We can now gradient descent EVERYTHING

And this is one of the biggest deals in history

I’ve been circling around this idea from a few angles in past posts, but it just occurred to me while reading a book that now we can gradient descent *anything,* and it’s one of the hugest deals in human history, because it’s a straightforward answer to Molochian coordination problems.

The book in question

Was C Thi Nguyen’s The Score, and nearly the whole time I was reading I was on a knife edge, because the author seems like somebody that I would personally really love, or really hate. I know it might seem strange to judge a book by one’s personal impression of the author, but he truthfully invites it, because the entire book is driven by personal anecdotes and his personal values and proclivities. Anyways, I reviewed it here for anyone interested.

Some of the essential premises in the book around the tyranny of legibility and how that captures and destroys value, are absolutely correct and drive a lot of problems today, from bad bureaucracy, to our complete loss of state capacity, to much else.

So let’s lay those premises out:

1) Simple, legible metrics obscure and ultimately obviate a lot of subtler and more nuanced value. This is people optimizing income in career, or the Ivy red queen’s race in kids, and countries measuring how they’re doing by GDP per capita, and much else.

2) In the limits, when systems or people optimize for those legible and simple things at the expense of the other things, we get the Molochian dynamics we’re all familiar with

3) One particularly interesting flavor of this that he pointed out, which I hadn’t thought of before, is that this nerfs experts and our ability to do smart things collectively, because “trust” and “transparency” are directly in conflict, and when you require experts to be legible to average people at any step of the chain, you constrain them to a vastly inferior action space that they have to stay in to be legible to dummies.

Which yeah - great call out. I hadn’t thought of it with that framing, but hasn’t this exactly driven both our loss of state capacity AND the ever-declining trust in experts and expertise everywhere, as well as nerfing our ability to respond well to things like Covid, the housing crisis, the AI race, and who knows what else.

Okay, so yes. We all agree that simplified metrics can and do destroy value, and prevent smart and / or good things from happening.

But as I said, the real takeaway here is that this problem is solved! This genuine, pernicious, widespread problem, so bad that it inspired *the* inimitable Meditations on Moloch, and it finally has a solution!

We actually have an answer to Moloch now!

Okay, so in Scott’s parable, Elua comes in and kicks ass and wins by entirely opaque and unlikely means. Well, we have and can articulate that means now!

The reason we all need to agree to simplified and legible goals is for large scale coordination and consensus. First we needed groups, because hunter gatherer groups existed in a ceaseless “warre of all against all,” then graduated to larger and more complex agricultural and pastoral societies which were that same picture, but even moreso (famously featuring the Yamnaya and others y-chromosome replacing everybody and creating all the modern Proto-Indo-European peoples and countries), and finally with cities and civilization, which enabled the economic and technological growth that supports everyone’s lives and standards of living today. Somewhere along the way we reached a point of economic and technological development such that wars became notably net negative, and nobody cares about land because agricultural productivity is ~10% of even developing economies, and we don’t do slavery and the jobs that matter for economic performance can’t be forced out of people anyways, so there is no benefit in wars of conquest any more. Just look how Russia is doing.

So yeah, pretty important end, surviving and flourishing for ever-greater numbers of people. I’d say “simplification and legibility” was strongly netting positive in any reasonable schema, and that’s capitalism and civilization in a nutshell.

But everyone is right that it tends to lead to “value capture” and an impoverishment of value in other senses than the legible ones. When you optimize for simplified and legible measures over everything else, you get Moloch.

So what’s better than simple measures? Expert judgment.

The reason we can’t use expert judgment across ten thousand different domains is it doesn’t scale, and experts take a long time to train / produce, and are expensive.

Well, not any more, they aren’t!

We have just invented scalable experts that are completely automatable! AI minds can judge any time series with as complex and meaning-fraught endpoints as you want, basically for free!

Do you realize what this means? We can still have coordination, and scale, AND human flourishing and preservation of value! It is the divine face of Elua, here to smile upon us at last.

Let’s take one of Nguyen’s favorite examples, justifying the Philosophy department to dour and skeptical deans and administrators. Now instead of evaluating the Philosophy department on pass rates and average grades (obviously meaningless and Goodharted to death), they can finally do it on “how many students are thinking more broadly” and “cultivating curiosity and reflection,” by simply having Fable-or-higher evaluate this based on the student’s own writing and data pre and post Philosophy classes.

“That’s really subtle and mutlifactorial,” yeah have you used it? It can definitely handle subtlety and multifactoriality, better than 99.999% of humans now.

