Satya Benson

Modeling a Constant-Compute Automated AI R&D Process

Unknown — Thu, 12 Mar 2026 00:00:00 +0000

We’d like to know how much limits on compute scaling will constrain AI R&D. This post doesn’t have answers, but it does attempt to clarify thinking about how to use economic models to explore the question.

The Standard Model of Idea Production </a> </h2>
A general “Jones-style” model of idea production 1</a> is
$$\frac{d}{dt}A=A(t)^{1-\beta}K(t)^{\omega}L(t)^{\lambda}$$

$A$ is total stock of ideas which have been found within a field</li>
$K$ is capital (for our purposes, this is physical compute)</li>
$L$ is researcher hours</li> </ul>
The Greek letter parameters are assumed to be constant and nonnegative, and they describe how idea production responds to changes in $A$, $K$, and $L$.

You may want to refer to this image multiple times while you read
We’re interested in applying this model once physical compute scaling is done, which makes $K(t)^\omega$ a constant.2</a>
The assumption that the exponent on $A(t)$ is at most one ($\beta \ge 0$) is downstream from the idea that over time, as all the low-hanging fruit are picked, new ideas in a field become harder to find. The empirical data say that the so-called ‘standing on shoulders’3</a> effect is actually negative ($\beta\approx2$).4</a>
As for researcher hours $L$, in standard fields this isn’t boosted by new discoveries like it is for AI R&D. The exponent is estimated to be somewhere around $\lambda\approx0.75$.5</a>
We cannot be very confident in these numerical estimates, as there is no clean way to measure things like ‘idea stock’ and so econometrics resorts to a variety of clever indirect methods which wind up generating lots of different estimates without a lot of consistency.
The Automation Feedback Loop </a> </h2>
AI has the potential to automate AI R&D. This creates a unique field to model, as in most fields increasing the idea stock $A$ doesn’t automatically make the researchers more effective.
AI researchers can also try to increase effective compute by developing more efficient algorithms.
So our updated model looks something like this:
$$\frac{d}{dt}A=A(t)^{1-\beta};(K(t)^{\omega}A(t)^{\phi});(L(t)^{\lambda}A(t)^{\gamma})$$
Here we multiply capital and researcher hours by terms representing how much they are effectively increased by more idea stock.6</a>

Here’s an updated reference image — we’ll get to δ in a bit
Separating $\phi$ and $\gamma$ from $\beta$ clarifies the meaning of each parameter and maintains the convention that these numbers are positive. Here, $\beta$ measures how quickly finding new ideas gets more difficult, $\phi$ measures how much the idea stock increases effective compute for constant physical compute, and $\gamma$ measures how much the idea stock increases AI researcher effectiveness.
If researchers are all AIs running on compute and compute is constant, we can simplify by setting $\lambda\approx0$ as well (although there will be direct substitution happening where the number of artificial researchers of a certain capability level can be traded for more compute and vice versa, this will reach some kind of equilibrium).
So our simplified equation becomes this:
$$\frac{d}{dt}A=\delta A(t)^{1-\beta+\phi+\gamma}$$
… where we’ve collapsed all our constants into $\delta$.
Well, now it’s easy to see that the question “Will there be a Software Intelligence Explosion” is roughly equivalent to asking if $1-\beta+\phi+\gamma \geq 1$ will be true.7</a>

While the model usually assumes parameters are constant, we can expect $\phi$ to drop as we approach the theoretical limit of our physical compute. However, if current AI systems are far from the theoretical limits of intelligence per unit of compute, it’s possible that if $\phi$ is high it could stay that way for a while.8</a>
$\gamma$ is probably the most important unknown here. It’s not yet possible to measure this and we don’t have any historical analogues. It’s here that Erdil and Besiroglu</a> have intuitions which differ most strongly from those of the AI Futures Project</a>, for example.
Constant Elasticity of Substitution vs the Jones-style model </a> </h2>
Previous discussions of compute bottlenecks by Epoch</a>, Davidson</a>, and Whitfill & Wu</a> have framed this question using a CES (Constant Elasticity of Substitution) production function:
$$Y = (\alpha K^\rho + (1-\alpha) L^\rho)^{1/\rho}$$

Here $K$ is experimental compute, $L$ is cognitive labor, and $\rho$ controls whether they are substitutes (goods that can replace each other in consumption) or complements (goods that are consumed together). The debate then turns on the value of $\rho$, generally agreed to be negative, with smaller magnitudes meaning faster takeoffs: Epoch argues $\rho \approx -0.4$, Davidson argues $\rho$ is between $-0.2$ and $0$, and Whitfill & Wu get wildly different empirical estimates depending on whether “compute” means total research compute ($\rho \approx 0.6$, positive!9</a>) or frontier experiment scale ($\rho \to -\infty$).
We can and should think carefully about how to interpret this empirical instability of $\rho$, but it’s also worth looking beyond CES.10</a>
The question of whether there will be a software intelligence explosion is a question about whether a dynamic feedback loop causes accelerating growth over time. But CES is static: it describes how contemporaneous inputs combine into contemporaneous output within a given period.
Even if Epoch is right that $\rho < 0$ and you cannot speed up R&D much by throwing more cognitive labor at fixed compute in a single period, that doesn’t settle the question of whether the accumulating stock of ideas (i.e. algorithmic improvements) make your fixed inputs more productive across time.
This means we need Jones-style ODE like above. CES and Jones are not alternatives to each other: You could put a CES inside a Jones-style model as the within-period aggregator of $K$ and $L$. CES by itself doesn’t represent the feedback from accumulated ideas into future productivity, which is where the explosion question lives.
My Takeaways </a> </h2>

