideas because it’s organized around existing products. Every terminal node maps to something you could order from a catalog or find in a review paper. The splits encode what the field has done, not what the field could do.

Third Tiling Tree

Suppose we try a new split - say, by “living organism” vs “non-living assembly of multiple macromolecule types” vs “individual macromolecule.” Now, living organisms are subdivided into bacteria, yeast, mammalian cells, and so forth, according to standard taxonomy. Non-living assemblies might include liposomes, hydrogels, crystalline lattices, or even cork or cellulose structures. Individual macromolecules could be viral capsids, engineered proteins, or a strand of DNA that folds into origami.

That split works because the categories are quite distinct - they don’t overlap. (Indeed, as we write these words, we see - this could easily be two splits in a row - living vs. not, and then if not living, made of one thing or many.) The tree reveals gaps: we use mammalian cells, in therapeutics, but not nematodes. Why? Could nematodes have useful drug delivery properties or mechanisms? Curiously, a recent study proposes to use the toxoplasma parasite to deliver therapeutics to the brain2 - showing how vast the true space of possibility might be.

We thus created a final tree, using our insights from the first few iterations. We aimed to make every split as clearly defined as possible, using words with as precise a meaning as possible, and making the tiles as exclusive as possible. Using this tree, we came up with several new ideas (using yeast, nematode cells, and other novel ideas, for effective gene delivery).

Tree 3 attempts to split, whenever possible, by physical properties that don’t care about preexisting categories. The first split: cellular vs acellular (does the vehicle metabolize and replicate as a unit, or not). This forces you to consider whole organisms as delivery vehicles, not just biomolecular assemblies. Within cellular/eukaryotic, we split by organism body length: <1cm, 1-30cm, >30cm. Why? Because this determines practical lab handling, whether something would need to be introduced by surgery vs. bloodstream injection, and immune response scale.

Then, the uni- vs. multi-cellular sequence reveals a gap of possibility: we use bacteria (prokaryotic) and mammalian cells (large eukaryotic multicellular). But what about small eukaryotic unicellular organisms? That branch produces yeast—metabolically complex like mammalian cells, but with bacterial-scale ease of culturing and genetic manipulation. Yeast naturally produce vesicles and have well-characterized secretion pathways. Could yeast be engineered as a therapeutic?

The <1cm multicellular branch produces nematodes. C. elegans is 1 mm long - and, while a standard genetic model in biology, has never been (to our knowledge) used as a delivery vehicle itself. Could you engineer nematodes to migrate to target tissues and secrete therapeutic genes or gene products? Or some other organism, with the complexity of a nematode? The organism’s size makes it imageable in vivo, and its lifespan (weeks) matches many treatment timescales. Curiously, there is a field called “helminthic therapy” - which we only started to discuss after making this tree - in which parasitic worms are used to try to alter the immune system of human patients! Could such worms be used to deliver novel therapeutic agents too?

For acellular paths: “carbon-containing,” followed by “no replication”, followed by “1 component” raises the question - what single-component materials can encapsulate genes? If glucose-containing: perhaps cellulose or starch capsules. These are biocompatible, biodegradable, cheap, and can be chemically modified. On the carbonless branch, we split by a quantitative measure of material rigidity (Young’s modulus), another example of a strict and well-defined split.

In short, if you do not create the perfect tree to represent your problem space at first, do not fret. Spend some time critiquing your tree and thinking from first principles about whether your tree actually accomplishes the goal you set out for. From there, take all the time you need to retile your space and draw new trees. Repeat until you have an end product you’re satisfied with, and you’ll be on your way to solving some big problems!

Subscribe now

Over the years, the Synthetic Neurobiology group at MIT has practiced many learnable and teachable creativity and problem solving skills. Our core hypothesis is that creativity, and problem solving, like any other skill, can be learned and taught. But how often is it that someone takes a class on creativity? Or attends a seminar on problem solving? Such skills seem to be the kind of thing that one is simply expected to pick up, along the way. We want to change that, by making it possible to study, and practice, creativity and problem solving, like any other skill.

In this essay we discuss one of the most powerful methods our group practices - one that we call the “tiling tree” method. In the tiling tree method, one constructs a diagram of all the possible ways of solving a problem. The instructions for doing so are simple to explain, but to follow them well - helpful for generating truly new, impactful, creative ideas - requires attention to detail, logical consistency, and questioning everything one reads or thinks. To make a tiling tree of the solutions to a problem, take the set of all possible solution ideas, and split that set into subsets that do not overlap, but that together comprehensively cover the original set - analogous to the way that tiles might cover a wall completely, without overlapping. Then take each of the new subsets, and split it into yet smaller subsets - again, that don’t overlap, but that together cover all the possibilities.

One way to represent this process, as you split sets into subsets over and over again, is as a tree diagram. Each split of a set is depicted as a set of “branches” radiating downwards from that set, each leading to a smaller subset. Eventually, you might get to sets that contain just one idea - the “leaves” of the tree, which represent the final outcomes of the iterative splitting process. These ideas might be actual projects that one could do. One can then evaluate each candidate project in terms of its potential impact, uniqueness, and feasibility (or whatever other criteria one prefers). Perhaps one could perform a mathematical calculation or computer simulation to assess an idea’s feasibility, or conduct a pilot experimental study to see whether a project might work.

If done properly, the tiling tree method forces one to think of ideas that one may not normally think of. It also guarantees a kind of completeness - you can have the satisfaction that you have not missed out on a possible path. One simple way to split a category into subcategories - although it’s not always the best way of doing so - is a “binary chop” method. Take the set of all of the ideas for how to solve a problem, and split that set into two subsets - one subcategory where each idea has property A, and another subcategory where each idea does not have property A. Let’s consider a toy problem - every possible way of generating energy:

We first split the entire set into renewable and nonrenewable subsets. The renewable subset, in turn, we split into solar and non–solar subsets. What is a renewable, non-solar method of generating energy? Suddenly, we are in a regime of ideas that most of us, perhaps, do not think about every day. Maybe ocean tides could move masses in harbors near a city, driving downstream electricity production. Geothermal energy comes to mind - could there be versions of that, that could be implemented anywhere on earth? The tree is forcing us to think about ideas that we may not have been thinking about before. (As an exercise, you might try to continue extending the tree above.)