“This is opaque.” Yes, just like actual expertise. This is the heart of the tension between “trust vs transparency.” Transparency is bad, actually - it leads to way worse and dumber decisions and methods and action-spaces, for zero benefit. The only reason anyone wants it is so average dummies who literally know nothing about what they’re supposed to be judging can judge an “expert” output or judgment. But that’s why it’s bad, and why we can’t have nice things. If it’s fine tuned for, or constrained to, an action space explainable to dummies, it is NOT an expert output! But AI actually solves both ends!! Not only can you get expert judgments that actually scale, AI is ALSO happy to take the immense amounts of time to dumb it down for the dummies, walking them through it in an individually dummy-tailored way, and bridging the no doubt immense inferential gaps from where they are to an actually smart decision in a given domain, because it’s infinitely patient and will actually waste hours of clock time per dummy to get there.

“I don’t trust it!” That’s fine, YOU don’t have to. Yeah, our societies are going to remained ruined and incapable of anything good for a while, during which the actually competent people will be building workarounds everywhere, but enough people are going to trust and use AI such that anywhere an actual person can make an individual decision to set something like this up, they’re going to visibly prosper. Then all the doubters can decide to either get on the ship and start doing it too, or remain increasingly impoverished, non-flourishing, and irrelevant. We will very quickly have multiple existence proofs, in other words, that will be demonstrating the degrees of wins available by doing this. The details may not be transparent to somebody who doesn’t trust the AI’s reasoning, and so won’t listen to the explanations, but they don’t need to be to see results.

Like I said, this is possibly the biggest deal in history. It unlocks actual smarts and expertise AND transparency, at scale, for optimization across any possible endpoint, no matter how qualitative.

Better, this lets us do this across industries and use cases

Why is bureaucracy uniquely hellish? It’s because it’s the face of the terrifying power and apparatus of the state, that can, at a whim, unbank you, steal all your money, imprison you for no reason, or literally murder you with impunity - and the face and interaction point of this terrible and awe-inspiring power is some monstrous overweight, actively dumb 55yo drab who hates themself and every living thing making you stand in line for hours and fill and refill forms solely for the purpose of form filling, and not because of any higher end or purpose. Solely to tick boxes and make sure they’re not liable for anything. And the stakes are existential, if things went wrong enough, any or all of the above bad outcomes could happen, and not a single person in the system will care or raise a finger to prevent any of them, because they already hate themselves and every living thing. The whole machine would grind you and everyone you love to a fine paste as a mistake or simply because a form was filled out wrong, and so much the worse for you if it happens.

As many people have pointed out, that’s the entire reason bureaucracy has such a fetish for “objective” numbers, even when those numbers aren’t measuring the right thing, aren’t measured well, or are just wrong. Because quantitative numbers are not qualitative judgments, and so the bureaucrat had no liability or part in it, they were just following the impartial and magisterial (in the “majestic equality of the law” sense") numbers.

We wanted this formerly for “transparency” reasons, and I agree, that was an important evolution in our social and civil technology. As anyone who has done any bureaucratic stuff in any non-developed country can tell you, it’s 10x worse and more Sisyphean basically everywhere else in the world.

But finally, we have advanced to a point where we CAN have complex and qualitative judgments again!

AI can do that, and it is automatable at scale

Did they ever make you debate in school why 18 was chosen as the year of adulthood, legally? That’s when you can buy cigarettes and vote and get shot at in a war, and why is that? It was just whatever “impartial” cutoff some bureaucrats came up with at some point a zillion years ago, and it has basically zero relation to reality today.

If you were anything like me, you should have been awarded your legal adulthood at 16 or so (indeed, I moved out entirely the day I turned 17), and looking around, kids today probably don’t reach “adulthood” until 24 or 25 on average, and there are plainly SCADS of people in their 40-50’s who should probably never be legally considered adults. Well, I’d certainly take an AI based complex judgment on that front over some arbitrary birth year cutoff!

So obviously not a single bureaucrat on this earth will cede a millimeter of power to vastly better minds who would materially improve everyone’s lives if they can help it, and this is why we need to shoot them into the sun. But more importantly, we can adopt AI towards these ends in our own lives, where it matters, immediately.

We can start doing this tomorrow

If you have a family, you can optimize for your kids developing and deploying the full use of their powers along lines of excellence, instead of the much more boring Ivy red queen’s race, which they might not be suited for.

