Another, Smaller Return

After letting it hang over my head for the last five years, I am finally going back to finish my graduate degree. I was one class (plus my thesis) away from finishing when I accepted my first full-time offer, and I made only marginal progress in the interim.

Given all that’s happened, I cannot in retrospect express any real surprise over that. But now that things have been reasonably stable for awhile—arguably, across the relevant axes, the longest stretch since my engineering career began—taking the time to finish up is reasonable. I’m not thrilled to be back in school again, not the least because it’ll mean significantly less free time. Fortunately there was a relevant course available both online and in the evening, which avoids a lot of the issues that were holding me back from trying to do this sooner.

I’ll still need to finish my thesis at some point, but for now, getting these last few credit hours taken care of is the priority. We’ll see how much homework this class ends up being. Perhaps I’ll have time for thesis work, perhaps not. I never wanted to be a student while also holding down a full-time job, but this is where I ended up. At least I can pay my own tuition this time around.

Believe it or not, I don’t actually like deadlines

If you go through the archive of this blog, you’ll notice that the vast majority of my posts fall in the last few days of the month—usually, the last day of the month. The fact that WordPress displays the post archive on the sidebar by month provided, at the start, a significant motivation to keep the streak running once I finally got started. It’s held for almost a decade, which is longer than almost anything else I’ve done consistently in my life beyond things like brushing my teeth twice a day.

I don’t actually like this approach.

Among other things, it leads to vacuous posts like this, rather than writing something more interesting. But, unfortunately, if I don’t have some sort of deadline, then nothing gets written at all. This is, more than anything, a consequence of our situation of extreme information overload, where optionality about what to consume (and, for that matter, what to produce) is extremely high. Barring some sort of civilization-ending catastrophe or transhuman ascendance, I don’t see this situation significantly changing anytime soon.

I would like to be writing routinely because I want to say something interesting or at least helpful (for me, if no one else). That does not seem to be the behavior I’m actually exhibiting. There have been various mitigating factors, but at this point such conditions seem to be closer to the norm than the exception. Projecting my life forward, I need to be designing a future that can handle such factors if I want to get anything (of any sort) done at all in the timeframe I desire. Again, transhuman ascendance would help here. The fact that my body is going to fall apart and I’m going to die (and that everyone I love and care about is going to have their body fall apart and make then die) is one of the bigger pressing issues facing both me personally and our society as a whole.

See, there’s a post there. But I’m not going to be able to do the subject justice, so I’m not writing it. Not today, and it’s safe to assume not anytime soon.

I am attempting to introduce some small, sustainable changes at present as part of my seasonal theme, but I won’t make any promises about how that will affect writing output. In all likelihood, other concerns will eat up any gains those changes make. The motivation for the Year of Less was, among other things, to focus on what I’m doing (or could be doing) and select a more manageable set that I actually want. Right now, writing this blog is a small time footprint overall, but reaching the levels of output I long desired would require a substantially higher output, and that would necessarily entail not doing something else. Perhaps, unless I want to commit to writing more frequently and deeply, those hours—and mental overhead—could go somewhere else.

I’m not committing to anything now, but definitely still thinking about the question. If I can’t write a post unless the end-of-month deadline is coming up, though, that’s a strong piece of evidence in the “anti-” direction.

Halfway Through

Believe it or not, 2026 is halfway over—49.6%, if we’re abusing the rules of sigfigs. I wouldn’t believe it myself, if I didn’t stop and reflect on everything that’s happened in my life over the last six months.

The halfway point is a natural time to reflect on my yearly theme, but I’ve never explicitly done so before. The last few years, it wouldn’t have made a great deal of sense, but this year, it does, so here we are.

My theme for 2026 is the Year of Less. I chose this because it was both sufficiently meaningful and sufficiently flexible that I could apply it in a lot of different areas (which is the point of annual themes). Unlike resolutions, well-crafted themes basically can’t fail.

That’s the idea, at least. So far, the Year of Less sure looks like a failure. Rereading my post from January, I haven’t really tackled many of the things I was thinking about at the time, and instead added—temporarily or permanently—a number of new things to my life. I certainly don’t feel like my life is “less” right now.

Such things always make sense at the time. Several were effectively outside of my control. Some, I had simply planned to do should the opportunity presented itself, and then the opportunities presented themselves. A few others, I had already agreed to before the year started, and simply had to continue on with. Looking back, it’s hard to identify very many specific decisions where I realistically would (or could) have taken another option.