Let’s go through some examples that exemplify different aspects of the process. Then, you might try the method yourself, with problems you struggle with - that way you will hone your skill in this practice.

Let’s start by thinking about a cutting-edge area of research, full of excitement and risk, opportunity and innovation - the field of brain-machine interfaces (BMI, also called brain-computer interfaces (BCI)). This may get a bit technical, so we’ll try to give some background information to get everyone on the same page, but don’t worry if some technical details seem too far afield - we’ll emphasize the core take-home messages. By way of background: brain cells, called neurons, compute using (amongst other signals) electrical pulses that last about a thousandth of a second long. Neurons are small - about 100,000 neurons might be packed within a cubic millimeter of the human brain. The ability to record neural electrical pulses allows for extraction of information from the brain, helping us understand how the brain works, and also could help humans control computers and other machines according to their thoughts.

How can we record the high speed electrical activity of neurons in the brain? Because different neurons have different roles, we want to do such recording scalably - recording as many neurons as possible. Perhaps we want to aim for recording the activity of neurons throughout the whole brain, at single cell resolution. This might be especially useful if we want to understand how the brain computes (there is evidence that a single neuron can influence the activity of the entire brain12), although for brain-machine interfaces, such a goal might be less critical - you might be able to control a computer with a tiny fraction of the information of the entire brain. Whole brain neural recording is a very ambitious goal. A first split might be to consider whether the neural data is relayed out of the brain through wires (physical conduits for information-carrying particles or energy to be conveyed), or through wireless (information-carrying particles or energy are conveyed through free space) means:

Now, at any point in making such a diagram, you can evaluate the “leaves” of the tree, at the bottom, for whether the projects associated with them are aligned with your success criteria. In this case, which method of data transmission, wired or wireless, can carry more information, so that we could potentially scale to recording data from a whole mammalian brain (a mouse brain has perhaps 70,000,000 neurons; a human brain, perhaps 90,000,000,000)? One can calculate how much information would be coming out of the entire brain if you recorded every neuron in the whole mouse brain (mice being commonly studied in neuroscience), or throughout the whole human brain. If one does such a calculation for ultrasound, optical, radio, and other wireless modalities of information transfer, one finds that they convey, generally, far less information than with wired means (considering optical fibers to be a kind of wire, using the definition of wire given above). We did such calculations in 2006-2008, and concluded at the time that if we wanted to record from the whole mouse brain, we’d need to go with wires - which led to a project to build 3-d, scalable, nanofabricated electrode arrays, that would convey information out of the brain through wires (published in this series of papers in 2015-2018345). Of course, it’s useful to revisit a tree every now and then, since as underlying or complementary technologies evolve, a bad idea might turn into a good one - this tree above is very useful to redo, at least once a year!

Optogenetics, the use of light to control neural electrical activity6, began in 2000 with a tiling tree of sorts. The brain contains thousands of kinds of cells - small vs. large, excitatory vs. inhibitory, shaped in different ways, and containing different kinds of molecules. Different cell types also change in different ways in different diseases. How can you determine which type of cell contributes to a given emergent function or dysfunction of the brain? Ideally, we could activate or silence the electrical activity of just one kind of neuron, and not its neighbors, upon delivering some kind of energy to the brain. If we could express a gene in the targeted set of neurons (in particular, different cell types differ in their molecular composition, making them targetable through gene therapy methods), and not its neighbors, and the gene product (or protein) could convert delivered energy into electricity, then we could control the activity of just those neurons, with that form of energy. By activating a specific set of cells, you could figure out what behaviors, or pathologies, or therapeutic effects, it was sufficient to trigger. And by silencing a set of cells, you could determine what behaviors, pathologies, or therapeutic effects it was necessary for.

We first set out to determine: what form of energy should we deliver to the brain? We compiled a list of various modalities of energy delivery, including magnetic fields, mechanical force, light, and so forth. The list of the possible forms of energy that you can deliver to the brain is pretty short - the laws of physics are very concise. Then, for each case, we did a calculation: what kind of spatial resolution, temporal resolution, and size of electrical effect, could be achieved by that form of energy, perhaps coupled to an appropriate genetically encoded protein? Light, of course, would have higher spatial resolution than mechanical force or magnetic fields. We also calculated how much effect each form of energy would have on an appropriate genetically encoded protein - for example, how much force a magnetic field could apply to a genetically encoded magnetic bead, fused to an ion channel. Our calculations were helpful in ruling out certain ideas - for example, one of our calculations suggested that for a small magnetic protein connected to an ion channel, the magnetic field might need to be inconveniently large, to actuate the ion channel.

Next came a choice - do we use an existing light sensing protein, or make a new one? Microbial rhodopsins, light-driven ion transporters from single-cell microbes, seemed to be able to create sizable currents in response to modest light fluxes. We started requesting genes encoding for these proteins, from colleagues, and very quickly we were off to the races7. We did a pilot study with one opsin, channelrhodopsin-2, in cultured neurons - and it worked, enabling them to be electrically activated by pulses of blue light, pretty much on the first try. Now, thousands of studies have used optogenetics (opto meaning light, and genetics because these are genetically encoded tools) to study how specific cell types contribute to brain functions or disorders, and some of our molecules are even showing promise in clinical trials for helping people with certain forms of blindness8. Here is the tree:

It can be useful to redo a tree every now and then: perhaps an idea that was bad last year, is now a good idea, because of some change in the world (e.g., AI tools are now available, or genome sequencing is now cheaper, or computers are faster, or some other trend). The tree that yielded optogenetics is such an example: in the years since we did the original thinking, people have characterized many new force-sensitive, magnetic field-sensitive, and other proteins, and therefore some branches of the tree are now more impactful and/or feasible than they appeared to be, in the year 2000. New technologies are emerging, such as sonogenetics and magnetogenetics, which seek to activate neurons expressing mechanical force- and magnetic field-sensitive proteins, respectively. Regarding finding vs. making the protein - in this era of generative AI, perhaps novel protein architectures could exhibit features not achievable with purely natural proteins.

