Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Bugs or Features?, published by qbolec on September 3, 2022 on LessWrong.

We often laugh over human-specific "bugs" in reasoning, comparing it to a gold standard of some utility-maximizing perfect Bayesian reasoner.

We often fear that a very capable AI following strict rules of optimization would reach some repugnant conclusions, and struggle to find "features" to add to guard against it.

What if some of the "bugs" we are looking at, are actually the "features" we are looking for?

We seem to distinguish "sacred" and "non-sacred" values, refusing to mix the two in calculations (for example human life vs money). What if this "tainted bit", "NaN-propagation", is a security feature guarding against Goodharting leading to genocide or dissolution of social trust? What if utility is not a single real number, but instead a pair? What if the ordering is not even lexicographic, but partial? What if it's a much longer tuple? Which brings me to next point:

We often experience decision paralysis apparently unable to compare two actions. What if this is simply because the order must be partial for security reasons? An alternative explanation of this phenomenon is that we implicitly treat "wait for more data to arrive and/or situation to change in tie-braking way" as an action available to us - is that bad?

We often decide which of the two end-states A vs B we prefer based on the path leading to them, amusingly favoring A in some scenarios and B in others. What if this is because we implicitly assume that end-state contains our brain with the memory of the path leading there? Isn't this a cool feature to treat the agent as part of its environment? Or what if this is because we implicitly factor in considerations of "what if other society members would follow this kind of path, or decision-making algorithm?". Isn't this a cool feature to think about second-order effects, about "acausal trades", and to treat own software as perhaps shared with other agents?

At least some long-term stable cultures have norms requiring children to follow adults' advice even if it conflicts their own judgment and more importantly said children apparently follow along instead of revolting and doing what (seems) good for them. Isn't that corrigibility a feature we want from AIs we plan to rear? Shouldn't there be safe-guards against child-knowing-better-than-parent in any self-modifying system spawning new generations of itself?

The whole sunken cost fallacy/heuristic. Isn't it actually a good thing to associate cost with each deviation from the original plan? Do we really want to zig-zag between more and more shiny objects with no meta-level realization that there's something wrong with this whole algorithm in general if it can't keep its trajectory predictable to itself? Yeah, sunken cost is more than that - it's not just fixed additional cost of decision - it's more like the guilt for not caring about your past self being invested into something. But again, isn't that a good thing from a security perspective?

I anticipate that each of these examples can be laughed at using some toy problem simple enough to calculate on the napkin. Sure, but we are talking about producing agents with partial information about very fuzzy world around them with lots of other agents embedded in them, some of them sharing goals or even parts of source code - we will rarely meet spherical cows on our way and overfitting to these learning examples is the very problem we want to solve. Do we really plan to solve all of that with a single simple elegant formula (AIXI style), or the plan always was to throw in some safety heuristics to the mix? If it's the later, then perhaps we can take a hint from parents raising children, societies avoiding dissolving, and people avoiding mania? Thus, what I propose is to take a look at the l...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AXRP Episode 18 - Concept Extrapolation with Stuart Armstrong, published by DanielFilan on September 3, 2022 on The AI Alignment Forum.

Google Podcasts link

Concept extrapolation is the idea of taking concepts an AI has about the world - say, “mass” or “does this picture contain a hot dog” - and extending them sensibly to situations where things are different - like learning that the world works via special relativity, or seeing a picture of a novel sausage-bread combination. For a while, Stuart Armstrong has been thinking about concept extrapolation and how it relates to AI alignment. In this episode, we discuss where his thoughts are at on this topic, what the relationship to AI alignment is, and what the open questions are.

Topics we discuss:

What is concept extrapolation

When is concept extrapolation possible

A toy formalism

Uniqueness of extrapolations

Unity of concept extrapolation methods

Concept extrapolation and corrigibility

Is concept extrapolation possible?