The halfway point is also the natural time to decide whether to continue with a theme, or pick a new one. It might seem that I’d be keen to try something else. But really, the underlying motivation is stronger than ever. I still think paring down my life a bit would be a good idea.

Indeed, I used my theme to make a significant decision to that end in the last week. I had been vaguely planning to do Camp NaNoWriMo this year, but thinking through what it would entail, I came to the inexorable conclusion that I shouldn’t bother. The time commitment is quite significant, and I have other projects which I can’t afford to defer. Thinking about it in the context of the Year of Less made it significantly easier to accept this decision, which is exactly what a theme is supposed to do.

And, of course, time is not the only part of Less. It is most salient right now, perhaps because its the area where I feel like things are going the worst. Some of the other aspects I discussed in January are going well enough. I wouldn’t say I’m knocking any of them out of the park, but the financial and health items I mentioned seem to be progressing acceptably.

I still haven’t decided where the blog fits into all of this—indeed, before reviewing the January post last night, I’d forgotten that I’d even been thinking about it. Perhaps that’s the sort of thing to revisit at the end of the year. December will mark a full decade of posting at least once a month. If I decide to call it quits, that’ll be a nice round number to end it on.

Most Humans Probably Don’t Exist Yet

Yesterday, I led a discussion about longtermism at the Houston Effective Altruism meetup. My preparation felt a little half-baked compared to where I wanted to be, primarily because laptop shopping during my week off ending up taking even more time than expected.

One idea that I really didn’t get to is that, when projecting the future population of humanity, fairly reasonable assumptions imply that the vast, vast majority of humans haven’t been born yet. While I don’t think glossing over this assumption necessarily invalidates any of the takeaways from our discussion, it is definitely one of the motivation notions behind the longtermist school of thought.

In Will MacAskill’s book What We Owe The Future, he lays out a calculation that finds over 99.9% of humans could live in the future. The calculation assumes that Homo sapiens last as long as a normal mammalian species (approximately 1 million years), and the human population stabilizes around its current level (8 billion), and human lifespans don’t dramatically change. Since we’ve been around about 300,000 years—during which time about 109 billion humans have lived and died—that implies we have about 700,000 years to go. If the human population of 8 billion continues for that long, then something like 80 trillion people will live in the future.

Whether or not we include the present and past population of Earth in that figure doesn’t really matter. 117 billion is less than 0.15% of 80 trillion.

You can play with the assumptions can get different numbers. Our World in Data did a similar calculation, assuming the human population stabilizes at 11 billion and lasts 800,000 years, and comes up with 100 trillion future humans. The basic logic remains the same: if humanity does not go extinct relatively soon¹, and the population growth of the last few centuries does not dramatically reverse², then we can reasonably assume that most people who will exist do not already exist.

This has significant implications for how we think about our place in the story of humanity. There are good reasons to suspect that the decisions made during this era will be unusually impactful compared to others in history, to the extent that some people argue plausibly that we are living in the most important century.

While any given person may not have all that great of an impact on the course of history, basically no individual could have significant impact during the long hunter-gatherer period, or perhaps even during the pre-civilizational agricultural period. Compare to the impacts which individuals had during the Industrial Revolution (and since), for better or for worse. Now, you can argue that if Thomas Midgley Jr. hadn’t developed tetraethyllead gasoline and chlorofluorocarbon refrigerants, someone else would have. Perhaps so! The fact that any one individual could develop two such things causing massive—negative—environmental impact on a global scale and within a normal human lifetime tells us that something is very different about this period compared to earlier ones.

Longtermism emphasizes the impact, then, which our decisions today will have upon those future people. Now, whether we can predict and choose what impact we should have, whether we can actually achieve an impact, and whether we should be attempting to have an impact, are all subject to vociferous debate. You can probably guess my point of view simply by knowing that I volunteered to lead a discussion on the subject, but I think that any reasonably debate about our relationship to future people needs to begin by acknowledging that there’s going to be a whole lot of them.

---

¹It is not a given that humanity will last that long, of course. Some of the longtermist’s major projects are in the field of existential risk reduction, to ensure that that many people get the opportunity to exist in the first place. I think this is also where I’m supposed to mention Brandon Carter’s Doomsday Argument, which asserts that us existing so early in the distribution of possible humans is unlikely given a random sample. I don’t find this argument particularly convincing, simply because somebody has to be first, and what does it even mean to be “first” in the context of continuous hominid evolution?