Really stretching oneself to think - did I really tile the space? - is useful. As an exercise - let’s go back to the wired vs. wireless neural recording example. Is there a third alternative to wired and wireless? Defined as above, there are definitely options not covered by the categories we were calling wired and wireless (as you may recall, we were defining wired as relying on “physical conduits for information-carrying particles or energy to be conveyed,” and wireless as “information-carrying particles or energy are conveyed through free space”) - for example, what if we don’t rely on particles or energy at all? Hmm… what about secreting information out of the brain, in chemical form, through the bloodstream, where it can be easily harvested through a simple blood draw?

Curiously, a few groups have recently showed that having the brain secrete chemical payloads into the bloodstream, from a given point in the brain, is indeed feasible910. One of these studies applied this strategy to a gene that is expressed after neurons are electrically active, so that a blood draw could then eavesdrop on the level of neural activity in a given region of the brain. What other possibilities can you come up with?

This example highlights an important strategy for making a tiling tree as useful as possible: make sure that the words used to describe sets and subsets, are very precisely defined. Writing out the definitions of words, as concretely as possible, can help. Otherwise, it’s very easy to make splits that result in categories that subtly overlap, or to make splits that result in categories that don’t actually together tile the space of possibility. Forcing yourself to write out the definition of each word used to describe a category, can reduce the risk of making such incorrect splits.

It can also be helpful to redo a tree over and over again, by splitting a given set into subsets in multiple, distinct ways. Perhaps one might choose different attributes by which to split a category of ideas into subcategories. Different kinds of splits might provoke very different kinds of ideas, because they force you to evaluate an idea space from different angles. A good split should be comprehensive - it might not be very useful to make a long list of individual possibilities within a category, because you might miss something important. Physics and math-oriented splits can be useful, because there is often a very finite number of ways to do something, considered from a physics or math standpoint (the above examples, of wired vs. wireless, and listing forms of energy that can be delivered to the brain, are of this flavor). Other ways to vary the kind of split pursued would be to consider different spatial scales, different temporal scales, the amount of energy involved, or the number of components involved. A good split might expose the trade-offs that govern decision-making in a solution space. Again, it’s important to make sure that the splits are “real” - that the subsets that result don’t overlap.

Can we try a radically different split, for the neural recording example? Brain data must eventually be in digital form, if we are to analyze it in an implanted microchip, or an external computer, or some other digital device. Do we digitize neural data in the brain, and relay out just the bits, or do we digitize the data outside the brain, after relaying out analog info?

This leads to some interesting ideas about whether we put digital electronics inside the brain or not. But what if we take it one level further? Can we digitize neural data, inside the neuron itself?

Hmm… how can we do this? If one could put an electronic device inside a cell, perhaps it could record neural activity into digital form - an idea that some groups are exploring. But what about non-electronic ways of digitizing information, inside a cell? Storing digital information can be done without electronics, by using as the digital medium information-encoding polymers like DNA, RNA, and protein. There are 4 letters (nucleotides) that make up strings of DNA and RNA, and 20 letters (amino acids) that make up proteins - in other words, such biomolecules are digital storage media. Several groups (including ours11121314) independently devised ways to store a history of neural activity in DNA, RNA, or protein form.

It is worth noting that when we originally looked at non-electronic digitizing-in-neuron strategies, we heavily explored DNA- and RNA-based methods. However, the world changed, and advances in protein engineering eventually made a protein-based approach, in which proteins self-assemble into a ticker-tape-like chain to encode the timing of recorded neural events, quite promising. Maintaining a tiling tree and revisiting it periodically as synergistic fields advance, can help reveal when tree branches that might have previously seemed like dead ends, might be ready to yield fruit. If a branch seems especially worthy for some reason, but is blocked by a missing capability, it may be worth revisiting, especially when technologies relevant to that missing capability are rapidly changing. Or perhaps going after that missing capability, through inventive work, would be a worthy project in its own right.

Many people, when they try tiling trees for the first time, end up only summarizing the ideas of the past, and not generating new ideas. That’s not a bad way to start, and can be a useful learning tool - it helps reveal the intellectual structure of past knowledge. But the exercise is most powerful when we force ourselves to think up new ideas. Done properly, you can end up with a roadmap for the future - as we did in the context of neural recording, generalizing some of the approaches talked about above15. One of us (Adam) is now extending variants of this philosophy at new organizations like Convergent Research to do “gap mapping,”16 revealing missing engineering pushes needed across science, to help science funders and researchers choose major unmet needs towards which to invest their time, effort, and resources.

The tiling tree method, not surprisingly, has been invented and reinvented by many people over the years, in different areas of the human enterprise. Fritz Zwicky, an astronomer who made many ahead-of-his-time contributions to the field of astronomy, including proposing the mechanism of supernova formation, neutron stars, and dark matter, came up with many of his ideas with a similar method, that he called “morphological analysis”17. Zwicky himself noted that similar ideas go back centuries, at least to Paracelsus in 153018, and perhaps even earlier. Another example is from the field of consulting - which unfortunately makes it sometimes the butt of jokes - namely, the concept of “mutually exclusive, collectively exhaustive,” or MECE, which sounds similar to the above concept of “tiling”19. Clearly such strategies have been of use in many disciplines and professions over the years. Please give them a try!

In the realm of science, people often talk about Bell Labs, in the 20th century, as a vanguard of innovation, having developed the transistor, the laser, the solar cell, information theory… the list goes on and on. One of the reasons given for Bell Labs’ success was its ability to foster collaboration, maximizing collisions between people in its legendary cafeteria, and encouraging people to keep their doors open. The Bell Labs model worked well for a vast number of 20th century problems. But what about 21st century problems, which often seem to be of much greater complexity? Many of these problems are “network problems” - problems involving complex systems, made out of many building blocks, which interact in complex ways, making it hard to study or fix such systems. Such problems can be overwhelming when treated reductionistically (people often resort to focusing on an ultra-narrow hypothesis in a vast sea of possibilities), yet too approximative when treated phenomenologically (which discards mechanistic detail and results in nonrobust conclusions). Biology and medicine are full of such problems. Can we design a collaborative framework to facilitate great innovations, in such fields?