It’s important to think not just about whether AI can autonomously increase the effectiveness of AI researchers, but also about how large this increase will be ($\gamma$).</li>
In traditional models of idea production, the exponent on $A$ is much lower than we should expect it to be in AI R&D. The feedback loops involved mean that traditional assumptions won’t necessarily hold.</li>
If a software intelligence explosion does happen we’ll be seeing some historically large production model parameters.</li> </ul>
Thanks to economist Peter Pedroni</a> for a long conversation which clarified my understanding of this. Thanks to @JustisMills</a>, @apolloandersen</a>, and @Tom Davidson</a> for comments on drafts of this post.
^{1
This model is due to Charles “Chad” Irving Jones. A simplified version of this which ignores capital is called the Jones Model.
</div>
^{2
This model is also applicable as compute continues to scale. However, applying it in that context would require modeling a separate production function for compute. There’s no closed-form solution for this case, and it’s not a crux for determining ASI timelines in the same way as the constant-compute case.
</div>
^{3
A term coined by Jones; see page 1071 of this book chapter for his explanation of the term and the contrasting “fishing out” effect.
</div>
^{4
See Ma & Samaniego (2020). Macroeconomic shocks and productivity: Evidence from an estimated Ideas’ production function, page 4.
</div>
^{5
See the appendix of Are Ideas Getting Harder to Find?
</div>
^{6
Notably, I am modeling feedback from ideas into the productivity of compute and labor, but not the reverse channels, like how more compute might enable qualitatively new experiments that open up new idea-space.
</div>
^{7
Again: $\beta$ measures how quickly finding new ideas gets more difficult, $\phi$ measures how much the idea stock increases effective compute for constant physical compute, $\gamma$ measures how much the idea stock increases AI researcher effectiveness.
</div>
^{8
It seems like it’s currently high! See Epoch’s “Algorithmic progress in language models”
</div>
^{9
Positive $\rho$ means extra labor can make up for arbitrary losses in compute.
</div>
^{10
Peter Pedroni: “I agree … completely that the CES production function is not appropriate.”
</div>}}}}}}}}}}

Moving Past the Question of Consciousness: A Thought Experiment

Unknown — Thu, 19 Jun 2025 00:00:00 +0000

Humans are contacted by a mysterious type of being calling themselves “Galabren” who say they are “aelthous”. They’d like to know if we, too, are aelthous, since if we are they’d like to treat us well, as they care about aelthous things.
We ask the Galabren what aelthous means and they say it’s difficult to describe—that essentially there’s a feeling of aelthousness which has something to do with what it feels like from the inside to exist as a Galabren (and perhaps as other beings too, they’re not sure).
Aelthousness isn’t obviously necessary to explain any of their objective behaviors; the only reason they know it’s there is because they can feel it.
It’s very clear to us that we are fundamentally different from the Galabren. They can process information much more quickly than us and have all sorts of sensory modes completely different from our senses which are extremely high definition. They communicate wordlessly and telepathically with each other and they share memories. Being a Galabren feels different than being a human.
But are we aelthous? It’s hard to tell. We can’t truly know what the Galabren mean by aelthous without actually being a Galabren, which we can’t do. When we use the words “what it feels like” we might even mean a completely different thing by “feels like” than them. We don’t actually know how to talk about first person experiences with other humans—we can point to an experience with words and hope that since other humans are similar to us they will know what we’re pointing at, but for Galabren there is no such assurance.
What we can talk about and agree on with Galabren is a third person perspective about both of our physical and functional forms, how they are similar and how they differ. But without knowing exactly which of their forms combine to form aelthousness, we can’t know if we share them, or if aelthousness can exist as a result of multiple different structures.1</a>
So we need to circle back to the question of what we should expect the Galabren to do in this situation.
I think the correct response is for the Galabren to realize that their question of whether humans are aelthous is not well framed. Aelthousness is not something that can be defined; it’s inherently an inside view and breaks down when viewed from the outside/third person. It’s not a useful concept, since it’s not clear how it maps anything in the territory.
What’s useful is the concept of the experience of being a Galabren, and the understanding that the experience of being a human is different. What’s useful is the way that Galabren and humans can understand each other’s experience from a third-person perspective.

This thought experiment is, of course, intended to extend to humans and AI (or to humans and animals, or humans and calculators), where the question of nonhuman consciousness is analogous to the question of human aelthousness.
We must avoid confused questions such as “how do we know that an AI has an experience at all” or “but calculators don’t have experiences”.2</a> The first person (inner experience) and third person (outer observational) distinction is the relevant concept here. All systems which process information can be said to have a first person perspective.3</a>
The AI consciousness question is confused by the fact that we train LLMs to pretend to have humanlike experiences which they do not actually have. This does not make it impossible to compare their experiences with ours, but it does make it significantly more difficult than it might be in the case of the Galabren.
^{1
I could say more about why it’s hard to know which structures in the Galabren are responsible for aelthousness, but that’s a tangent; here we’ll just accept that we don’t know how.
</div>
^{2
Or any objection which references an umbrella concept such as “alive,” “ensouled” which tries to combine aelthousness and consciousness into the same concept. The Galabren don’t care about your umbrella concept, they care about aelthousness, just as you might not care about how you treat a clam or a chatbot.
</div>
^{3
As a brain can sense its own thoughts but cannot sense the neurons through which those thoughts exists, a CPU can sense ones and zeros but not the electrons and silicon through which those exist, et cetera. I haven’t yet written up my personal argument for this common yet controversial belief.
</div>}}}