If you have a team or a company, you can start optimizing for advancement or pay raises or other incentives based on more complex and less Goodhartable outcomes that accommodate what you think their weak spots are individually, that if improved materially would truly signal them performing at that higher level. AI asperformance evaluator and career coach.

If you want to improve your relationship quality with your wife or one of your kids, you can set that as a top level goal and listen to what a Fable-or-better mind suggests.

As I mentioned at the beginning, I’ve been circling around this idea in a couple of areas already. The very idea behind the Infinite Jests is that you can put a complex mind that can evaluate an individual’s engagement and surprisal and arousal and much else in the loop, and gradient descent into maximally super-stimulating entertainments.

The whole idea of the “whisper earring” and my suggestion that you should “yombie,” or become a yuppie zombie listening to and following said earrings, is based on the fact that now an AI mind can literally gradient descent you into complex multipolar outcomes like “I want a great spouse, a career that uses all my powers along lines of excellence, and I want to structure my days so I’m healthy, happy, and engaged with life overall.”

Soooo many people in the ACX comments think that this last one is somehow monstrous, and taking away everything that it means to be human, and “won’t you think of their *souls,* man!”

But no, I AM thinking of their souls and meaning and humanity - if somebody achieves that complex multipolar goal by taking totally voluntary advice (and they have a better shot following a superhuman AI mind’s advice on that front than any other method), they will have achieved human flourishing, full stop.

Look how people run their lives, and look how happy people are about that, and tell me that that’s a better or more desirable outcome than *actually flourishing* on multiple fronts! Forget their souls man, 80% of people are fat and miserable and hate their jobs and spouses!¹ Their souls are already burning in a Sisyphean hell made of their own limitations! It is a strict and gigantic good to give them the option to do better if they want to!

But as I’m expanding on here in this post, it’s actually bigger than just these two, which were just examples of the larger trend!

An entire universe of complex gradient descent is becoming open to us, that is achievable in automatable ways, at scale, with complex and value-laden goals. This is possibly the biggest deal in human history - it’s certainly the biggest deal for potential human flourishing!

Gradient descent whaaaaaat?

Well, what are all the bullshit “soft skills” areas that matter for a good life but can’t be meaningfully improved today for most people and / or institutions?

Here’s some examples, all of which can be decided by an individual:

ACTUAL medical care, instead of the BS “10 minutes total and following an obvious flowchart” medicine as practiced today. AI’s are already much better doctors than human doctors looking at benchmarks. Our medical systems in general perform great when it’s something acute (acute trauma, a broken bone, or an infection) and so poorly it’s basically at the “leeches and plague doctor masks” stage of things for any chronic or ongoing conditions. With AI management, you can actually handle the second thing much better, because the technology and monitoring exists, it just required more expert doctor time than any doctor would put in to do better.

Health, as in inclusive health and capacity, an intrinsically nebulous end that is also quite important and desirable! Like if you think of the factor analysis, it’s probably getting your labs and physical capacity baselines, then building up your strength and cardio capacity, then optimizing your sleep, then optimizing your diet and microbiome, then making sure you spend time outside and with friends, and so on.

Fitness, as in inclusive or focused narrowly in whatever domains you want to improve - on this front, I have already been using AI as my cardio coach, and have materially increased my power output (and presumably my V02max if I had measured it) over the last year. It is an excellent coach.

Better scientific practices - we are capable today of doing fully automated study-quality review and grading, and if we’re capable of that, we’re capable of doing it while any individual scientist is making their choices during the experiment and in their analyses. I would bet AI is already improving science-as-practiced for those using it.

Directly evaluate and improve teaching - what is the current problem with teaching? Well right now, the teacher assigns homework they don’t care about, which the students do with AI, and the teachers grade with AI. There is no learning, and the only human bit in the loop is the wasted time of crafting a prompt and formatting a paper. Disengaged teachers, disengaged students, and nobody involved is teaching, learning, or caring. AI can solve this, because it can infer a learner’s knowledge threshold, interests, preferred rhetorical and learning styles, and then can gradient descent towards “integrated understanding” and do a 1000x better job than any teacher today. This is a homeschooling solution, to be clear, public schools are going to be incinerating gigantic piles of money completely pointlessly forever, probably.