²Which could very well happen! Recent data suggests that we’re much closer to the “low fertility” scenarios than the Soylent Green-type catastrophes many were imagining fifty years ago. The exact trajectories remain to be seen, but my personal prediction is that low birth rates will be a much bigger social problem in the the 21st century than overpopulation. These will be modulated by the exact impacts of AI replacing human workers and longevity research, of course.

Beware Planning Fallacy

Epistemic Status: Search your feelings, you know it to be true.

I’ve had planning fallacy on the mind recently, for a combination of both personal and professional reasons. This cognitive bias has been a focus of mine, again for both personal and professional reasons, for many years. It’s perhaps ironic, then, that I experienced it so greatly, in multiple areas, between the time I decided to write about it and the time I actually sat down to write. And that motivates me to write about it even more.

Planning fallacy usually refers to the consistency tendency of humans to underestimate how long it will take to complete a task. The term is also sometimes applied to underestimates of cost and risk, because those often see similar failures of estimation. I suspect that all my readers (yes, both of you) can think up multiple times you underestimated how long something will take, despite having all the relevant knowledge and experience to make a more accurate prediction.

Interestingly, planning fallacy is mostly an egocentric bias. We typically do better when looking at the timelines for others’ tasks, even overestimating how long such things will take.

Various explanations have been proposed, but I’m less confident that any particular explanation survived the replication crisis than the concept of planning fallacy itself. A number of the proposed explanations relate specifically to project management, which is of course a major motivation for my studying the phenomenon, but which is also heavily dependent on the organizational context, and thus less generalizable.

On the other hand, there are a few basic strategies for combating the planning fallacy, and these are applicable in most contexts.

Implementation Intentions are concrete plans for action, focused on the where, when, and how. All too often, we have a clear answer for who, what, and why, but neglect the other questions. If, for instance, you want to write a novel, you need to decide where the actually writing is going to take place (your house, a coffee shop, on your patio, etc.), and when (in the morning, late at night, only on weekends, etc.), and how (in a Microsoft Word document, on a typewriter, with quill & parchment, etc.).

There’s some evidence, though I haven’t checked whether these studies are still replicating, that simply thinking things through explicitly helps one commit to actually completing the task. (It’s easier to get up and go if you’ve already planned everything out.) Additionally, this explicit breakdown leads to the second major technique…

Segmentation refers to explicitly breaking down the task into subtasks, and estimating the time (or cost) of these subtasks before adding them up. Humans are notorious for giving a top-level estimate significantly smaller than the sum of its parts. This is less a matter of fudging, though that certainly happens, especially in contexts with pressure from above. Rather, we simply come up with a more optimistic overall answer than when we stop and think through everything that the larger goal entails.

This assumes, of course, that we have a clear sense of what said goal entails. Sometimes, starting out, that’s not the case. Either we don’t know the breakdown, or we don’t know how long some of the individual components will take. (For large and complicated projects, it’s probably some of both.) In this case, we can use reference class forecasting to get a starting value to iterate around.

We use this a lot in engineering design, and I realized the application for combating planning fallacy years before I formally encountered the term. Take the relevant figure-of-merit (usually time or cost, in project management) and check what the number ended up being for a prior comparable project.

If a comparable project took two years, maybe revise up your estimate of six months. Maybe there’s a good reason to believe you can get it done faster—but maybe you’re full of it. Unless you can very clearly spell out what you’re going to do different and how it will make things better, basing your estimates on past performance is safer.

For a concrete example, when my parents decided to move, they could have used the recent experience they’d had cleaning out my grandmother’s house as a starting point. That took them nine (9) months, with considerable assistance from me. Moving to a new house is a different but somewhat comparable experience, so use nine months as a starting point. But then consider the two storage units of stuff from my grandmother’s house that wouldn’t fit in our old house, and that I was attending college full-time in Lawrence instead of living at home, and it’s not easy to see that it could end up taking quite a bit longer that that.

In the end, it took 26 months—far longer than my parents remotely expected when they chose to undertake the project. The last seven months featured full-time contributions from me, because I graduated college and moved back home. I was never able to prove it, but dumping all of my breaks during Junior and Senior year into unpaid household labor (instead of, you know, resting after finals) probably had something to do with my abject lack of job offers on graduation.