Let’s examine how great biological and medical inventions and discoveries arose in the past, and see if patterns emerge. If they do, then perhaps we can create learnable, teachable versions of those patterns - aiming to do on purpose, what previous generations did by accident. One observation is that many great biological innovations arose because of different people who contributed unique and independent insights, often over an extended period of time. This observation may sound trivial, but it’s worth exploring how it might guide us to proactively choose our behavior. As one example, take the green fluorescent protein (GFP), which was discovered in a species of jellyfish by a marine biologist in the early 1960s. The gene was then isolated and proposed as a tool - namely, to enable cells, or specific proteins within cells, to glow green, so that they could be tracked under a microscope - by a molecular biologist in 1992. The gene product was shown to work in living cells, by a neuroscientist, in 1994. The protein was made brighter, and mutated to form more colors (e.g., yellow, cyan), by a chemist in the mid-to-late 1990s. In 1999, red fluorescent proteins were discovered, in corals, by another team. In each case, the person or people involved worked at a different institution than the others, and sometimes were even from different fields than the others. Furthermore, the work was done over a period of decades. Today, millions of studies have benefited from using fluorescent proteins to tag cells, and proteins within them, to follow them over time.

As a second example, take CRISPR, which revolutionized genome editing, and is being explored for correcting disease-causing mutations in humans. The original characteristic DNA sequence of CRISPR was discovered accidentally by a molecular biologist in 1987. Around 2005, the protein Cas9, which performs the enzymatic action associated with CRISPR, was first described and identified as a nuclease protein. In 2007, scientists studying bacteria relevant to yogurt manufacture showed that the CRISPR system could help bacteria defend against viruses. In 2012, researchers showed that you could use the Cas9 protein and a specific accessory known as a guide RNA to cut a targeted piece of DNA, demonstrating the programmability of the system. And in 2013, scientists showed that such genome editing could be done in a living human cell. (As with any scientific discovery, there were several other critical steps not mentioned here, key to the ultimate success of the discovery; our goal in this essay was simply to illustrate the distributed nature of discovery, and not to provide a comprehensive history.) Again, people working at different institutions, and approaching the problem from different fields, working over a period of many years, were key to making CRISPR useful.

As a final example, let’s take the case of GLP-1 drugs, which are sweeping the headlines as they help human patients with diabetes, obesity, and other problems. The peptide GLP-1 was discovered in 1986, and its effects on metabolism characterized throughout the late 1980s and beyond. But the natural peptide has a very short half-life in the body, making it difficult to use as a drug. In the early 1990s, work on Gila monster venom, of all things, revealed GLP-1-like molecules that could last longer in the body - long enough to make for a meaningful drug. In the mid-2000s, GLP-1 agonists with ever-increasing efficacy received FDA approval, and by the mid 2010s to early 2020s, easy-to-administer forms, that last for very long durations, went mainstream in medicine. Again, different steps were achieved by different experts, working at different places, and over a long period of time.

Could any of these biology achievements have been conducted under one roof, by a fixed set of people? It is hard to see how. In short, the Bell Labs of the 21st century is the whole earth. In such a situation, how should you structure your work? For the purposes of today’s article, we’ll stick with biology, but one can perhaps imagine similar strategies applying to other 21st century problems (the success of open source software endeavors, such as the Linux operating system, comes to mind).

In biology, people often perceive a dichotomy between basic science, driven by curiosity, but often of unclear application, vs. translational work, driven by a medical or application goal, but sometimes exhibiting high risk (e.g., clinical trials are expensive and have a high failure rate). Both kinds of science, of course, are important. What if, as a third path, one could invent a tool that enables many biologists to go after basic science discoveries, perhaps ground truth-oriented ones (i.e., focused on fundamental building blocks and their interactions), deploying it freely to the community? Those scientists can apply the tool to different kinds of problems, making foundational and mechanistic discoveries. Then, one can design a translationally optimal implementation of a particularly interesting discovery, to solve a long-felt need. This “invent, deploy, discover, design” (or ID³ for short) strategy may be useful, when approaching a complex problem space. It’s a way to form involuntary collaborations, simply because the incentives are aligned - people will work together, even if they never meet, to tackle a problem space systematically.

Let’s consider an example of such an ID³ project: optogenetics. In optogenetics1 (“opto” for light, and “genetics” because the technology is genetically encoded), we control the electrical activity of brain cells, or neurons, with light. By activating the electrical activity of specific neurons, say in a mouse, we can see what behaviors, or pathologies, or therapeutic states, they might cause. By silencing the electrical activity of specific brain cells, we can see what behaviors, or pathologies, or therapeutic states, they are needed for. We achieve this by borrowing proteins from the natural world, that normally convert sunlight into electrical current in single celled microbes. Such microbial rhodopsins (which were discovered starting around 1971, with many distinct classes being reported over the decades following) are used by organisms either to store solar energy in chemical form, or to navigate around bodies of water for optimal photosynthesis.

In the year 2000, one of us (Ed) along with a fellow student decided to try to use these microbial rhodopsins to mediate the optical control of neural electrical activity. In 2005, we published the first report on using a specific microbial rhodopsin from single-celled green algae, to make neurons activatable by pulses of blue light (the invent phase). We disseminated the molecule freely, to thousands of scientific research groups (deploy, with great help from the nonprofit DNA-distributing service Addgene2), and people started applying the molecule to a diversity of scientific problems (discover). Importantly, optogenetics was easy for people to use, which helped the technology spread widely. We, and others, found more and more microbial rhodopsins, with different properties. One group in Europe was trying out optogenetic molecules in the retina, desiring to to treat forms of blindness in which photoreceptor cells die, like retinitis pigmentosa. The core idea was to make cells of the retina which did not die, into artificial photoreceptors, by equipping them with rhodopsins. We helped them figure out which molecule to use (design). In 2021, the European team reported that this molecule could be expressed in the retina of a blind patient with retinitis pigmentosa, and helped him achieve a partial restoration of functional vision. Clinical trials are now being conducted by many companies, of such optogenetic tools, in the context of blindness.

Another collaborative set of groups, in 2009, used an optogenetic molecule to discover neurons that, when activated in the mouse brain, caused brain circuits to undergo electrical oscillations, or brainwaves, preferentially at a gamma (~40 Hz) frequency. In 2016, one of the groups collaborated with our group to find that stimulating such brainwaves in Alzheimer’s model mice (i.e., mice engineered to get Alzheimer’s-like symptoms) caused the brain to clean up its molecular pathology. Later, that group, collaborating with our group and other groups, designed flickering lights and clicking sounds that, because they ran at 40 Hz, could clean up the molecular pathology of Alzheimer’s, and improve cognition, in mouse models (2016-2019; note that optogenetics was no longer needed). Now, human trials are underway with Alzheimer’s patients, with a startup company co-founded by the group leaders (including one of us (Ed)) aiming for FDA approval. Small-scale trials have shown promise in slowing the progression of Alzheimer’s. By following ID³ we facilitated multiple discoveries with practical implications for human health, that we would have probably never solved on our own.