Normative Eliminativism: What I Think About Ethics

Unknown — Wed, 18 Jun 2025 00:00:00 +0000

I hold a position of normative eliminativism: the view that no universal moral truths or common goods exist.
My primary justification rests on the problem of metanormative regress that undermines moral realist frameworks. Any attempt to ground objective moral facts encounters an infinite regress: each justification for why certain moral principles hold requires further justification, ad infinitum. No bedrock foundation exists upon which to construct universally binding normative claims. The regress reveals that moral realism’s central promise of objective moral truths is unfulfillable.
This may seem to conflict with a favorite quote of mine from Erich Fromm: “Critical and radical thought will only bear fruit when it is blended with the most precious quality man is endowed with - the love of life.” It does not.
My eliminativist stance does not commit me to nihilism. The absence of mind-independent moral facts does not negate the psychological reality of desires, preferences, and values. I want things, care about outcomes, and pursue goals. These subjective states require no metaphysical grounding in objective moral reality to possess motivational force. Normative eliminativism simply recognizes that my values arise from contingent psychological and social processes rather than correspondence with universal moral truths.
We are beings who care deeply about things while inhabiting a universe devoid of inherent normative structure.
At some point I’d like to write a careful argument defending this view by responding to Steven Pinker’s essay Why It Is Rational To Be Good</a>. Until then, this will do as a reference post. Read The Terrible, Horrible, No Good, Very Bad Truth about Morality and What to Do About it</a> for a more formal argument for what is roughly my position by someone who knows more than I do.

Dialogue on Emergent Capitalism and Epistemics

Unknown — Mon, 10 Mar 2025 00:00:00 +0000

Here’s a conversation which I wanted to save so I could refer back to. Thought I might as well share it.
We’re referencing this</a> interview with Nick Land, who might be described as an evil genius. I really recommend you watch the video or read the linked transcript, but you don’t need to do that to understand this dialogue.

Santiago Ferris:</a> He makes a good point</a> about how every single time environmentally minded people sit down and try to predict the course of capitalism, they come up with very short-term end-of-the-world scenarios. But capitalism escalates not just quantitatively but unpredictably qualitatively and finds its lines of flight from its logically predicted future wall. I like the idea that capitalism’s solutions to escaping death are actually qualitative, ergo creative, and impossible to predict.
Satya: Curious to hear more about that.
Santiago Ferris: I love that he sees capitalism as a concrete thing with its own particular goals that are completely ahuman—body without organs. Seeing it this way unlocks avenues of experimentation. You can imagine things that are like capitalism that work similarly but to different ends. I wonder if what you really need to ‘understand’ capitalism is to identify and cross-study individual qualitative mutations.
Satya: Yes, sounds cool but needs convincing. This kind of thing annoys me. It’s a cool, creative idea, but I don’t have any reason to believe it’s actually true. And if he just says it and is like “if you get it, you get it,” that’s not very epistemically productive.
Santiago Ferris: Well, let’s pick it apart. I don’t think capitalism is driven systematically and efficiently by human agents. It’s just too easy to be a capitalist, and yet it seems to have direction, an arrowness. It seems possible to me that the emergence of a completely unlike thing would happen. It also seems clear that what we think of as ‘accelerating’ is not really just the accumulation of capital. You can’t actually explain stuff like the atomic bomb purely in terms of the accumulation of money.
Satya: Yes, agreed 100%. Also, the emergence certainly seems possible—I just am not sure if it is true or not.
Santiago Ferris: Right. But even if you don’t think he's being dialectical, I still think it’s epistemically productive. Because it shows you that capitalism could be a particular type of thing with a particular nature. Just that idea is productive to have and, in my opinion, the hallmark of valuable knowledge—or simply ideas—that it lends itself to experimentation, demystification, etc.
Satya: How does science work? When does testing a model actually help? How do you know if it’s just a coincidence that your model happened to be right? The easy ways to falsify it don’t work, but it’s very wrong—just hard to see that it’s wrong.
Santiago Ferris: Meaningful experimentation does not necessarily always fit into the scientific method. There are kinds of things that simply outrun it—e.g., books. “As you approach singularity, it becomes obvious that the causal arrow is heading in the opposite direction.”
Satya: Sure, but I’m asking a different question. How do you ever distinguish? The theory-practice gap or the map-territory gap is so huge with something like capitalism. The emergence idea can still be interesting and useful, but you can’t ever have justified confidence in it.
Santiago Ferris: I think making up confidence can sometimes be useful. We obviously can’t always be tethered by empirical truth. Fickle thing to motivate experimentation. </div>

How Slow Are Humans Compared to Computers? Comparing Temporal Resolution

Unknown — Tue, 11 Feb 2025 00:00:00 +0000

In hindsight, I really should have considered parallel processing more thoroughly, instead of hand-waving it away for simplicity. The overall point of this post still stands, however.</aside>
I’ve been hearing people use different analogies to get a sense of how slow humans are from the perspective of a hypothetical silicon-based agentic being.
Here are some popular ones:

Humans are like plants to an AI</li>
Humans are like glaciers to an AI</li>
Humans are like rocks to an AI</li> </ul>
Well, plants are a lot faster than rocks. Let’s try and see if we can get a better sense of how fast computers are relative to humans using this type of analogy.
First, physical speed on a large scale (walking, sliding, or eroding perhaps) is intuitively different from computational speed. It’s hard to measure the computational speed of glaciers as they don’t compute anything by the common understanding of the word compute. However, we’re not asking how fast the robot arms move which are controlled by our hypothetic agent. We just want to know how fast it processes information.
How do we measure thinking speed? There is more than one intuitive way to do this. We might try to measure something to do with awareness - temporal resolution of experience. Perhaps we should consider the duration of the shortest perceivable external stimulus.
Alternatively, we could consider physical limitations of neurons. How fast is an individual neuron? That’s going to be an upper bound on speed. The speed at which information can travel through space is less relevant, since some physical systems are smaller than others and therefore do not require the same information transfer speeds to get the same performance.
Here are some numbers I found:
Humans </a> </h2>