Misunderstandings of Stuart’s approach

Following Stuart’s work

Daniel Filan: Hello, everybody. In this episode, I’ll be speaking with Stuart Armstrong. Stuart was previously a senior researcher at the Future of Humanity Institute at Oxford, where he worked on AI safety and x-risk, as well as how to spread between galaxies by disassembling the planet Mercury. He’s currently the head boffin at Aligned AI, where he works on concept extrapolation, the subject of our discussion. For links to what we’re discussing, you can check the description of this episode, and you can read the transcript at axrp.net. Well, Stuart, welcome to the show.

Stuart Armstrong: Thank you.

Daniel Filan: Cool.

Stuart Armstrong: Good to be on.

What is concept extrapolation

Daniel Filan: Yeah, it’s nice to have you. So I guess the thing I want to be talking about today is your work and your thoughts on concept extrapolation and model splintering, which I guess you’ve called it. Can you just tell us: what is concept extrapolation?

Stuart Armstrong: Model splintering is when the features or the concepts on which you built your goals or your reward functions break down. Traditional examples are in physics when (for instance) the ether disappeared. It didn’t mean when the ether disappeared that all the previous physics that had been based on ether suddenly became completely wrong. You had to extend the old results into a new framework. You had to find a new framework and you had to extend it. So model splintering is when the model falls apart or the features or concepts fall apart and concept extrapolation is what you do to extend the concept across that divide.

Daniel Filan: Okay.

Stuart Armstrong: Like there was a concept of energy before relativity, and there’s a concept of energy after relativity. They’re not exactly the same thing, but there’s a definite continuity to it.

Daniel Filan: Cool. So you mentioned that at some point we used to think there was ether, and now we think there isn’t. What’s an example of a concept or something that splintered when we realized there wasn’t an ether anymore, just to get a really concrete example.

Stuart Armstrong: Maxwell’s equations - Maxwell’s non-relativistic equations are based on a non-constant speed of light. Maxwell’s equations are not relativistic, though they have a relativistic formulation.

Daniel Filan: Hang on. I thought they were. Isn’t that why you get the constant speed of light out of them?

Stuart Armstrong: Okay. If that’s uncertain, then let’s try another example.

Daniel Filan: We could do energy, once you discovered general relativity, or.

Stuart Armstrong: Energy, inertial mass, for example: those concepts needed a Newtonian universe to make sense, when it wasn’t so much the absence of ether, but it was the surprisingly constant speed of light that broke those. So when you m...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: An Update on Academia vs. Industry (one year into my faculty job), published by David Scott Krueger on September 3, 2022 on The AI Alignment Forum.

I've been an assistant professor (equivalent) for ~1 year now at Cambridge. Shortly after accepting the position, I wrote AI x-risk reduction: why I chose academia over industry. Since then, I've had a lot of conversations on academia vs. industry with people getting into AI x-safety (e.g. considering applying for PhDs). This post summarizes that experience, and describes a few other updates from my experience in the last 1.5 years.Summary of recent conversations:

Most people haven't read my previous post and I point them to it.

A common attitude is: "Industry seems better, WTF would you do academia right now?"

A perhaps equally common attitude is: "I am new and have no strong opinions, just trying to figure out what people think."

The main reasons I hear for industry over academia are:

Short timelines

Need to access Foundation models

Academic incentives to work on topics less relevant to x-safety.

I think these are all valid points, but I there are countervailing considerations. My response to these 3 points has generally been something like:

Academia still seems like the best option ATM for rapidly training and credentialing people. Even under fairly short timelines, this seems likely to be more valuable than direct work. Furthermore, it is a mistake to simply focus on efforts on whatever timelines seem most likely; one should also consider tractability and neglectedness of strategies that target different timelines. It seems plausible that we are just screwed on short timelines, and somewhat longer timelines are more tractable. Also, people seem to be making this mistake a lot and thus short timelines seem potentially less neglected.

There are and will be open source foundation models. Organizations like Anthropic seem keen on collaborating and providing access (although we haven't yet pitched them anything concrete). It's not clear that you have that much better access when you are working out one of these orgs (I'd be curious what people at the orgs think about this one!): my impression is that it is still clunky to do a lot of things with a large model when you are at an org, and things like retraining are obviously very expensive; these orgs seem to also favor large group projects, which I assume are directed by leadership; so in practice, there might be less difference between being entry-level at an org vs. being in academia.