Let us pick a second example - expansion microscopy (ExM)3. In ExM, a preserved biological specimen is chemically permeated by a dense mesh of swellable polymer - basically, the active ingredient in baby diapers. Adding water to such a polymer-embedded specimen causes it to swell by ~10x or more, in linear dimension. Such samples can then be imaged with nanoscale precision on ordinary microscopes (which are otherwise limited in resolution to about 300 nanometers or so, perhaps 100x bigger than a biomolecule). Our lab announced ExM in 2015 (invent), and quickly people started to learn it from our papers, websites, and protocols (deploy). After all, the building blocks of life are nanoscale, and interact over nanoscale distances, so there was much pent-up demand for an easy-to-use nanoimaging technology. At the time of writing of this article, perhaps 800 experimental studies have appeared using ExM - many discoveries have been made with ExM. And, as with optogenetics, ExM is easy to use (one ExM pioneer announced4, “It’s so easy, even my nine-year-old can do it”). Interestingly, as the discoveries flowed, people started modifying and adapting ExM for their own purposes, sometimes working with us. As just a couple examples of this (design) - in 2024, a group announced that they could expand fragments of a protein away from each other, evenly, so that you could see the shape of the protein on a light microscope - quite a surprising outcome. In 2025, another group announced they could break the genome, while still inside a cell, into pieces, expand the pieces away from each other, and then sequence them - enabling nanoscale visualization of both genome sequence and structure, while still inside a cell. Finally, in 2025, two of our group’s alumni, now running their own labs, led the way on a project to expand the whole body - bones and all - by adding in a softening step for bone, before the expansion step began. ExM benefited from the fact that ExM was not only easy for others to apply, but also easy for others to improve - in other words, there was a bit of an “ID³ chain reaction” here, where a design step could also serve as an invent step, triggering another round of innovation.

In short, ID³ might be a good model, for many kinds of scientific problem where you want to achieve a practical outcome, without losing sight of underlying ground truth. One key insight is that ID³ reduces risk by having many people work on the middle discover step, using technology that has been deployed to them. Of course, for this to happen, an invention must be easily applied, and ideally inexpensive. A technology that is not easy to use, may not be amenable to this model. (For those of you who have read the book “The Innovator’s Dilemma,” which discusses how new startups can disrupt industries ruled by old companies, there might be parallels: ID³ might be a sort of scientific cousin of that idea.) Another reason ID³ works well is that collaborating too actively might sometimes be bad for innovation: if experts work too closely together, they might not give a risky idea generated by one of them, a safe space to grow: perhaps some of the other experts might judge it to be infeasible, according to established but inaccurate, or outdated, dogma. Rather, sometimes you need a “skunk works” where a team can work on an idea without being told by others (e.g., upstream inventors, downstream designers), that their idea is stupid. There is a certain kind of strategic ignorance that can help with serendipity - you don’t know that something is impossible, so you give it a try anyway - and because the experts were wrong, you succeed. Finally, ID³ helps increase the number of stakeholders who want to see a technology succeed. They can find many applications of it to different problems. They can support each other, sharing wisdom. By sharing ownership of a technology with many, it can become “too big to fail” and take on a life of its own. From the tool inventor’s point of view - there is an opportunity cost if they stop invention, and instead switch to application. If a tool inventor pursues one application, out of 1000 possibilities, and that application fails - that could result in far less impact than if they simply gave the tool to 1000 people to pursue those 1000 possibilities, some of which may prove revolutionary, even if most fail.

Can we create a system, sort of a decentralized worldwide Bell Labs, to achieve ID³ at scale? Imagine an organization that could help define problems, facilitate invention of tools, orchestrate deployment of technology, nurture discoveries, and help with the design phase - all in a completely globally distributed fashion. Perhaps this organization could identify appropriate people to tackle different tasks, rapidly allocate funding and other resources to them, and upon project completion, allocate credit and reward appropriately. In contrast, current incentive structures are not necessarily a great match for ID³. A group might hold back on publishing a tool, hoping to make a big discovery on their own, rather than deploying it - depriving the world of the tool. The ID³ process can take years, and sustained activity in science is challenging, as any one advance might seem incremental, and thus unappreciated. (Many of the stories above encountered such bumps in the road along the way - optogenetics was a hard sell, and resulted in one of us (Ed) struggling to find a job5, and expansion microscopy struggled to get funding6.) Much of the credit, and/or profit, may go to the people who take the final step, with the initial steps being hard to fund (because they are so remote from the value, which only becomes apparent after the final step is taken), and hard to attribute credit for (for similar reasons).

Perhaps a blockchain-like structure could help - tracking all the value generated, from initial invention, to later deployment and discovery, to final design, routing resources to those who could make contributions, and helping assign rewards to them later - perhaps when their work might have been otherwise forgotten, because the latter steps captured perceived value and obscured key initial steps. If the profits from a successful therapeutic, for example, could flow upstream to those who made enabling discoveries, and the inventors of the tools who empowered those discoveries, perhaps that would align the incentives of biology better. As a special case, right now there might not be much incentive for people to publish negative results. But if someone were to put a negative result on the blockchain, and when someone uses the information later for a successful project, a portion of the reward could flow upstream to the originators of the negative result that helped the later project head in the right direction, that could be beneficial to all. For this to work, access to the blockchain of knowledge would require some kind of enforceable agreement about sharing of future credit and profits. But perhaps someone could figure that out.

In short, involuntary collaboration allows for innovation to flourish, and perhaps there is a way to perform it, consciously and deliberately. There’s an old saying, if you want to go fast, go alone. If you want to go far, go together. Here is a way to do both.

Attending TAMS was life-changing. Indeed, I’m amazed more states don’t make their own versions of TAMS. I got to do chemistry research on the origins of life (working on the hypothesis that DNA was assembled by clay), study advanced mathematics, volunteer to help build a playground, dive deep into philosophical and humanistic questions, and make lifelong friends.

