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: Moore's Law, AI, and the pace of progress, published by Veedrac on December 11, 2021 on LessWrong.

It seems to be a minority view nowadays to believe in Moore's Law, the routine doubling of transistor density roughly every couple of years, or even the much gentler claim, that There's Plenty [more] Room at the Bottom. There's even a quip for it: the number of people predicting the death of Moore's law doubles every two years. This is not merely a populist view by the uninformed. Jensen Huang, CEO of NVIDIA, a GPU company, has talked about Moore's Law failing.

"Moore's Law used to grow at 10x every five years [and] 100x every 10 years," Huang said during a Q&A panel with a small group of reporters and analysts at CES 2019. "Right now Moore's Law is growing a few percent every year. Every 10 years maybe only 2s. ... So Moore's Law has finished."

More academically, the International Roadmap for Devices and Systems, IRDS, warns that logic scaling is nearing certain fundamental limits.

After the “1.5nm” logic node goes into production in 2028, logic dimensions will stop shrinking and improved densities will be achieved by increasing the number of devices vertically. DRAM will continue to shrink CDs [Critical Dimensions] after that, but the minimum lines and spaces will only shrink modestly and should be reachable by improved EUV and EUV double patterning. The large number of masking levels and the many steps for 3D stacking of devices will make yield and cost high priorities.

This claim is not based on tooling limits, but a projected minimum useful size of transistors.

Lines and spaces are the flagship pattern of lithography. [...] Note that the logic node names are the commonly used names for each node but are not the same as the minimum half pitches of those nodes. Resolution improves to 12 nm half-pitch in 2022. This corresponds to the logic “3 nm” node. The IRDS expects that this resolution will be achieved through EUV double patterning. Then there is a further decrease in line and space resolution of 2 nm per node until 2028, when minimum line and space resolution is expected to reach 8 nm half-pitch. The 8 nm half pitch could be achieved with EUV double patterning, but there is time to develop other methods also, such as high-NA EUV lithography. After that, no further improvement in required resolution is projected, although this is due to projected device requirements, not expected limitations in patterning capability.

Computers are made of stacks of wires in a dense 3D network, and line and space pitch is a measure of how close parallel lines can be packed.

Besides mere physical inevitability, improvements to transistor density are taking an economic toll. Building the fabs that manufacture transistors is becoming very expensive, as high as $20 billion each, and TSMC expects to spend $100 billion just over the three years to expand capacity. This cost increases with each cutting-edge node.

This bleak industry view contrasts with the massively increasing demands of scale from AI, that has become a center of attention, in large part due to OpenAI's attention on the question, and their successful results with their various GPT-derived models. There, too, the economic factor exacerbates the divide; models around GPT-3's size are the domain of only a few eager companies, and whereas before there was an opportunity to reap quick advances from scaling single- or few-machine models to datacenter scale, now all compute advances require new hardware of some kind, whether better computer architectures or bigger (pricier) data centers.

The natural implication is that device scaling has already stalled and will soon hit a wall, that scaling out much further is uneconomical, and in conclusion that AI progress cannot be driven much further through scaling, certainly not soon, and...

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: Moore's Law, AI, and the pace of progress, published by Veedrac on December 11, 2021 on LessWrong.

It seems to be a minority view nowadays to believe in Moore's Law, the routine doubling of transistor density roughly every couple of years, or even the much gentler claim, that There's Plenty [more] Room at the Bottom. There's even a quip for it: the number of people predicting the death of Moore's law doubles every two years. This is not merely a populist view by the uninformed. Jensen Huang, CEO of NVIDIA, a GPU company, has talked about Moore's Law failing.

"Moore's Law used to grow at 10x every five years [and] 100x every 10 years," Huang said during a Q&A panel with a small group of reporters and analysts at CES 2019. "Right now Moore's Law is growing a few percent every year. Every 10 years maybe only 2s. ... So Moore's Law has finished."

More academically, the International Roadmap for Devices and Systems, IRDS, warns that logic scaling is nearing certain fundamental limits.