The incentives are real, and may be worth playing into somewhat if you want to get a faculty job. However, it is becoming easier and easier to work on safety in academia; safety topics are going mainstream and being published at top conferences. Work that is still outside the academic Overton window can be brought into academia if it can be approached with the technical rigor of academia, and work that meets academic standards is much more valuable than work that doesn't; this is both because it can be picked up by the ML community, and because it's much harder to tell if you are making meaningful progress if your work doesn't meet these standards of rigor. There's also a good chance that within ~5 years they will be mainstream enough that it's now a good career move to focus on safety as an academic. Also, depending on your advisor, you can basically have total freedom in academia (and funding tends to give freedom).

Finally, the incentives outside of academia are not great either: for-profit orgs are incentivized to build stuff, whether it's a good idea or not; because LW/AF do not have established standards of rigor like ML, they end up operating more like a less-functional social science field, where (I've heard) trends, personality, and celebrity play an outsized role in determining whi...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Chesterton Fences and EA’s X-risks, published by jehan on September 2, 2022 on The Effective Altruism Forum.

TL;DR: EA lacks the protective norms nearly universal in mature institutions, leaving it vulnerable to the two leading causes of organizational sudden death: corruption and sex scandals.

A foundational belief in EA is the importance of unintended negative consequences; a famous actionable heuristic around this is Chesterton’s Fence:

Let us say, for the sake of simplicity, a fence or gate erected across a road. The more modern type of reformer goes gaily up to it and says, "I don't see the use of this; let us clear it away." To which the more intelligent type of reformer will do well to answer: "If you don't see the use of it, I certainly won't let you clear it away. Go away and think. Then, when you can come back and tell me that you do see the use of it, I may allow you to destroy it.

It’s not to say that you can never take down fences, only that you recognize what you’re losing so that you can weigh it against what you stand to gain. However, I believe EA has carelessly dismantled valuable fences while gaining comparatively little. At the very least there are low-hanging fruit fences that could have outsized returns with few costs.

1) Dismantled Fence: Undervaluing Outside Experience

This has been discussed extensively, so I won't comment further except to note that outside experience is a rite of passage in a wide range of communities.It may be that this general lack of outside experience has prevented EA from adopting hard-won norms found elsewhere:

2) Dismantled Fence: Excessive Fraternization

Meme in EA: All your friends, roommates and partners are EAs.

This is not a good thing. Many have commented, including Will MacAskill, that this is personally suboptimal and have tried to correct it. The EA-consistent reason for this seems to either be work-life balance or viewpoint diversity.

These reasons are valid, but not the reason nearly every successful organization has rules around this: because it's an existential risk for The Mission.Not in some second-order way either, excessively close and incestuous relationships cause dysfunction directly and in particular leave it open to perhaps the modern era's leading cause of organizational sudden death: sex scandals.Successful institutions have rules around fraternization and power relationships, EA seems to largely lack explicit norms here-if anything the norm seems to actually go the other way from the mainstream.I understand that relationships are what makes a community, but bad relationships will make a bad community while also endangering the mission.

3) Dismantled Fence: Lack of Explicit Anti-Corruption Rules

You may read “corruption” and think “graft”, but that's typically the last rung on the corruption ladder. It's not so overt or conscious at first, so more like self-dealing, conflict of interest, and nepotism.Few explicit rules address the problems above, and there is at least the perception that within EA they exist (but this perception could be mitigated via visible norms).

Mature institutions have rules preventing these sorts of failure modes. EA however is now combining fraternization with few rules around self-dealing and favoritism to create a massive opportunity for organizational death with no comparable upside.

Fences as an Immune system

These fences not only prevent existing members from inadvertently sliding into scandal, but also act as an immune system for bad actors.

EA now has an enormous amount of power, money, and influence. When it was small and unappealing it may have gotten away with security through obscurity, but it's now far too big of a prize to exist so immunocompromised.(Cross-posted here: www.atvbt.com/eacrit)

Examples: Mormon missionaries, family businesses, New Zealand

"But...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Igor Kiriluk (1974-2022), published by turchin on September 3, 2022 on The Effective Altruism Forum.