Here is what I wrote (slightly edited to accommodate the transition from spoken to written form; original video here; longer, more science-oriented version with slides, delivered at U Iowa in 2018, posted here):

It is a great honor to be here with you today. 30 years ago, I was sitting where you are now - about to cross the stage, and receive my diploma from TAMS.

Today, I’m a professor at MIT, where I lead a research group inventing tools to understand and repair the brain. The brain is a mystery. We can’t explain how the brain generates our thoughts and feelings. We can’t fully cure any brain disease – which affect over a billion people around the world. The group I lead has developed molecules that let us control brain cells with light, to repair the brain. We’ve used the swellable material in baby diapers, to make brain specimens a thousand times bigger, so we can trace how brain cells form networks.

Our inventions are leading to therapies showing promise in human trials – helping blind people see again, revealing how to clean up junk in the brain in Alzheimer’s disease. We share all our tools freely, helping thousands of scientists confront brain diseases, and to explore the mysteries of thought and feeling. My hope is that we can build computer simulations of the brain, and watch thoughts and feelings as they take shape. Maybe we can confront the nature of consciousness. And, maybe we can end brain diseases for good.

But 30 years ago, as I was graduating from TAMS, the path was murky. There was no textbook, or recipe, telling me what to do. I was fascinated about confronting philosophy through science. I wanted to understand the nature of human existence, and to engineer improvements thereupon. But, when you are confronting a problem that big, where do you even begin?

When you do anything for the first time, there is no recipe, no textbook. You’re at the mercy of luck. I began to wonder – can you increase your luck, through skill? Can you engineer serendipity?

I think you can – and in a post-internet, AI-embracing, world, it might be the most valuable skill you can have. Going forward, following recipes and textbooks just might not be as interesting, anymore. Nor is it up to the task of confronting the intractable problems of our world, the list of which seems to grow in number every day.

Today I’ll tell you three short stories about ways to be lucky on purpose. As a professor, I teach, and I like to think that you can take art forms, like being lucky, and turn them into learnable, teachable skills. Today I hope to persuade you that some forms of luck can be learned and taught.

My first story is about optogenetics, how we control the brain with light. Brain cells compute by firing electrical pulses. In your brain right now, as you hear me speak, certain cells are firing electrical pulses, in language areas of your brain. Brain diseases often involve corrupted electrical pulses. If you could control them, you might be able to repair the brain. But how can you do that?

There have been many attempts to control the brain. Drugs can alter brain activity, but can be slow and messy. Electrical stimulation can help, but electricity spreads in all directions. Can we be more precise? In optogenetics, we deliver light to the brain. And to make brain cells sense light, we borrow molecules from nature. Some microbes have special molecules, that act like little solar panels, and convert sunlight into electricity. We took one of these molecules, put it in a brain cell, and aimed light at it. Amazingly, the cell fired an electrical pulse. Now, this was serendipity. It didn’t have to work. The molecule could have been toxic, and killed the brain cell. Or it might have done nothing at all.

Today, thousands of scientists use this technology, which we call optogenetics – opto for light, and genetics because the molecule is encoded by a small piece of DNA - to study the brain. Using animal models common in neuroscience, people have used optogenetics to find cells that trigger the recall of a memory, cells that when activated can slow the progression of Alzheimer’s, and cells that when activated can improve mood. It’s even showing promise in human trials, as a treatment for blindness – helping people whose eyes have lost their light-sensing cells, to sense light again.

How did we make ourselves so lucky? Well, we used a strategy so powerful, that I use it to this day. Simply put, try to think of every way of solving a problem. This is easier than it sounds! In this case, we made a list of all the forms of energy you can deliver to the brain – there’s light, sound, radio waves, a few other things. You can write the whole list down in a couple minutes. I liked light because it’s faster than anything else, and you can aim it precisely. Next question: how do you make brain cells sense light? Well, you can either design a tiny solar panel, or you can try to find one. That’s the whole list, just two cases. Finding one sounded easier. So we started emailing people, asking anyone who would listen - could you send us the light-driven molecule that you are studying, so we could put it into brain cells? And some people replied. We were in business! We took one molecule, put it in a brain cell, and as I told you, we could activate it with light. By writing down every way of solving a problem, in a systematic way, you can hone in on the best path. You may even find ideas you wouldn’t ordinarily think about. It helps you make a map of your own, when none is given to you.

My second story is about failure. Failure can be useful, if it’s the right kind. It needs to be a constructive failure – a failure you can trust. If you fail because you lacked basic skill, or didn’t try, that may not be helpful. But if you failed because you encountered a fundamental difficulty, you might have found a secret of the universe.

We wanted to map the brain. That way you could figure out where in the brain to intervene, to help somebody. Maybe we could simulate thoughts and feelings in a computer. This is hard - the brain is wired up via tiny, nanoscale connections. The brain itself, of course, is gigantic, relatively speaking. We thought – there are lots of high-resolution imaging methods out there, let’s use them to map the brain. How hard could it be? A year later, we were struggling – the methods we tried were difficult, slow, and expensive. How could you ever map a whole brain?

We were failing. But we trusted our failure – we had learned a secret, which is that nanoimaging, despite all the big announcements, was really hard. And it was hard because of fundamental physical limitations. That means that really new ideas would be needed. This was an example where failure helped us learn the truth. So, we tried another of my favorite serendipity-boosting methods – imagining what would happen, if we did the opposite of everyone else.

For literally 300 years, the way that you imaged in biology, was with some kind of lens. You magnify an image of an object. We wondered – what is the most opposite thing we could do? What if we made the object, itself, bigger? We decided to take a brain - preserved, not living - and infuse it with the kind of material you find in baby diapers – swellable polymers. Do it just right, add water, and the baby diaper material would swell, making the brain bigger. We call this technology expansion microscopy. It allows you to map the brain with cheap imaging devices. And not just the brain – all of biology is made of nanoscale building blocks, or biomolecules, which interact with each other with nanoscale precision. Now anyone can map these building blocks. It’s being used to analyze parasites, in labs in resource-poor environments. It’s being used to see if you can detect cancer in a biopsy, earlier. And of course, we’re heading down the path of using it to make a map of the brain so detailed, that we can simulate it in a computer.