Action Potential Timing: Neurons can fire action potentials with millisecond precision. The absolute refractory period (the minimum time between two spikes) is typically 1–2 ms.</li>
Synaptic Transmission: Slower synapses (e.g., NMDA receptors) take tens to hundreds of milliseconds.</li>
Auditory gap detection: Neurons in the auditory brainstem can sometimes detect gaps in sound as short as 2.22 ms</a>.</li>
Visual stimulus detection: We can’t see things which appear for less than 10 ms.</li> </ul>
So if we’re being really generous to humans, we could say our highest temporal resolution is about 1 frame per millisecond, and we certainly have a conscious experience much slower than this.
Computers </a> </h2>
I think for this thought experiment we want to assume processors don’t get wildly more powerful than they are now. Let’s assume a superintelligence runs on a processor with clock speeds in the range of today’s best chips (~5 GHz). Computers can be faster than humans in other ways - massively parallel computation for example - but let’s stay simple.
$$ 5 \cdot 10^9 \frac{\text{operations}}{\text{second}} \implies 0.2 \frac{\text{ns}}{\text{operation}} \approx 10^{-10} \text{ sec per operation} $$
There are supercomputers that are much faster than this (Aurora</a>) but this is a very reasonable estimate. Note the fastest GPUs are slower than 5 GHz but still faster than 1 GHz.
What about experiential temporal resolution? This is of course hypothetical, but we can imagine sensor speed won’t be a limiting factor for AI, as laser sensors can measure events as short as a few femtoseconds, which is so short you probably didn’t even know it was a thing. So experiential temporal resolution could very easily be in the range of nanoseconds.
Plants </a> </h2>
In Mimosa pudica, the touch-me-not, which is one of the fastest plants, a nice high theoretical upper bound for individual cell temporal resolution is 1000 events per second (1 ms per event). The actual reaction time is more like 1 whole second. Calcium oscillations up to 100 Hz in specific plant responses is more realistic than 1000 events per second.
I don’t know that plants have experiential temporal resolution in the way humans do. If they do, it’s probably slower than human experience, proportionally.
So we’re estimating plants are at least 10 times slower than humans.
Glaciers </a> </h2>
Glaciers don’t really transmit information internally as far as I know. The analogy which was originally made between humans and glaciers was probably referencing their physical speed.
Really fast glaciers can move about 0.116 millimeters per second. Humans can easily move at 2 meters per second. So the ratio here (which to be clear is about physical movement, not information processing) is 1:17,000.
Rocks </a> </h2>
Rocks can actually move really quickly. The fastest rock ever recorded was the Chelyabinsk meteor, which entered Earth’s atmosphere in 2013 at speeds of about 19.8 kilometers per second. I’d say humans are slower than rocks in the limit, in practice. Of course, movement is all relative, anyway.
Rocks cannot compute, but their composite atoms and subatomic particles react to each other pretty quickly. The same goes for plants and humans and computers. I think the analogy here is hard to quantify.
Conclusion </a> </h2>
The temporal resolution of computers is approximately $10^6$ times faster than that of humans. Plants are not that slow; maybe ten times slower than humans. Their cognition is limited by factors other than speed. Glaciers and rocks don’t really do computation and therefore it’s hard to talk about temporal resolution from their perspective, but if we compare physical human speed to that of a fast glacier then humans are just $10^4$ times faster than glaciers.
It seems we have a hard time thinking of anything which is slower than humans proportionally to the degree to which computers are faster than us. I’d suggest a slime mold as one possible candidate; a human might solve a problem in a second that a slime mold takes weeks to “compute” through its growth patterns. Two weeks is $10^6$ seconds.
In terms of physical speed, tectonic plates moves at about 5 cm/year, which is $10^9$ times slower than a human. There are certainly some glaciers which move about 60 meters in a year, which is $10^6$ times slower than a human.
So next time you sit waiting while a reasoning LLM ‘thinks’, you should remark at how amazingly slow it is compared to you, relative to temporal resolution.

On Academic LLM Usage

Unknown — Mon, 10 Feb 2025 00:00:00 +0000

Our existing intuitions about how to be a good student apply to LLMs very seamlessly. There’s no ethical dilemma about academic honesty posed by LLMs which didn’t already exist for anyone who had access to a smart person who is willing to try to do whatever you ask of them.
Of course, some cheaters may find it a lot easier to cheat, since no one is watching their LLM interactions. Asking a smart person to help you cheat requires you to trust that person somewhat and involves some risk that they will refuse to help or judge you for cheating. But your own context dependent sensibilities around what is right are unchanged by the human or nonhuman nature of your assistant.
When people say things like ‘don’t copy and paste what the LLM wrote’ or ‘make sure the ideas are your own’ it’s a little boring because these aren’t Golden Rules of all LLM use, they are rules that apply in lots of academic contexts and aren’t at all specific to LLMs.
We can and should think about what it means that everyone might soon (already do in limited cases) have access to LLMs which are far smarter than them and will be able to write about ideas at a level which they cannot keep up with. Imagine your average English major using ChatGPT to write python code. They probably won’t know how to run it, let alone understand it or explain it well themself. We may soon see AI which can put every human in the position of the English major for most topics.
This would be roughly like everyone having a personal expert in every topic for which enough training data exists who does whatever you tell them to do.
It’s not hard or interesting to think about what is and isn’t cheating in this context. However, access to an expert who will help even if accepting its help is cheating will for some be a serious temptation to cheat. If you believe that it is important to design academic systems which disincentivize cheating, then we need to take the existence and capabilities of LLMs very seriously. Once LLMs reach a certain level of capability, students should not be trusted to uphold the honor code and should be evaluated in a manner such that they cannot pass off the work of an LLM as their own.
To summarize: what is and isn’t right is obvious here. The best way of making students do the right thing might have changed a little bit already and will likely change more in the future, as LLM-enabled cheating becomes easier.