If I sound bitter, it’s because I am. I predicted this would go poorly, that no lessons had been learned from the preceding comparable operation, and that I would have a miserable time for no personal benefit. My concerns were dismissed, and then borne out even worse than I’d imagined. Applying some simple reference class forecasting, using the previous move as the starting point, would have gotten them a lot closer to the actual ~3x factor than the wildly optimistic vague intentions they used to make a major life decision.

Now, sometimes it’s not clear what the relevant reference class actually is. This is a legitimate limitation to reference class forecasting, especially when you’re trying to do something new. Cleaning out a dead person’s house is not exactly comparable to moving your own house, but it clearly in the same general ballpark. But if you’re trying to do something almost entirely new, then you have a bigger challenge. When trying to scope out a Mars mission architecture, for instance, there’s not really a directly comparable project to work off. We can learn lessons from previous infrastructure and aerospace projects, but they aren’t exactly the same thing.

That’s no excuse for throwing up our hands and pulling numbers out of thin air. Setting arbitrary targets is a recipe for disappointment. Instead, we use these techniques to come up with our best estimate based on the information available, however incomplete, and build in a healthy margin (especially when succeeding on-time is important). Your idea of “a healthy margin” will almost certainly be too small, as I learned the hard way this last weekend. Make it bigger.

In a professional context, upper management won’t like it. But they’ll like it less if they get fired for being late. As Akin’s 23^rd Law of Spacecraft Design states, “[t]he schedule you develop will seem like a complete work of fiction up until the time your customer fires you for not meeting it.”

The application of these principle is left as an exercise to the reader.

A Blg st Wtht Jst 4 Letters

The keybard n ths lat s startng t g. S far t’s jst art f ne rw, and even then t smetmes wrks and smetmes desn’t. Bt tnght s ne f the days when t desn’t wrk, s y’re gettng ths nstead f a rre st. Gd lck decherng t.

The Inferential Distances Really Are Long

Expecting Short Inferential Distances is an underrated classic of Yudkowksy’s Sequences. It’s one of many ideas which I first encountered in the Sequences that significantly shifted my perspective on living in a human society. If it’s overlooked by other readers, I rather suspect it’s because similarly-powerful insights are common in the Sequences (especially compared to where the culture was when they were written).

Yudkowsky points out that for most of human history, most of the people you would have interacted with did not have a large degree of specialized knowledge. Almost everything they knew was the same stuff you knew, at least once you’d gotten old enough to absorb all of it. It was, in all likelihood, quite rare to find yourself in a situation where you were two or more steps removed from the general knowledge which you shared with everyone around you.

That’s not to say that there wasn’t ever specialized knowledge, far from it. Especially as societies became more complex, there were kings and generals and priests and tradesmen and so on who knew things that the farmers they lorded over did not. But even during those later periods—the hunter-gatherer phase of human existence lasted far longer than the agricultural phase—the vast majority of the population consisted of those farmers, not the lords.

I’ve run in to this a couple of times in the last week, but I’m only going to talk about the fun example.

As Yudkowsky elaborates, the expectation of short inferential distances poses serious problems for science communicators (and, for that matter, any other form of specialist trying to get a point across to a non-specialist). It is genuinely hard to put themselves back in a position of ignorance, to abandon all the assumptions and ideas and facts that they’ve accumulated in the process of becoming a specialist.

Or, as Randall Munroe put it:

Now, I can relate to this experience viscerally, because I make a living working as an aerospace engineer. “Rocket science” is perhaps one of the most pervasive cultural stereotypes of esoteric knowledge. I find this amusing for multiple reasons, including the fact that there are plenty of types of science and engineering that are far more intellectually demanding. (Brain surgeons, I suspect, really are Built Different™.) I maintain that our field is coasting on a reputation earned back when most American adults didn’t have a high school diploma.

But the effect, I have to admit, is real. I’ve experienced it many times, talking to intelligent and well-educated adults, even ones who are interested in science and space exploration. There are things which I figure are common knowledge, at least among that subset of the population, and discover are emphatically not.

The local rationalist group does occasional PowerPoint parties (we have a better name for them, naturally), to which I’ve brought various presentations on aerospace topics. A few recent presentations I’ve done on spacecraft propulsion have involved extensive question-and-answer sessions on background material, because I vaulted into specialized topics and didn’t realize that my friends didn’t know about some of these things.