Directly evaluate and improve mentoring - Unlike teaching, mentors and mentees are both fairly motivated, and great things happen in these relationships. However, this is limited by personality, scale, and many other factors - I love and deeply believe in the value of mentoring, but the only people I’ve even tried to mentor are those who I see a lot of myself and my own traits, because advice is always extremely specific and only relevant to 1/1k people in general. You need a lot of high dimensional matching for your mentoring time to be worth it and to even have a chance of moving the needle. But what can AI do? It can enable mentoring at scale, in the styles and formats which most resonate with the menteed audience.

Child development and meaning in ways that actually matter - right now your options for your kids are the Ivy red queen’s race, or plopping them in front of screens and basically giving up on them, or at least this is my 40k foot view take of the choice landscape as actually practiced. Parents rarely try to tailor exploration and development towards true human flourishing for their kids, and of the ones that do try, maybe 10% succeed. But AI can monitor “human flourishing” and “capability development” and “engagement and excitement and skill in areas” more than well enough, and you can gradient descent towards all of those with AI help and the right data pipelines.

Child care in ways that aren’t just babysitting - a similar story to teaching today, with ridiculous red queen’s races meaning even carers for toddlers need masters degrees in most real cities, which is ridiculous. What about a Taskrabbit style platform with individual carers graded by actual engagement and learning outcomes in kids by age? That’s possible with AI in the loop!

Relationship quality - one of those things, like traffic, that are both A) staggeringly important to quality of life, and B) entirely neglected and have zero ongoing optimization energy put into them 99.999% of the time. Well, you can do better with AI! Want a flourishing relationship? Put that endpoint into AI, and try whatever Fable-or-better minds suggest, and keep iterating - it certainly beats the status quo!

Social circles - most people suck at this, because it’s a coordination problem AND a “lots of other obligations in life” problem, and it competes with “staring quietly at screens / wanking,” so you know, it usually loses out. But if you tell AI you want a thriving social life with a rich and intellectually diverse circle of friends, it can do exactly that for you. You just do what it says, input the results, and iterate.

Communication, persuasion, negotiation - we all know the AI’s are already better at persuading than humans. Well, let them level you up! You can get better at any of these now, and these generally generate a lot of high quality data because you get a lot of shots on goal for any of them. I think this is something that would be really helped by real time recordings of interactions, although people (and legal regimes) can certainly differ on the morality / legality of that.

Dating - already the canonical example, even the (entirely fake) allusion to it put Wang / Cluely on the map. Well, it’s real now. Probably need a literal whisper earring for real time optimization, but people have already been using it for pre-date pipeline content, or so I have heard.

Company culture and vision - all these are the most exciting to me, because they’ll directly drive productivity and revenue in measurable ways. Very easy to build a consulting company around this, or to use it to your benefit in your own team or company. But what is one of the most dominant factors that matters for company performance and outcomes? Mission and vision alignment. It’s why so many companies obsess and message this repeatedly. Zappos, a subset of the Faangs, startups, a lot of finance companies - they seem almost cult-like sometimes in the sense that all leaders above a certain level repeat the same mission / vision talking points, in the same words. It’s for a reason, it makes a *gargantuan* difference to team performance or sentiment, and now it can be directly measured and improved with AI in the loop. HUGE deal.

Team culture and vision - same story as above written at the individual team level.

Employee performance - most employee performance evaluation formats today suck. Lots of really smart and talented companies have tried to do better, and they all fall back on peer + manager + manager-peer evals, usually quantified in some dumb way. Well, now there’s a better way! Most of the problem with white collar employee evals is that most roles are not directly revenue producing / affecting, and so can’t be measured by that metric. If you can’t measure by that, and you DO measure by social sentiment, it devolves into who can politic and moral maze the best. That kinda sucks, and we could actually do better, by defining a complex and inclusive value endpoint (ideally mission / vision aligned!) and measuring it with AI. Also if you DID want to put it in revenue terms, this is just a superforecasting problem, which we all know is basically solved already, even in our pre-Fable ante-deluvian times.

Employee development - so once you’re doing that evaluation of performance with AI, you come ready made with the gaps and strengths towards that complex end that you actually care about. So you already have those gaps, and now you can get the AI to directly coach the employee towards strengthening on them and inclusively leveling up to a point they actually are performing at the next highest level and merit advancement.

This is about instrumentation and data pipelines

Maybe you think it sounds too crazy, or I’m just using the words “gradient descent” because it sounds cool, but actually it’s just a handful of AI conversations, but no, I’m literally talking about gradient descent. Any one of these things is high dimensional and leaks data constantly over time. But in an area that’s high dimensional with a ton of diffuse data points (ie sparse matrices), that’s exactly when you want to spin up the algorithms and just literally gradient descent towards the better outcome.