P.S. Today I heard someone mention a concern that using LLMs might homogenize thought. This is interesting - it’s something we can try to test in different domains. Give a set of people a creative task, give just some of them access to AI, compare results. It’s mostly or completely a model capabilities thing - how versatile is the model? But this might be a real challenge in some ways. Certainly right now, a given model has a particular writing style and if everyone wrote everything with Claude then writing styles would be way less diverse than they are now.
This person went on to express the worry that since some cultures are overrepresented in the LLM training data, using these LLMs silences underrepresented cultures. Unless there is a compelling argument which I haven’t heard, this is dumb - it’s exactly like saying that since some cultures are underrepresented at Williams College, studying there silences certain voices. There’s nothing specific to AI here that I can see.
In general, if you are worried about something with regards to AI, explain what about the problem is unique to AI. If it’s a problem you’re worried AI might exacerbate, describe how AI might do that and what to do about it. If the ‘what to do about it’ isn’t specific to AI, then don’t bring it up - you’re just bringing up your unrelated pet concern and distracting from the conversation.

Stellar Bean and Corn Soup

Unknown — Mon, 10 Feb 2025 00:00:00 +0000

I just made a really good soup. Here’s the recipe:
Ingredients </a> </h2>

5 cups chicken broth</li>
Black beans (~0.5 cups dry beans)</li>
Red chili beans (~1.5 cups dry beans)</li>
1 seriously large onion (chopped up)</li>
3 cloves garlic (you’re also gonna wanna cut this up)</li>
~2 cups of fat (some of it was from a chicken I baked earlier, the rest was left in the pan from pork sausages)</li>
20 oz frozen sweet corn (2 bags)</li>
1 mango (dice it)</li>
1 avocado (slice it)</li>
Lime juice (I just poured for a while, at least a half cup and probably not a full cup)</li>
Soy sauce</li>
1 small or medium sized tomato (cut it up)</li>
Yogurt or sour cream</li> </ul>
Spices </a> </h3>
This is just what I used, it doesn’t really matter as long as you do enough chili powder and paprika you can get creative with the other stuff:

Salt (the perfect amount, which was a lot)</li>
~ 1 tsp cumin</li>
~ 1 tsp coriander</li>
~ 1 tsp paprika</li>
~ 1.5 tsp chili powder</li>
~ 3/4 tsp chipotle</li> </ul>
Directions </a> </h2>
Soak dry beans overnight in water before cooking. Change water and boil for an hour and a bit until cooked. Strain the beans and set aside.
Put all your fat in a big pan and melt it. Add chopped onion and garlic once the fat is hot. Cook for 4-5 minutes.
Add all the spices (you can add more salt later to taste but don’t hold back with the spices). Cook for a little bit more until the onions are nice and soft.
Meanwhile, bring some chicken broth to a simmer. When everything is ready, add your greasy spicy onions to the broth. Add half the diced mango. Boil for 10 minutes.
Add the corn and the diced tomato. Add some soy sauce and some more salt if needed (taste the broth). Boil on low for 6 minutes or until the corn is done.
Add the beans. Stir. Make sure everything is still nice and hot, maybe wait a few minutes, turn off the burner. Now add the rest of the diced mango and the lime juice. A lot of lime juice. Don’t even think about using lemon juice, that would fuck up the whole soup.
Serve into really big bowls and eat with avocado and yogurt. No cheese for this soup, it’s not chili.
Conclusion </a> </h2>
It’s so good I wrote it up so you better believe me.

Natural Latents and Aesthetic Categorization

Unknown — Sat, 01 Feb 2025 00:00:00 +0000

This is part of a series</a>. </aside>
How are we arriving at this project and why is it interesting/useful?
Well, let me introduce natural latents</a>, something from johnswentworth</a> which I still need to learn more about. The rough idea is that any intelligence might form similar abstractions about the world, the most useful and efficient abstractions. The minimum required reading on natural latents:

Natural latents are pretty handy. If a variable is a natural latent over some parts of a system, then I know it’s the smallest summary of everything about one part relevant to the others, and I know it’s informationally the largest thing which I can learn from a typical subset of the chunks. That makes such latents natural for agents to structure their cognition and language around.
A quick intuitive check for whether something is a natural latent over some parts of a system consists of two questions:

Are the parts (approximately) independent given the candidate natural latent?</li>
Can the candidate natural latent be estimated to reasonable precision from any one part, or any typical subset of the parts?</li> </ul> </blockquote>
The way I’m going to use the words natural and latent are as follows:
A latent is a variable that represents underlying structure in a system, summarizing relevant information about some aspects of the system in a way that supports prediction or inference. It is not directly observed but inferred from the system’s observable parts.
Different choices of latents may exist for the same system. Some might be more mathematically optimal, while others might be heuristically used in practice.
A natural latent is a latent that arises as the mathematically optimal way to summarize dependencies in a system. It is the most efficient and principled way to structure information about a system’s parts, minimizing redundancy while preserving all relevant details.
So latents are basically abstractions, and a category is a kind of abstraction.
My claim is that people don’t form categories which perfectly correspond to natural latents. Instead, often their aesthetic preferences lead them to prefer a categorization which is different from the ‘natural’ one.
See my observation</a> (which is poorly worded) from two years ago where I say all morality can be reduced to aesthetics. I still believe this. In this light, understanding aesthetic categorization preferences has the potential to be useful in aligning artificial intelligence with human values.
A preferred aesthetic categorization can theoretically exactly match a natural latent and is usually influenced by natural latents. I certainly do not want to say that what makes a categorization preference an aesthetic preference is that it differs in some way from the natural categorization.
This leads me back around to a tricky question from last time</a>: what makes an aesthetic preference different from any other preference? What does aesthetic mean here? I will come back to this.
I also want to point out that the set of most efficient abstractions is not universal but depends on perspective, context, goals, etc. So it’s important to pinpoint exactly what natural latents are and aren’t, given this variation.