At this point, I’ve decided bring a more basic topic on rocket propulsion each month, until I finally get to the point where I’m no longer talking above their level of background knowledge on the topic. This is to my benefit in multiple ways. First, of course, is doing the exercise, taking the advice Yudkowsky gives of backing up multiple steps back when running into communication difficulties. Second, it’ll make these presentations easier in the future, if more of my friends do know the basics of the subject, I can talk more freely on more involved topics. (They’re going to come to me with their questions about space news anyway, so having more background information speeds things up, and means I can answer their questions in a more comprehensible manner.) Third, it’s a good idea to revisit the fundamentals from time to time. Just because I aced the class in undergrad doesn’t mean I understand the topic well enough to apply it now. The occasional technical refresher is a good idea for any specialist, to ensure their own understanding is sharp.

Our next get-together is tomorrow, so I’d better leave it at that and get back to working on my presentation.

My Theme for 2026

Usually I’d do my annual theme post at the beginning of January, not the end. This is actually the latest I’ve ever done it—with the exception of 2025, when I never manged to write the theme post at all.

Last year, my theme was health. This didn’t really feel like a choice on my part, because nothing in my life a year ago really felt like a choice for me. My mother’s health problems had dominated my life since June of 2024, and would clearly continue to dictate my high-level decisions for some time going forward. But health is a classic example of a yearly theme because it is quite flexible. It can apply to yourself, or to those around you, and can apply to mental or financial as well as physical health. Given the situation at the time, it was about as close as I could get to a failure-proof theme. Despite considerable frustrations, by the middle of the year, my mother’s health had improved to the point where she could live by herself (with some assistance). In December, I was able to take her to attend her father’s 100th birthday party, something that would have been impractical (if not impossible) six or nine or probably twelve months earlier.

The purpose of themes, as noted proponent CGP Grey says, is to build a life that you want to live. The last couple of years have been so far from that that it’s hard to fully articulate. There has been some recent progress, but after getting whacked so many times in a row, I’m naturally quite hesitant to make any sort of plans under the assumption that things won’t fall to pieces in the immediate future.

I therefore want to capitalize on my recent freedom of action so that, come what may, I will be better positioned for whatever events occur, and my life overall will be better than it otherwise would be. Unfortunately, this made picking a theme for 2026 a bit more challenging than one might expect. A lot of the ideas which resonated with me weren’t particularly good themes, in part because they tended to be too specific. The two particular ideas which I considered the longest were downside mitigation and attack surface reduction. Those names might give you a clearer picture of my broader mental state.

More recently, after the trip in December and other events in the last few weeks of the year, I considered more conventional names like resolution or consolidation, but both of those still felt like they depended a bit too much on particular conditions and outcomes which I could hardly guarantee.

Themes are better than New Year’s resolutions precisely because a year is just too long of a period of time to set realistic goals over. It’s extremely difficult for an individual to construct SMART goals in January that will even make sense in December. Instead of dictating a state you wish to achieve, a theme can provide guidance when making decisions all through the year. A well-crafted theme can help you make decisions across many axes of your life that push your behavior and situation in a consistent direction.

I’m not sure it rises quite to the level of “well-crafted”, but my theme for 2026 is the Year of Less.

Less can, and does, refer to a lot of different things here. I have a lot of different goals, intentions, and desires—of various shapes and sizes—which can be reasonably described as having less.

This includes my own goal of maintaining a reasonable state and size, by having a little less weight (or at least gaining less than I otherwise might).

It also includes things like less debt (at levels, I should be able to pay off both my car loan and student loans this year), and less backlog (I’d like to make significant progress on my read/watch lists). But those aren’t really the idea I’m aiming for.

Rather, I’m trying to cut out the cruft, to get my life a bit closer to the theoretical minimum, so I can focus my energy and attention on the fundamentals. I’m aiming for less distraction, less dispersion of time and effort. Generally speaking, our culture seems to demand more, more, more from us. Every request is for more, to increase the amount of stuff we’re doing and thinking about. Heck, most New Year’s resolutions are structured as something you want to do more of (or some enjoyable thing you want to do less of). The usual result is spending all your time and energy on a myriad of tasks and activities to ultimately accomplish nothing. Bringing more resources to bear on a more limited number of tasks will almost invariably yield better results.

There needs to be some degree of diversification—to use an extreme example, you won’t be able to keep working if you don’t take the time to eat and sleep and otherwise look after your physical health. Recreation and so on are also valuable for your mental health, even before we consider the potential cross-pollination value to any sort of creative work.