After the “1.5nm” logic node goes into production in 2028, logic dimensions will stop shrinking and improved densities will be achieved by increasing the number of devices vertically. DRAM will continue to shrink CDs [Critical Dimensions] after that, but the minimum lines and spaces will only shrink modestly and should be reachable by improved EUV and EUV double patterning. The large number of masking levels and the many steps for 3D stacking of devices will make yield and cost high priorities.

This claim is not based on tooling limits, but a projected minimum useful size of transistors.

Lines and spaces are the flagship pattern of lithography. [...] Note that the logic node names are the commonly used names for each node but are not the same as the minimum half pitches of those nodes. Resolution improves to 12 nm half-pitch in 2022. This corresponds to the logic “3 nm” node. The IRDS expects that this resolution will be achieved through EUV double patterning. Then there is a further decrease in line and space resolution of 2 nm per node until 2028, when minimum line and space resolution is expected to reach 8 nm half-pitch. The 8 nm half pitch could be achieved with EUV double patterning, but there is time to develop other methods also, such as high-NA EUV lithography. After that, no further improvement in required resolution is projected, although this is due to projected device requirements, not expected limitations in patterning capability.

Computers are made of stacks of wires in a dense 3D network, and line and space pitch is a measure of how close parallel lines can be packed.

Besides mere physical inevitability, improvements to transistor density are taking an economic toll. Building the fabs that manufacture transistors is becoming very expensive, as high as $20 billion each, and TSMC expects to spend $100 billion just over the three years to expand capacity. This cost increases with each cutting-edge node.

This bleak industry view contrasts with the massively increasing demands of scale from AI, that has become a center of attention, in large part due to OpenAI's attention on the question, and their successful results with their various GPT-derived models. There, too, the economic factor exacerbates the divide; models around GPT-3's size are the domain of only a few eager companies, and whereas before there was an opportunity to reap quick advances from scaling single- or few-machine models to datacenter scale, now all compute advances require new hardware of some kind, whether better computer architectures or bigger (pricier) data centers.

The natural implication is that device scaling has already stalled and will soon hit a wall, that scaling out much further is uneconomical, and in conclusion that AI progress cannot be driven much further through scaling, certainly not soon, and...

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 Emergency Fund for Effective Altruists, published by bob on December 11, 2021 on The Effective Altruism Forum.

I am considering setting up a charity. Its goals are:

To help effective altruists in financial need by giving back part of their donations

To promote large donations to EA-related causes by providing donors with a safety net, minimizing risk

To foster a healthy EA community

To achieve this, the charity maintains an emergency fund. If effective altruists are in financial trouble, they can recoup up to 50% of their EA-related donations from that fund.

By default, the charity acts as a proxy for GiveWell. That is, if you donate $100 to the charity, we will put $10 in the shared emergency fund and forward $90 to GiveWell. Then, if you ever need $50 back, we can easily verify that you donated $100 in total.

Direct donors (i.e. people who did not donate through to the fund, but donated directly to GiveWell or its recommended charities) are also eligible for support. However, as a fraud prevention measure, direct donors will have to get their charity to state to us that they will never return any funds to the donor, and we will need a confirmation of the exact amount they donated. Of course, if a lot of people ask for support without contributing to the fund, we will go broke. So please donate through the fund once it exists.

If our balance declines, we might ask for donations that go entirely to the fund. If we ever cannot help everyone who needs help, we will proactively raise funds wherever we can.

The recoupment process will be similar to filing an insurance claim, but less strict. If someone who donated through the fund asks for a fraction of what they donated in total, they will probably get their money back without any friction. We do not have to worry about fraud as much as insurers, because donors can only recoup half of what they originally donated.

As far as I know, this type of charity does not exist yet, and there are a lot of things that still need to be figured out. More extensive discussion can be found here (me) and here (Alicorn). Note that both linked essays discuss a fund that only offers support to contributors (excluding direct donors). I no longer think that is desirable, and it will be very difficult to convince governments to grant such a construction tax-exempt status.