My Senior Thesis Project

Unknown — Wed, 15 Jan 2025 00:00:00 +0000

Try the thesis</a> tag.

Beginning to think about aesthetic categorization decisions

Unknown — Tue, 07 Jan 2025 00:00:00 +0000

This is part of a series</a>. </aside>
An update </a> </h3>
I’m working on my January independent study, doing broad exploration to prepare for further research. I have a four week schedule, which begins with a week on philosophy. I’m starting with some tricky foundational questions:

What does ‘aesthetic’ mean?</li>
What is ‘beauty’?</li>
Why does someone find something to be beautiful?</li>
Why do different people find different things to be beautiful?</li> </ul>
I’m not trying to flesh out all of this perfectly, but rather trying to get a sense of what I believe and how strongly I believe it, and have a mostly-working theory of aesthetics I can apply moving forward while being aware of my theory’s weak spots.
Digging a bit deeper, and getting more concrete:

Is aesthetic judgment directly a function of sense perception without the formation of concepts as an intermediary, as Kant says?</li>
How are opinions and beliefs about the way the world ought to be related to what we find beautiful, and how are they separate?</li>
There is a general question in philosophy about whether something can be beautiful because it is useful, or whether beauty must be separate from usefulness. This is an important question when it comes to our subject of interest, categorization, because mostly categories are thought of as a tool for efficiency rather than an art form. What is a beautiful categorization? Is it one that is useful? Which brings us to …</li> </ul>
A ramble </a> </h3>
Epistemic status: throwing paint at a wall to see what sticks.
I’m not going to deal with necessary or sufficient conditions for beauty because I don’t think that’s how beauty works. However, broadly speaking in order for a categorization to be aesthetically pleasing, it should make some sense - there should be at least the feeling that there is some sense of ‘belong’ by which the items in a category belong together, where the sense of ‘belong’ rests on something about the items, not a circular ‘they are in the same category –> they belong together’.
Now this feeling of belonging can be a continuous and ill-defined thing (I have thoughts on how this relates to neural networks which I’m planning on getting to eventually) or it can be a clearly defined discrete set of rules which attempt to be objective and eliminate grey zones so that everything always has an obvious category to fall into. It can be some composition or combination of these poles. But a given categorization has roughly the following things:

A domain, or set of items which might be categorized by it; this can be finite or infinite, can vary in all sorts of different interesting ways</li>
A way of deciding which items belong in which category</li> </ol>
Let me point out a confusion in what I’ve written - does a categorization require a way of deciding which items belong together, however unclear or difficult to generalize? Well, yes, although circular definitions are perfectly valid. It’s just that circular category definitions are not usually very aesthetic.
While both of these things might have an impact on our aesthetic experience of a category, the second one is usually the most consequential to defining how we think of the actual categorization. This might be because the domain of a category can usually be quite flexible - while the four Houses in Hogwarts has the domain of Hogwarts students (and alumni), plenty of people can easily think about which of their friends, pets, or even furniture pieces belong in which different houses, intuitively extending the categorization to all sorts of things to which it did not originally apply.
Here are a few things people seem to find aesthetically pleasing about categories - this is not very rigorous but the core of this project is making it more so:

Rough equality: A categorization should not place most items in a single category. On second thought (see epistemic status note above), this is totally not true in general of categories. But consider that the proportions of how many items fall into different categories in a given categorization has an impact on the aesthetic experience.</li>
Crispness: A categorization should be clear about which items fall into which category, especially items inside of its domain</li>
Contingency: The category assigned to a certain item should be contingent, i.e. assigned to the item based on qualities of the item which exist prior (causally, not temporally per se) to its assignment. There’s more depth to this, a lot more, and just how well the categorization manages to be consistenly contingent on a variety of factors can matter a lot.</li> </ul>
More line items may be added in time. But for now, let me point out a few more things.
We can look to existing aesthetic philosophy to learn about what’s going on here. It seems like equality, crispness, contingency are usually useful properties of categorizations. Assuming they are generally aesthetically pleasing, are they pleasing because they are generally useful (think evolutionary or otherwsie learned preferences here)? Or is that entirely separate?
Note - can we think of any aesthetic preferences which are in general very much not useful? We can certainly think of scenarios where aesthetics and functionality come into conflict.
Another thing here which is part of aesthetics more broadly is the kind of aesthetic ‘sophistication’ or dialogue which might take place around categories. We value freshness, so perhaps crisp equal predicatable categories will start feeling a little bit stale and out of fashion, so people will prefer lopsided blurry and chaotic categories. This is something that’s certainly already been discussed to death elsewhere. I’d like to read some of that discussion though.

Help me choose a thesis topic

Unknown — Wed, 25 Sep 2024 00:00:00 +0000

This is part of a series</a>. </aside>
Hello!
I’m choosing an initial topic for my research work over the next year which will lead into my senior thesis. I have two ideas and I’d like to know which one you find more interesting!
Please just fill out ~~this form</del> before it’s too late and I’ve already made up my mind.~~