There are, undeniably, many of things which I want to be doing more. Unfortunately, I have no more hours in the day that I did last year. To do more of one thing, I will necessarily have to do less of something else. This is what I mean by trimming my life down to the the essentials. I don’t expect to get to the true minimum, nowhere close, but I think having the Year of Less as a theme can help guide those decisions, to make choices that reduce rather than increase my obligations.

This means, of course, saying “no” to more things, especially marginally valuable things—that is, things which are “on the fence” of being worth pursuing. This will sometimes be uncomfortable, telling people I don’t want to do something they’re proposing, and dropping or reducing my engagement with things which I’d previous been involved with. That won’t necessarily be easy, but having this be my explicit theme for 2026 will probably make it easier than it otherwise would be, especially as the year goes on and I do it more often.

Now, for the obvious question: will I stop writing this blog? Is that something I should cut out, since I don’t seem to be really going anywhere with it? I can’t say for certain. Maybe ten years is long enough, and it’s time to call it quits. Maybe I’ll finally start writing more than one post a month (usually on the last day or two of the month). I won’t commit either way. Generally speaking, themes are scalar, not Boolean. Perhaps, if I went really, really hard on the Year of Less, I’d decide to delete every social media account I’ve ever had and become an Engineering Hermit™. That doesn’t seem likely, and almost certainly wouldn’t be the optimal balance for achieving the full suite of my life goals. If anything, writing more is more likely, but continuing about the same way I have been is by far the most probably outcome.

The future is uncertain. That’s one of the core observations motivating the idea of annual themes in the first place. So, we’ll just have to wait and see.

My Theme for Winter 2025-2026

I’ve discussed the idea of yearly themes on this blog for…actually, most of the time I’ve been writing it. For much of that time, I’ve been interested in the idea of seasonal themes, in addition to the annual kind, but never really had a good one in mind. Part of this has been my tendency to tie the themes too closely to particular goals, which is one of the reasons I’ve managed to fail themes, despite the whole point of a theme being that it can’t be failed the way a New Year’s Resolution can be failed (and usually will).

I’ve found an appropriate theme for the current season, however, so for the last couple of weeks, I’ve been trying out the Winter of Now.

The basic idea with this theme is to procrastinate less, to do things now instead of later. This includes both things that aren’t necessarily fun but need to be done (for instance, maybe I could get my taxes done early for a change), as well as things that are enjoyable yet I put off out of habit. This latter is perhaps even more important than the former.

See, the last couple of years, with all the family health problems, have only solidified earlier habits of mine to put things off till later. In some contexts, this is unavoidable. But, after something like a decade of being either in an active crisis or an interminable project—where my time is not actually mine to dispose of—it becomes reflexive. I’ll read it later. I’ll watch it later. I’ll play it later. I’ll write it later. I’ll do it later. At a certain point, the backlog becomes so tremendous that I have to ask myself if this mythical “later” is ever actually going to arrive?

Certain things have to wait at some times. But I’ve gotten to the point where I’m procrastinating things I want to do, or know I will find rewarding on the net, just because that’s what I always do. I’m hoping to at least modestly reduce those tendencies over the next couple of months. I’ve seen some success already, and think there’s room for several more relatively easy wins, just by getting myself out of this temporizing mindset.

We’ll see how it goes. I’m still finalizing my theme for 2026, but will have a post about that shortly after the start of the new year.

Where do we go from here?

Three years ago, on the 30th of November 2022, OpenAI released GPT-3.5, better known as the first implementation of ChatGPT.

Now, the potential of the Generative Pre-trained Transformer (GPT) models had been clear for several years before that point. Scott Alexander specifically wrote a blog post called GPT-2 As Step Towards General Intelligence in February 2019. At that time, the outputs were barely coherent compared to what Large Language Models (LLMs) are doing today. To those of us who remember what algorithmically-generated text was doing before, however, it was quite an impressive leap.

There were some online curiosities circulating in my general sphere around that time which, in today’s parlance, we would call “a wrapper on ChatGPT”, but ChatGPT didn’t exist back then. It was hard to get access to GPT-2, as opposed to today, where everyone’s shoving a dozen different LLMs in your face in every product connected to the Internet. Sending messages to an automated Tumblr account simulating a real Tumblr user was one of the better ways to try out the model.

ChatGPT was, undoubtedly, a paradigm shift along many axes.