This time, we failed. But you don’t always have to fail yourself - there’s failure all around. Revolutions are often hiding in plain sight - the big idea is out there, but one thing is holding it back. You fix that one thing, and you get the revolution. Facebook wasn’t the first social network. CRISPR wasn’t the first genome editor. Google wasn’t the first search engine. In each case, there were predecessor technologies. And an additional insight, caused the revolution. I call this method “failure rebooting.” Give it a try :)

My last story is about people. After I finished my PhD at Stanford, I applied for faculty jobs. I wanted to start a ground truth-oriented research group focused just on making brain technology. Neurotechnology is a cool, fast-growing field these days. But back then, neurotechnology was not yet cool. The very department at MIT that is my home base today, rejected my job application. Actually, most places I applied to, rejected my job application.

Then luck kicked in. Years before, I had been a teaching assistant for a professor in an interdisciplinary center at MIT, full of designers and multidisciplinary thinkers and creative misfits who didn’t fit into traditional academic buckets. He had an idea for how to teach quantum physics to freshmen, and I volunteered to help. In the lab of another professor there, I did research on quantum computing. These two professors alerted me, during my job search struggles, to a job opening they had, that was going unfilled. They encouraged me to apply, and I got the job. The lesson: if I was helpful to others, I could nurture serendipity in return.

I have countless examples of this kind– finding the optogenetic molecules now in human trials for blindness, getting funding for the expansion microscopy project, being invited to give my first TED talk – all of these resulted from people-serendipity, sometimes after chains of events transpiring over many years. Once I started my lab, I decided that our group would share all our tools as freely as possible – and as people used the tools, and benefited from them, more and more serendipity came my way.

The toughest problems do not come with textbooks or recipes to tell us what to do. We must master serendipity as a skill, to make progress on such problems. I hope I’ve shown you that some forms of luck can be taught and learned.

There’s an old saying, perhaps going back to ancient Rome: fortune favors the bold. But perhaps today we can do better: the bold can engineer fortune.

And so as we part ways today I wish you good luck. But don’t just wait for it – go make it happen. Thank you.

Or if you’ve dropped out of (or avoided) college, and are teaching yourself (the situation that prompted the title of this essay), what knowledge should you seek, and what skills should you practice?

Some people learn a skill early in life and then practice that skill throughout their career. This can be quite fulfilling if one loves the day-to-day act of performing that skill. But one is then often limited to solving problems that are well-addressed by that skillset, meaning that other problems might appear to be out of reach. Conversely, some people are obsessed with a specific problem and are excited about solving it over extended periods of time, perhaps even over a lifespan. But if such a person lacks the skills needed to solve that problem, the path could become frustrating, or even futile.

A hybrid model that may be useful for many people is to consider having a “skill” phase of one’s life, where one learns the fundamental knowledge appropriate to address a problem space of interest, practicing relevant skills as needed, followed by an “impact” phase of one’s life, where one applies the knowledge and skills learned earlier, to that problem space. How should one choose what to learn in the “skill” phase to best solve the problems of the “impact” phase?

Of course, you can jump right into a problem space and learn what’s going on: what the needs are, as perceived by current practitioners within that problem space, and what skills they currently possess to meet those needs. This is a common strategy for reasonably mature fields, such as math or computer science, and often works well. But if the problem space is from a less mature field, such as biology or bioengineering, and thus full of fundamental risk and ambiguity - the kind of problem space that many ambitious and well-intentioned people might seek to enter, and even transform - then there is no guarantee that practitioners currently in the space have the complete set of knowledge and skills needed to fully address the problems of their space. There is also no guarantee that current practitioners are perceiving the needs of their problem space in the most effectively solvable way. In such a situation, one might want to have studied something radically different before entering the problem space, to bring a fresh perspective to the table. In many scientific fields, big revolutions were initiated by outsiders. Is it possible to learn the right kind of knowledge in order to become that outsider, deliberately?

If one wants to bring a radically different perspective to a problem space, then one might be without clear mentors to copy or curricula to follow. In such a case, how does one figure out what to do? One possibility is to look at the problem space, which often relates to a complex system and all the ways it can go wrong, and examine the fundamental building blocks of the system and how they interact. Then, one could focus the “skill” phase of one’s learning on studying knowledge and skills related to those building blocks, and their interactions, so that later, in the “impact” phase, one could apply such knowledge and skills to the higher-level problems at hand.

Alternatively, one could try to break down the building blocks and interactions, as currently perceived by the field, into even finer-scale building blocks and interactions, aiming to reveal unexploited avenues for innovation. For example, suppose one wants to be a biologist or bioengineer, and solve intractable diseases such as those of the brain or aging. If you study such diseases at a phenomenological level, and ignore the relevant building blocks (e.g., biomolecules) and their interactions (e.g., chemistry), you could be missing out on critical mechanisms underlying the disease, or potential ingredients for new tools to confront the disease.

In contrast, if one thinks of the human body, or any biological system, as made of fundamental building blocks, biomolecules, which interact through chemistry according to physical laws, then one might be able to use that lower-level knowledge to support the design of radically novel, highly effective technologies. Think of magnetic resonance imaging, or super-resolution microscopy, or genome sequencing - in each case, a deep understanding of physics or chemistry, applied to biological building blocks considered at the molecular level, was critical for envisioning, realizing, and applying the technology. (The quote, attributed to Einstein but probably apocryphal, that “problems cannot be solved at the level of understanding that created them,” comes to mind.)

One advantage of studying the science of building blocks and their interactions, in the skill phase of your life, is that you are probably studying a science that is more mature than the science you work on in the impact phase. To continue the example from the previous paragraph, physics and chemistry are more mature sciences than biology and bioengineering. The good news is - by studying these more mature sciences, you will be learning things that will likely still be true, and even current, decades later. In contrast, a less-mature, less-fundamental field may have evolved, with new discoveries and inventions, so as to become perhaps unrecognizable relative to its initial state. (This doesn’t mean to avoid majoring in biology or bioengineering, by the way - but one should try to dive deeply into the mechanisms underlying all that one learns, and not always be content with the abstraction layers presented by an expert or class.)