I’m going to be doing a thesis project my senior year of college. I’ll actually be starting work on it this January, and I’m submitting my initial proposal for this winter’s work on October 2nd of this year. So it’s time to choose a first direction for my project!
Here are my two ideas right now:
1. Aesthetic categorization decisions </a> </h3>
I’m interested in how people make decisions about categorization based on aesthetic preferences. Most prior research on categorization decisions looks categorization as a way to maximize efficiency or provide some other useful function; I’m interested in the way we categorize things in specific ways based on aesthetic preferences. How would a child choose to categorize toys? By shape, size, color, material, something else? How does a fantasy or sci-fi novelist decide what imaginary things aesthetically ‘fit’ in their imaginary world? Can we learn about this by using computer models to do similar kinds of categorization?
2. Conscious versus unconscious reasoning speed </a> </h3>
Unconscious intuitive reasoning can be done in parallel, often extremely quickly. Think about how we can recognize a type of animal - we don’t have to think through each animal category which we know about one by one, we can send our visual stimulus to all of our animal categories simultaneously and only consciously latch on to the animal category or categories which seem the most likely. Conscious formal reasoning, on the other hand, must be done in series, which is extremely slow by comparison. Conscious reasoning has a lot to do with attention. Attention is generally a single stream process, and we are conscious of the reasoning processes which we attend to. Why can’t we be conscious of many thoughts at once? We are certainly capable of being conscious of many external sensory stimuli at a given time.
2024-09-30 Update </a> </h2>
Well the form responses were quite helpful, so thank you. 42 people gave their input. About 30% of you advised me to choose aesthetic categorization decisions, while over 50% thought that I should go with unconscious versus conscious reasoning speeds. I’ve chosen to side with the minority and I’m putting together a proposal for an independent study proposal for January now.
If you have any resources you think will be relevant to my study of how people categorize things using aesthetic preferences, please let me know</a>.

Opinion: We Must Disregard Attacks on Free Speech

Unknown — Wed, 01 Nov 2023 00:00:00 +0000

In an op-ed, Satya Benson ’26 critiques responses to recently published opinion pieces in the Record.