Within my specific subset of the rationalist community, the conversation about AI risk changed noticeably. While nobody I know personally was laboring under the misapprehension that GPT-3.5 was an Artificial General Intelligence (AGI), it was clear that this was a major step up in terms of what computers could do, a lot sooner than many of the skeptics had predicted. While it remains unclear how far we can scale LLMs alone, it was clear that you could do things with these sorts of models that many people previously hadn’t thought possible. Image generation was also improving drastically during this period; comparing AI images generated in 2023 with those generated at the same time in 2022 is a night-and-day difference.

We spent a couple months at the local rationalist meetup frequently discussing the implications of all thing. Jokes along the line of “no FOOM, we’re still here” came up multiple times; we didn’t think ChatGPT was going to FOOM, of course, but it was a lot closer to something that could conceivably FOOM than anything we’d seen up to that point. At a minimum, I contended, something like an LLM could certainly be a component in a more general AI system.

Now, as people who were aware of Yudkowsky’s arguments, several of us had already noticed the cognitive dissonance of planning for the longer-term future when it’s not clear that humans are going to be in the driver’s seat much longer. The human spaceflight program I work on is planned for the late 2020s and well into the 2030s; others had work experiences which similarly looked well into the future. With fast-takeoff scenarios suddenly feeling plausible, the dissonance only grew.

The next major model increment, GPT-4, debuted in 14 March 2023. But that’s not why my life changed forever that month. That same week, my father had a stroke. That would have been bad enough on its own, but which he was in the hospital he was also diagnosed with cancer. His persistent cough, it turned out, was caused by carcinoma of the lung, and had almost certainly caused the blood thickening that created a clot in his brain.

The second quarter of 2023 was when the cultural conversation on AI finally came to fruition, and I had to almost entirely sit it out because my father’s precarious health took priority. The FLI open letter and Yudkowsky’s Op-Ed were both published while he was still in the hospital.

I can’t say whether I would have been able to have even a minuscule impact on the conversation if that hadn’t been the case, just as I can’t say whether we would have had better outcomes fighting the cancer if it weren’t on top of stroke damage. But those are the sorts of things I’m always wondering, and the chronology is such that I find these questions tied together in my mind.

This is only a fraction of the unexpected disruption I’ve experienced in the last three years. The health of my parents (my mother was probably already getting sick when my father was still alive) has, obviously, dominated the impact for me personally. But the wider perspective, of technological and political change, hasn’t made things feel any more sane for me.

I don’t know what my life will look like in a year, because for each of the last four years (at least), I would have been wildly wrong making that prediction. This is largely because of family health crises, but far from exclusively. When I start projecting out where AI technology, in particular, is going, it becomes very difficult to conceive of what will happen. I am reasonably certain that there will be a November 2026, though I don’t know what it will really look like. I’m considerably less confident about November 2036.

The manufactured normalcy field is important here. If you told someone from 2015 that people would be talking to AIs all the time and self-driving taxis were rolling out in many major metropolitan areas, they would imagine a much more “futuristic” scenario than the one we’re actually living in. Some of that is because it’s basically illegal to update the built environment to match the century we’re in (as I’ve long complained), but some of it is that we just get used to things remarkably fast.

What comes next? Who knows. I can try to make projections, but there’s too many surprises to say concretely. Breakthroughs do happen, people do make mistakes. The best model I’ve developed for the behavior of intelligent adults—at least, those untrained in rationality—is that they are positively addicted to stepping on rakes. If we even slightly generalize this tendency, it suggests that AI development and alignment will not go well. And so far, it hasn’t!

I don’t know, not at the level we’d like to know, how my mother’s health will evolve. I don’t know how my health will evolve. I don’t know how Congress and the White House will change the funding for human spaceflight programs. And I sure don’t know how AI will develop. The last three years were a very strange journey. I fear the next three years will be even stranger, but don’t dare venture to hypothesize how.

I’m so tired of being in limbo. I never once chose this. When it began, it was just personal. But now, with the political and technological aspects so salient, even when the personal aspects retreat for a time, the uncertainty barely abates. Now, to be fair, I don’t trust problems to stay solved anymore, because everyone kept stepping on the aforementioned rakes and I don’t doubt they’ll start up again given the opportunity. But even setting that aside, I never expected the eschatology of the robot cult a 14-year-old cyberbullied me into joining would actually show signs of materializing! How do you respond to that?

How do you plan for the future when it’s no longer clear that there’s going to be a future?

If I figure it out before we’re broken down for our constituent atoms, I’ll let you know.