Igor was an organiser the first meet-up in Moscow about effective altruism around 2013. Today his body was found at his home. The day before he complained about depression and bad health. His cryopreservation now is being organised.

He was also a one of four organisers of Russian Transhumanist Movement, along with Danila Medvedev, Valeria Pride and Igor Artuhov around 2003.

His main topic of interest was paradise-engineering. He translated works of David Pearce.

He may look detached from reality but he was first to react on new ideas and has very large network of friends everywhere: between visionaries, scientists and officials. Being a great networker, he helped many people to find each other, especially in the field of life extension.

His FB page:

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The discount rate is not zero, published by Thomaaas on September 3, 2022 on The Effective Altruism Forum.

(Note: I privately submitted this essay to the red-teaming contest before the deadline passed, now cross-posting here. Also, this is my first ever post, so please be nice.)

Summary

Longtermists believe that future people matter, there could be a lot of them, and they are disenfranchised. They argue a life in the distant future has the same moral worth as somebody alive today. This implies that analyses which discount the future unjustifiably overlook the welfare of potentially hundreds of billions of future people, if not many more.

Given the relationship between longtermism and views about existential risk, it is often noted that future lives should in fact be discounted somewhat – not for time preference, but for the likelihood of existing (i.e., the discount rate equals the catastrophe rate).

I argue that the long-term discount rate is both positive and inelastic, due to 1) the lingering nature of present threats, 2) our ongoing ability to generate threats, and 3) continuously lowering barriers to entry. This has 2 major implications.

First, we can only address near-term existential risks. Applying a long-term discount rate in line with the long-term catastrophe rate, by my calculations, suggests the expected length of human existence is another 8,200 years (and another trillion people). This is significantly less than commonly cited estimates of our vast potential.

Second, I argue that equally applying longtermist principles would consider the descendants of each individual, when lives are saved in the present. A non-zero discount rate allows us to calculate the expected number of a person’s descendants. I estimate 1 life saved today affects an additional 93 people over the course of humanity’s expected existence.

Both claims imply that x-risk reduction is overweighted relative to interventions such as global health and poverty reduction (but I am NOT arguing x-risks are unimportant).

Discounting & longtermism

Will MacAskill summarised the longtermist ideology in 3 key points: future people matter (morally), there are (in expectation) vast numbers of future people, and future people are utterly disenfranchised.Future people are disenfranchised in the sense that they cannot voice opinion on matters which affect them greatly, but another way in which they are directly discriminated against is in the application of discount rates.

Discounting makes sense in economics, because inflation (or the opportunity to earn interest) can make money received earlier more valuable than money obtained later. This is called “time preference”, and it is a function of whatever discount rate is applied. While this makes sense for cashflows, human welfare is worth the same regardless of when it is expressed. Tyler Cowen and Derick Parfit first argued this point, however, application of a “social” discount rate is widely accepted and applied (where the social discount rate is derived from the “social rate of time preference”).

Discounting is particularly important for longtermism, because the discount factor applied each year accumulates over time (growing exponentially), which can lead to radical conclusions over very long horizons. For example, consider the welfare of 1 million people, alive 1500 years in the future. Applying a mere 1% discount rate implies the welfare of this entire population worth less than 1/3 of the value of a person alive today. Lower discount rates only delay this distortion – Tarsney (2020) notes that in the long run, any positive discount rate “eventually wins”. It is fair to say that people in the distant future are “utterly disenfranchised”.

Existential risk

Longtermism implies that we should optimize our present actions to maximize the chance that fu...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Peter Eckersley (1979-2022), published by Gavin on September 3, 2022 on The Effective Altruism Forum.

Security engineer, digital rights activist, AI safety and policy researcher. Beloved in these communities.