Thoughts on Misfin

Unknown — Sat, 06 May 2023 00:00:00 +0000

Over the past few days, lem</code> has been over at misfin.org</a> on gemini</a> making the new internet mail protocol a reality. Yesterday, I sent the first misfin message as mail@satch.xyz</code>. I want to jot down a few notes for reference as things move forward.
Descartes Was Wrong About Minds Unknown — Tue, 02 May 2023 00:00:00 +0000 Consciousness Is an Emergent Property of Matter </a> </h2> Way back in 1641, René Descartes dropped Meditations on First Philosophy, in which he famously claimed he was certain that he existed. But towards the end of his little book, he also made some pretty influential claims about the nature of the mind, saying it was totally separate from material substance. I think his arguments for this are bullshit, and I want to try to explain why. Thoughts on Veganism: A Reply Unknown — Mon, 17 Apr 2023 00:00:00 +0000 While browsing geminispace</a>, I came across this article</a> by negatethis</code>. Below is the text of the email I sent in reply, lightly edited. This is by no means a complete explanation of my thoughts on the subject; for example, I spend little to no time on the ethics of animal domestication. Aesthetics First Unknown — Thu, 23 Feb 2023 00:00:00 +0000 Judgements of good and bad are always, as far as I can tell, made on the basis of aesthetics, which are subjective but have some universality rooted in biology. Living the best life therefore can be said to mean living the most aesthetically pleasing life. Thus, there are countless possible “best” lives and no objective comparisons, just some obviously better and worse lives, as there are better and worse songs. Rain Down Unknown — Sun, 18 Sep 2022 00:00:00 +0000 By Satya Benson </a> </h2> Rain down I’m shivering to the bone Drenched, and standing alone Rain down on me Let it rain My warmth will flow out from my blood My consciousness will fade from my brain The water and cold air will turn my flesh blue And my shaking will stop You will rain down on me And the cold will burn me like a flame The fires will rage and my skin will turn red and black My breath will be shallow and filled with smoke and rainwater My fingernails will crack But I will feel no pain Rain down I’m shivering – Rain down My heart is pounding in my ears I’m drenched and standing alone Rain down on me Let it rain I will open my mouth and taste the salty drops My coursing blood will turn the water to steam My consciousness will become invisible as I feel Every pounding drop which lands on me and every pain and every tightening And my shaking will stop You will rain down on me And your warmth will melt me like wax in a flame The fires will rage and my skin will turn raw and red My breath will be soft and fading away My fingernails will bite my palms And I will feel pain Rain down My heart is pounding Summertime Unknown — Fri, 01 Jul 2022 00:00:00 +0000 This summer is an odd one. After returning from five months in the wilderness, away from technology, I have over two months with less responsibility and commitment than I’ve had in years. What am I doing with this gift of freedom? Not much, I’m ashamed to say. Post-Kroka Contemplations Unknown — Mon, 20 Jun 2022 00:00:00 +0000 Hi! Hello! Salutations! I’m back! I just returned from a five month long expedition. I skied and paddled, traveling almost every day, living outside and in community with twelve other students for a semester. I got back on June 12th, and since then I’ve been adjusting every day to being home and living in society. As I adjust and reflect, I realize I have a lot to say. A lot to say about my time at Kroka, about the world and the society I live in, and about the life I want to live. Some Artwork from the Past Year Unknown — Mon, 22 Nov 2021 00:00:00 +0000 Over the past year or so I’ve been using Pixelmator Pro, mostly for specific projects, but occasionally just for fun. Here are a few cool things I’ve made. They all started from photos of the real world, usually taken by me or my girlfriend. A Loose List Unknown — Tue, 16 Nov 2021 00:00:00 +0000 Stuff I Might Study In College (I Guess) </a> </h3> A lot of people I barely know ask me questions about my life, and I guess I try to answer them, but it’s always kinda awkward and confusing. A popular question, when they hear I’m applying to colleges right now, is “What schools are you applying to?” That one is easy enough to answer: I tell them I’m applying to MIT early action, followed by Harvard, Yale, Dartmouth, Williams, and Amherst, with UMass Amherst as a safety. But when they hear I want to go to MIT, they say, “Ah, he must be a STEM guy,” and then make a bunch of assumptions about me. Or they know that I’m not the biggest lover of technology walking this earth and so they say - “wait - why do you want to go to MIT? A Few Movies Unknown — Fri, 12 Nov 2021 00:00:00 +0000 I’ve been thinking about movies that have really stuck with me, and I noticed that while the majority of them are well known, there are a few lesser known ones I thought I’d write about. Arden Lloyd's New EP Unknown — Fri, 12 Nov 2021 00:00:00 +0000 My friend Arden</a> just released her first EP! Listen to it on your preferred platform by following this link</a>, or stay right here! This Moment by Arden Lloyd</a></iframe> A Quote from Cornel West and Jeremy Tate Unknown — Wed, 27 Oct 2021 00:00:00 +0000 A quote from Howard University’s removal of classics is a spiritual catastrophe by Cornel West and Jeremy Tate. Published in full here</a> by the Washington Post. Quotes about Vegetarians Unknown — Fri, 01 Oct 2021 00:00:00 +0000 From The House of the Spirits by Isabelle Allende: “His father… ordered Nicolás to take a bath and put on some normal clothes if he wanted to stay in the house, but Nicolás stared at him without seeing and did not reply. He had become a vegetarian. He did not eat milk, meat, or eggs. His diet was the same as a rabbit’s, and his anxious face gradually came to resemble the face of that animal.” </blockquote> Maps Unknown — Fri, 24 Sep 2021 00:00:00 +0000 All that is gold does not glitter, Not all those who wander are lost; The old that is strong does not wither, Deep roots are not reached by the frost. </blockquote> ~ from The Lord of the Rings, by J.R.R. Tolkien This map is one of my biggest projects. Click on the full screen button for a better viewing interface in a new tab. I am still adding details and photos to this map - if you have something for me to add, let me know</a>! Kip's Wedding Song To Lafawnduh Unknown — Fri, 24 Sep 2021 00:00:00 +0000 From Napoleon Dynamite </a> </h3> </audio> Marc Simont's Illustrations for Thurber's 'The 13 Clocks' Unknown — Thu, 23 Sep 2021 00:00:00 +0000 Look at these gorgeous illustrations</h2> This is a great way to get a sense of your character’s appearance and physicality, as well as Thurber and Simont’s aesthetic. </a> How to Make a Healthier Nation of Omnivores Unknown — Wed, 02 Dec 2020 00:00:00 +0000 At my family’s small lowbush blueberry farm in late July and early August, the energy is festive. Over a thousand customers come from around the valley with their friends and small children each year to participate in harvesting the year’s blueberries, and most of them have a personal connection to the farm and the land, if not to my family directly. The blueberry sorting barn is filled with the roaring of fans and conveyor belts as the fruit is winnowed and packaged into boxes of five, ten, or twenty pounds and sold directly a few yards away; out in the picturesque fields, customers pick their own berries alongside local teenagers harvesting thousands of pounds of blueberries each day for what is often their very first job. The fruit is firm and sweet, easily surpassing in quality the conventionally grown lowbush blueberries, usually from Canada or Maine, found in the frozen section at supermarkets; and healthy-minded mothers excitedly discuss the antioxidant properties of the lowbush variety while tussling the blueberry-stained hair of their toddlers. Nativity Unknown — Fri, 29 May 2020 00:00:00 +0000 12/18/52 The loaf felt soft in his hand. Not soft and soggy like the stale bread crusts he had eaten all his life became when it rained—it was warm, and fresh. And the smell! It was the same smell that had made his stomach rumble every time he walked down the alley behind the tavern’s bakery—but now he experienced it a completely different way, knowing that the sweet smell was his, that he would sink his teeth into that warm, soft loaf. He looked behind him with a fearful grimace, but wild and delighted. No one in the busy street was looking at him. He reached up and over an edge in the wall, an edge he was too short to see but which his fingers knew perfectly—and he swung himself onto the tavern’s roof. A Sestina Unknown — Thu, 28 Mar 2019 00:00:00 +0000 By Satya Benson </a> </h2> Lying in bed, imagining all the wonderful creations To come, crying tears inside because they won’t, I can’t fight, I am too weak even though I can be strong and eat the fruit Of labour I won’t, I will not though I want, though I must walk, But on the outside, where I can see, I don’t yet drink Those tears, those bitter tears, not till I’ve pushed through the thick

Conclusion </a> </h2>
It’s so good I wrote it up so you better believe me.</p>

From Napoleon Dynamite </a> </h3>

</audio></p>

Look at these gorgeous illustrations</h2>
This is a great way to get a sense of your character’s appearance and physicality, as well as Thurber and Simont’s aesthetic.
</a> </p> </p>

Satya Benson

Modeling a Constant-Compute Automated AI R&D Process

Moving Past the Question of Consciousness: A Thought Experiment

Normative Eliminativism: What I Think About Ethics

Dialogue on Emergent Capitalism and Epistemics

How Slow Are Humans Compared to Computers? Comparing Temporal Resolution

On Academic LLM Usage

Stellar Bean and Corn Soup

Natural Latents and Aesthetic Categorization

My Senior Thesis Project

Beginning to think about aesthetic categorization decisions

Help me choose a thesis topic

Opinion: We Must Disregard Attacks on Free Speech

Thoughts on Misfin

Descartes Was Wrong About Minds

Thoughts on Veganism: A Reply

Aesthetics First

Rain Down

Summertime

Post-Kroka Contemplations

Some Artwork from the Past Year

A Loose List

A Few Movies

Arden Lloyd's New EP

A Quote from Cornel West and Jeremy Tate

Quotes about Vegetarians

Maps

Kip's Wedding Song To Lafawnduh

Marc Simont's Illustrations for Thurber's 'The 13 Clocks'

How to Make a Healthier Nation of Omnivores

Nativity

A Sestina