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: An Emergency Fund for Effective Altruists, published by bob on December 11, 2021 on The Effective Altruism Forum.

I am considering setting up a charity. Its goals are:

To help effective altruists in financial need by giving back part of their donations

To promote large donations to EA-related causes by providing donors with a safety net, minimizing risk

To foster a healthy EA community

To achieve this, the charity maintains an emergency fund. If effective altruists are in financial trouble, they can recoup up to 50% of their EA-related donations from that fund.

By default, the charity acts as a proxy for GiveWell. That is, if you donate $100 to the charity, we will put $10 in the shared emergency fund and forward $90 to GiveWell. Then, if you ever need $50 back, we can easily verify that you donated $100 in total.

Direct donors (i.e. people who did not donate through to the fund, but donated directly to GiveWell or its recommended charities) are also eligible for support. However, as a fraud prevention measure, direct donors will have to get their charity to state to us that they will never return any funds to the donor, and we will need a confirmation of the exact amount they donated. Of course, if a lot of people ask for support without contributing to the fund, we will go broke. So please donate through the fund once it exists.

If our balance declines, we might ask for donations that go entirely to the fund. If we ever cannot help everyone who needs help, we will proactively raise funds wherever we can.

The recoupment process will be similar to filing an insurance claim, but less strict. If someone who donated through the fund asks for a fraction of what they donated in total, they will probably get their money back without any friction. We do not have to worry about fraud as much as insurers, because donors can only recoup half of what they originally donated.

As far as I know, this type of charity does not exist yet, and there are a lot of things that still need to be figured out. More extensive discussion can be found here (me) and here (Alicorn). Note that both linked essays discuss a fund that only offers support to contributors (excluding direct donors). I no longer think that is desirable, and it will be very difficult to convince governments to grant such a construction tax-exempt status.

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: [Linkpost] Chinese government's guidelines on AI, published by RomanS on December 10, 2021 on LessWrong.

China (PRC) published a new AI policy: "Ethical Norms for the New Generation Artificial Intelligence" (2021-09-26).

A surprisingly large part of the document's content is dedicated to the topics related to AI X-risk (although the existential risk is never addressed directly).

According to the document, it regulates AI research conducted by pretty much any entity in China, from companies to natural persons.

The original official document in Chinese is here (archive). Other AI policies are available here.

Some excerpts that could be relevant to AI X-risk:

AI activities should... follow the common values of mankind, respect... the fundamental interests of mankind,... promote human-machine harmony and friendliness.

Ensure that humans have full autonomy in decision-making,... the right to suspend the operation of artificial intelligence systems at any time,... ensure that artificial intelligence is always under human control.

Actively learn and popularize knowledge of AI ethics

objectively understand ethical issues, and not underestimate or exaggerate ethical risks

take the initiative to integrate AI ethics into the whole process of management

Strengthen risk prevention. Enhance bottom-line thinking and risk awareness, strengthen the study and judgment of potential risks of artificial intelligence development, carry out timely and systematic risk monitoring and assessment, establish an effective risk warning mechanism, and enhance the ability to control and dispose of ethical risks of artificial intelligence.

Strengthen the self-restraint of AI R&D-related activities, take the initiative to integrate AI ethics into all aspects of technology R&D, consciously carry out self-censorship, strengthen self-management, and not engage in AI R&D that violates ethics and morality.

In the algorithm design, implementation, application and other links, enhance transparency, interpretability, understandability, reliability, controllability

achieve verifiable, auditable, supervisable, traceable, predictable and trustworthy [AI].

Prohibit the use of artificial intelligence products.... that endanger national security, public safety

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: [Linkpost] Chinese government's guidelines on AI, published by RomanS on December 10, 2021 on LessWrong.

China (PRC) published a new AI policy: "Ethical Norms for the New Generation Artificial Intelligence" (2021-09-26).

A surprisingly large part of the document's content is dedicated to the topics related to AI X-risk (although the existential risk is never addressed directly).