Eckersley is most famous as an advocate/developer at the intersection of tech and legal activism. His work on the Let's Encrypt free certificate authority, HTTPS Everywhere, and the SSL Observatory made the internet significantly less terrible. Mizza:

Let's Encrypt is something we all came to take for granted very quickly, but lots of us remember when getting an SSL certificate was an expensive and tedious process. Deprecating a billion dollar industry overnight and providing better security for internet users everywhere is a hell of a legacy to leave behind, and I hope one that will be an inspiration for generations to come.

He identified "device fingerprinting" as a major privacy hole and spent a long time trying to mitigate it. His PhD thesis was a characteristic mix of software engineering, economic theory, and law. We'd call it Web3 these days.

He was the 8th person ever to take the GWWC pledge. A few places say he was on the board of CEA US at some point.

Around 2017 he went into AI risk. His work bridged AI ethics and AI safety, covering recidivism prediction, self-driving cars, security against ML hacking, and military AI but also big-brain topics like impossibility theorems in social choice and ML, the cybersecurity implications of emulated minds, and the queer theory of human alignment. At the end, he was running the AI Objectives Institute (notable for mobilising people already concerned about corporate maximisation) and mentoring for PIBBSS. He was on many of the big "technical AI policy" papers of the last few years. The Alignment Problem is dedicated to him, after he convinced his friend Brian Christian of it.

During covid he worked intensely on private contact tracing algorithms.

He received a cancer diagnosis a week before his death. A minor consolation of dying in the present age is people telling beautiful stories about you for strangers and for posterity. So you can know that he was fluent in French, that he organised guerilla meetups, that he was an avid cyclist, that he was a memorable remedial teacher of programming. This is an Eckersley quotes account. Similarly, I can tell you he was writing code until one month before his death. He lived sans LinkedIn.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: We may be able to see sharp left turns coming, published by Ethan Perez on September 3, 2022 on The AI Alignment Forum.

There's a lot of discourse around abrupt generalization in models, most notably the "sharp left turn." Most recently, Wei et al. 2022 claim that many abilities suddenly emerge at certain model sizes. These findings are obviously relevant for alignment; models may suddenly develop the capacity for e.g. deception, situational awareness, or power-seeking, in which case we won't get warning shots or a chance to practice alignment. In contrast, prior work has also found "scaling laws" or predictable improvements in performance via scaling model size, data size, and compute, on a wide variety of domains. Such domains include transfer learning to generative modeling (on images, video, multimodal, and math) and reinforcement learning. What's with the discrepancy?

One important point is the metric that people are using to measure capabilities. In the BIG Bench paper (Figure 7b), the authors find 7 tasks that exhibit "sharp upwards turn" at a certain model size.

Naively, the above results are evidence for sharp left turns, and the above tasks seem like some of the best evidence we have for sharp left turns. However, the authors plot the results on the above tasks in terms of per-character log-likelihood of answer:

The authors actually observe smooth increases in answer log-likelihood, even for tasks which showed emergent behavior according to the natural performance metric for the task (e.g. accuracy). These results are evidence that we can predict that emergent behaviors will occur in the future before models are actually "capable" of those behaviors. In particular, these results suggest that we may be able to predict power-seeking, situational awareness, etc. in future models by evaluating those behaviors in terms of log-likelihood. We may even be able to experiment on interventions to mitigate power-seeking, situational awareness, etc. before they become real problems that show up in language model -generated text.

Clarification: I think we can predict whether or not a sharp left turn towards deception/misalignment will occur rather than exactly when. In particular, I think we should look at the direction of the trend (increases vs. decreases in log-likelihood) as signal about whether or not some scary behavior will eventually emerge. If the log likelihood of some specific scary behavior increases, that’s a bad sign and gives us some evidence it will be a problem in the future. I mainly see scaling laws here as a tool for understanding and evaluating which of the hypothesized misalignment-relevant behaviors will show up in the future. The scaling laws are useful signal for (1) convincing ML researchers to worry about scaling up further because of alignment concerns (before we see them in model behaviors/outputs) and (2) guiding alignment researchers with some empirical evidence about which alignment failures are likely/unlikely to show up after scaling at some point.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Sticky goals: a concrete experiment for understanding deceptive alignment, published by Evan Hubinger on September 2, 2022 on The AI Alignment Forum.