Suppose you are a student in a university - what major should you pick, and which classes should you take? Suppose you are not in a university, or dropped out - what knowledge, online or otherwise, should you learn? In both cases, it can be dangerous to follow a cookbook-style path. For example, simply pursuing the requirements of a major in a university setting may not help you achieve your specific career goals. If you are self-taught or seeking out mentors, then there still remains the challenge of choosing what to learn or who to guide you. The arguments above suggest that it could be helpful if you (or a mentor) could, given a problem space, identify the building blocks (and interactions) to study, and then consider studying the science of those building blocks and interactions. Below, each of the three of us includes some thoughts on our personal path:

One of us (Ed) studied physics, and electrical engineering and computer science, and also quite a bit of chemistry. Now, I lead a neurotechnology group at MIT, aiming to deconstruct the brain into computational processes, that run on top of chemistry, according to the laws of physics - with one goal being to create biologically accurate computer simulations of the brain, and another being to understand consciousness. Understanding biomolecules and their interactions involves physical principles that operate over certain ranges of energy scale, size scale, and temporal scale. That means that knowing general relativity (which helps with understanding black holes) or plasma physics (which helps with understanding the sun) may not be as immediately useful as, say, thermodynamics, or quantum mechanics. During the “skill” phase of my career, I had the luxury (having been an undergrad for 6 years) to learn several fundamental sciences, perhaps to overcompletion. Not all my classes are equally useful to me now, although I enjoyed most of them very much. In the interest of efficiency, at the end of this essay, you can see my suggested “dropout curriculum” to learn how to be a neuroengineer, written up for someone who knows basic physics, math, and chemistry, and wants to know the essentials for broadly operating in the field of neuroengineering. It is a list of classes (or online resources associated with classes) that could be helpful to the quest, stripped to the minimum.

Then I had to pick a problem space to work on, for the “impact” phase of my career. I cofounded MIT’s autonomous submarine team, and a couple months later, we won the first world championship. That was, scientifically speaking, too easy a problem. Another project I worked on at the time, was building a quantum computer. Early quantum computing designs did not scale well; this problem was too hard, scientifically speaking, at the time. My next project was a neuroscience one - and this one was “just right.” The brain may appear complex and messy at first glance, but because of my skillset, it appeared to me to be a computer system running on chemistry according to physics. One core problem was the heterogeneity of the building blocks - unlike physics and chemistry, which involve a small number of building blocks (elementary particles and atoms, respectively), the number of kinds of building block in the brain (biomolecules) is perhaps in the millions. So I set out to build tools to see and control those things. (More on this in a later series of posts.)

Claire, another author, chose to study electrical engineering and computer science to build fundamental skills in physics, math, and the generally applicable field of computer science, rather than taking more special topics classes in neuroscience. I instead learn to apply these skills through hands-on work and research. For example, at E11 Bio, I have been able to understand the intricacies of the problems that connectomics poses. And, through working at various labs at the McGovern Institute, I have learned many detailed microscopy, circuit design, and wet lab techniques that build strong intuition when designing future experiments and goals. However, it is not just neuroscience research; by also working with AI and algorithm researchers, I have been able to apply tools from other fields to novel research problems.

Finally, Nina, another contributor, studies computational neuroscience, with the goal of deeply understanding the fundamentals of math, physics, computation, and their relevance to pursuits within chemistry and biology. I spend most class slots I have focusing on more mature sciences, with occasional neuroscience electives to inspire excitement and learn currently-relevant techniques. I take part in many extracurriculars to better expand my breadth and depth of relevant skills, with occasional impact-focused projects to dip my toes into the problem spaces I’m passionate about (e.g. Alzheimer’s and related neurodegenerative conditions). During my time working with the Tsai Lab in the Picower Institute for Learning and Memory at MIT (they focus on network-level approaches to studying neurological disorders) on cognitive resilience in Alzheimer’s, I have learned both relevant skills (math, bioinformatics, experimental design) and have been able to contribute knowledge to the field as a whole.

Of course, when one pivots from the comfort of the “skill” phase to the ambiguity of the “impact” phase - that can feel difficult, provide enormous stress and frustration, and make one feel truly stupid - even if one is well prepared. Difficult problems are, well, difficult. And pretending a problem is simple, does not make it simple. We will have many essays on how to tackle the art of problem solving, in the future. The “impact” phase often involves breaking down big problems into smaller ones, strategies such as systematic ideation and constructive failure, deep questioning and strategic collaboration, and the engineering of serendipity (which we think is a learnable, teachable thing). There may not be a roadmap for where you are going. But you can learn how to make one of your own. By beginning with skill and then progressing to impact, you will be well equipped with toolboxes to not only plan out your path but to adapt to inevitable surprises and tantalizing opportunities along the way.

Here are classes of the “Dropout Curriculum” for neuroengineering, labeled by MIT class number (as evaluated in 1995-1999, with links to current, closest-match, MIT Open Courseware sites):

Core Classes1

• 5.11 Principles of Chemical Science (posted as 5.111SC)

• 5.12 Organic Chemistry I

• 5.60 Thermodynamics and Kinetics

• 7.012 Introductory Biology

• 8.01 Physics I (posted as 8.01SC)

• 8.02 Physics II

• 8.03 Physics III (posted as 8.03SC)

• 8.04 Quantum Physics I

• 8.05 Quantum Physics II

• 18.01 Single Variable Calculus (posted as 18.01SC)

• 18.02 Multivariable Calculus (posted as 18.02SC)

• 18.03 Differential Equations (posted as 18.03SC)

• 18.06 Linear Algebra (posted as 18.06SC)

• 6.003 Signals and Systems

• 6.042 Math for Computer Science

• 8.044 Statistical Mechanics I

• 6.011 Communication, Control, and Signal Processing

• 9.00 Introduction to Psychology (posted as 9.00SC)

• 9.641J Introduction to Neural Networks

This site will host essays by members, alumni, and friends of the Synthetic Neurobiology group at MIT. (The views and ideas expressed in these essays are those of the named authors, and do not necessarily reflect the opinions or endorsements of all members, alumni, and friends of the Synthetic Neurobiology group at MIT.)

We plan to post essays that touch upon three themes - the engineering of serendipity (how to be lucky, through learnable, teachable strategy), the engineering of understanding (tools for ground truth understanding of complex systems, perhaps including consciousness and existence itself), and the engineering of existence (how to address human suffering and overcome the limits of the human condition).

Looking forward to this new adventure!