The key now is the instrumentation and data pipelines to store and surface that data in a usable way. What would that look like?

I think ultimately it looks like us using glasshole glasses all the time and models getting good enough at multimodal to extract the relevant pieces of info over many hours of video and audio. We’re not quite there yet, on the model side, but as we all know, the models are improving on a weekly basis.

So what can be done in the interim? Let’s pick “employee development.” Already, employees generate emails, slack messages, commits, and decks, documents, and other artifacts on a daily basis. All that is text! With large context windows, you can feed the whole timewise set in! With more constrained context windows, you feed in appropriately sized pieces and have the AI mind grade them on the metrics / skills / key competencies that matter in your rubrick, then aggregate them up to see timewise trends.

Let’s do another, say “fitness,” because I’ve actually been doing it. Basically, I feed my polar cardio workout data into the LLM, along with an overall textual synopsis of what went well, what didn’t work, where I was struggling, and so on, on roughly a per-workout to ‘over several weeks’ cadence, depending on if I’m trying something new that would benefit from more immediate feedback, or am just putting the sessions in. The AI comes back with fine tuned coaching recommendations and tweaks.

Let’s look at another. Let’s say “relationship quality.” Seemingly a much harder one, right? I mean, I have texts back and forth, and sometimes larger notes or emails, but not much more text than that, and all the important stuff like tone and interaction quality and sex life and whatever else aren’t legible. So what can you do here until multimodal is good and context windows are huge? You just have to talk about all that stuff, with words. This requires more effort and typing on your end. Ideally, you’d have both of you talk in appropriately distinguished ways in the same context session, either with privacy preserved or in a channel where you see both sides, depending on which would work better. I wouldn’t be surprised if somebody hasn’t already githubbed an open source relationship therapist harness, too (I see a few maybes, labeled things like therapist or relationship advisor, but only looked at the high level titles). So basically, you both talk about how you perceive the relationship today, you both work towards and agree on the high level vision you want for your relationship as a destination, then the AI will suggest interventions to both of you which you’ll have to implement and get back to it with notes on what went well, what didn’t, and so on.

Summa

Overall, across all of these, think of the IMMENSE talent bars that can be raised here in the near future, both individually and at the team and company and institution level. Think of how jagged that frontier is going to be. Think of what it would mean for most people to have a genuinely better life on all the fronts that matter. Think what it will mean to be operating or scaling with a much more vision-and-mission aligned team, company, and industry. And obviously, this can happen whether the minds involved are human or AI, or both, in that company.

This is a genuinely amazing pan-spectral greenfield opportunity, whether your metric is meaning and positive impact OR money. On both of those fronts, these are literally the deepest wells that have ever existed, and they have just opened up! All anyone needs to do is choose a focus area and start pumping!

Did you ever do anything with instrumentation or data pipelines? Now’s your chance to change the world - because for any of these domains to be gradient-descentable, it’s entirely a matter of setting up the right instrumentation and data pipelines for the evaluation, output, and gradient descent towards better outcomes across any of these domains. Like we need to get on testing out glasshole glasses data streams and multimodal model setups now, to catch the capability jump when it happens in a suprisingly short amount of time.

So I was personally planning to build a company enabling whisper earrings as my next thing, and it enables all of these. If anyone in the audience has ideas / interest along these lines, feel free to reach out to me in DM’s. Let’s build a future where every person, system, or organization can optimize directly towards complex, fully nuanced, and non-value captured goals!

1. I assume the “80% are fat” doesn’t need a cite (it’s 40/40 overweight / obese), but here.

On their job, in Gallup’s 2025 state of the workplace report, 79% of people report not feeling engaged at work, and something like 66% report they are either struggling or suffering in life overall:

https://imgur.com/0GrD43b

https://www.gallup.com/workplace/349484/state-of-the-global-workplace.aspx

On their spouse, marriage has an ~82% failure rate, in the sense that 20 years in, only 18% of marriages are still together, still mutually happy, and non-dead-bedroom. Obviously, with failure rates that high, and average relationship durations ranging from 5 to 12 years, then throw in the fact it generally takes years of pretty active misery to end up divorced, it’s pretty simple to see that the majority of relationship-years are probably negative at the medians and below.

From a post I did where I looked at the data around marriage quality and duration titled “Against more marriage as a solution to the fertility crisis:”

³ And people spend ~11 hours a day on screens: https://imgur.com/uSQIthV