Thanks to Sam McCandlish for a useful conversation that prompted this post.

In “How likely is deceptive alignment?”, I lay out the case for why you might expect deceptive alignment to occur in practice for two different stories of ML inductive biases: high and low path-dependence. In this post, I want to propose an experiment that I think would shed substantial light on the high path-dependence case specifically.

In the high path-dependence case, the story for why you would get deceptive alignment is essentially as follows:

The model learns some set of proxy goals.

The model learns enough about the training process that, if it were to use that knowledge to directly optimize for what the training process is trying to get it to do, it would get better performance than just using its proxy goals.

Gradient descent modifies the model’s proxies to become long-term goals, making the model deceptive such that it starts optimizing directly for its understanding of what the training process wants and thus gets better performance.

However, as I talk about in “How likely is deceptive alignment?”, there is another, alternative option available to gradient descent for step (3): instead of turning the existing proxies into long-term goals, gradient descent could just replace them entirely with a pointer to the model’s understanding of the training objective. In my opinion, I think it is currently quite unclear which one of these gradient descent would prefer.

One way to think about these two options is that gradient descent is effectively choosing between: repurposing the existing terminal goal to care about the training objective instrumentally vs. modifying the existing terminal goal to care about the training objective terminally. Thus, we can conceptualize the question of which would be most likely to happen as: how “sticky” are an agent’s existing terminal proxies?

Now, one thing I will say is that, since the difficulty of changing the proxies into long-term goals seems to be essentially constant regardless of the complexity of the training objective—whereas the difficulty of creating a pointer to the training objective scales with that complexity—I think that, if we try to train models on increasingly complex goals, eventually deception will win almost regardless of the baseline “stickiness” level. But understanding that baseline could still be quite important, and it’s something that I think we can quite readily experiment with right now.

Furthermore, understanding of goal stickiness could also help shed additional light on the broader question of whether machine learning inductive biases look more like the high or low path-dependence world to begin with. If we see very little goal stickiness, that might just be because essentially nothing is sticky and we shouldn’t be conceptualizing anything as path-dependent to begin with.

Here’s an example experimental setup for trying to understand goal stickiness:

Train an agent (via RL, via a decision transformer, etc.) to accomplish some complex goal in a simulated environment requiring intermediate resources.

Concrete example: Train an RL agent to build diamond houses in a particular Minecraft chunk.

Fine-tune in an environment with two distinct stages. In stage 1, fine-tune the agent to get as much of the intermediate resource as possible and prevent the agent from being able to achieve the complex goal. In stage 2, provide no rewards (except maybe something like next state prediction) and allow the agent to do whatever it wants. Iterate.

Concrete example: Fine-tune the agent to get as many diamonds as possible while being locked out of the special chunk in stage 1, then let it do wha...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: The impact we achieved to date: Animal Advocacy Careers, published by SofiaBalderson on September 2, 2022 on The Effective Altruism Forum.

Hi all,

On behalf of the team at AAC, I wanted to share a new page that summarises our impact to date. It includes a breakdown of the impact we achieved via the following programs:

Introduction to Animal Advocacy Online Course - 50+ new relevant positions gained by participants (such as jobs, internships, and volunteer placements)

Job board - 7 people have found their roles there so far, and we get 1300+ monthly unique visitors

Fundraising Work Placement - 9 placements at 3 months FTE each in 2022

Career advice service - 20+ new relevant positions gained

Recruitment service - 4 hires

Let us know if you have any questions!

For those who doesn't know us: Animal Advocacy Careers (AAC) is an organisation that seeks to address the career and talent bottlenecks in the animal advocacy movement, especially the farmed animal movement. We are providing careers services for individuals at all levels of experience with animal advocacy.

We have room for funding and are funding-constrained. If you would like to support us, please donate here or reach out to us at [email protected].

Sofia and team AACInternally, we're relying more on our ICAPs system (see 2022 plans post for explanation), but it's complex so we also wanted to share some easier-to-grasp numbers with stakeholders.

Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.