According to the document, it regulates AI research conducted by pretty much any entity in China, from companies to natural persons.

The original official document in Chinese is here (archive). Other AI policies are available here.

Some excerpts that could be relevant to AI X-risk:

AI activities should... follow the common values of mankind, respect... the fundamental interests of mankind,... promote human-machine harmony and friendliness.

Ensure that humans have full autonomy in decision-making,... the right to suspend the operation of artificial intelligence systems at any time,... ensure that artificial intelligence is always under human control.

Actively learn and popularize knowledge of AI ethics

objectively understand ethical issues, and not underestimate or exaggerate ethical risks

take the initiative to integrate AI ethics into the whole process of management

Strengthen risk prevention. Enhance bottom-line thinking and risk awareness, strengthen the study and judgment of potential risks of artificial intelligence development, carry out timely and systematic risk monitoring and assessment, establish an effective risk warning mechanism, and enhance the ability to control and dispose of ethical risks of artificial intelligence.

Strengthen the self-restraint of AI R&D-related activities, take the initiative to integrate AI ethics into all aspects of technology R&D, consciously carry out self-censorship, strengthen self-management, and not engage in AI R&D that violates ethics and morality.

In the algorithm design, implementation, application and other links, enhance transparency, interpretability, understandability, reliability, controllability

achieve verifiable, auditable, supervisable, traceable, predictable and trustworthy [AI].

Prohibit the use of artificial intelligence products.... that endanger national security, public safety

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 Plan, published by johnswentworth on December 10, 2021 on LessWrong.

This is a high-level overview of the reasoning behind my research priorities, written as a Q&A.

What’s your plan for AI alignment?

Step 1: sort out our fundamental confusions about agency

Step 2: ambitious value learning (i.e. build an AI which correctly learns human values and optimizes for them)

Step 3: .

Step 4: profit!

. and do all that before AGI kills us all.

That sounds. awfully optimistic. Do you actually think that’s viable?

Better than a 50/50 chance of working in time.

Do you just have really long timelines?

No. My median is maybe 10-15 years, though that’s more a gut estimate based on how surprised I was over the past decade rather than a carefully-considered analysis. (I wouldn’t be shocked by another AI winter, especially on an inside view, but on an outside view the models generating that prediction have lost an awful lot of Bayes Points over the past few years.)

Mostly timelines just aren’t that relevant; they’d have to get down to around 18-24 months before I think it’s time to shift strategy a lot.

. Wat. Not relevant until we’re down to two years?!?

To be clear, I don’t expect to solve the whole problem in the next two years. Rather, I expect that even the incremental gains from partial progress on fundamental understanding will be worth far more than marginal time/effort on anything else, at least given our current state.

At this point, I think we’re mostly just fundamentally confused about agency and alignment. I expect approximately-all of the gains-to-be-had come from becoming less confused. So the optimal strategy is basically to spend as much time as possible sorting out as much of that general confusion as possible, and if the timer starts to run out, then slap something together based on the best understanding we have.

18-24 months is about how long I expect it to take to slap something together based on the best understanding we have. (Well, really I expect it to take <12 months, but planning fallacy and safety margins and time to iterate a little and all that.)

But iterative engineering is important!

In order for iterative engineering to be useful, we first need to have a strong enough understanding of what we even want to achieve in order to recognize when an iteration has brought us a step closer to the goal. No amount of A/B testing changes to our website will make our company profitable if we’re measuring the wrong metrics. I claim that, for alignment, we do not yet have a strong enough understanding for iteration to produce meaningful progress.

When I say “we’re just fundamentally confused about agency and alignment”, that’s the sort of thing I’m talking about.

To be clear: we can absolutely come up with proxy measures of alignment. The problem is that I don’t expect iteration under those proxy measures to get us meaningfully closer to aligned AGI. No reasonable amount of iterating on gliders’ flight-range will get one to the moon.

But engineering is important for advancing understanding too!

I do still expect some amount of engineering to be central for making progress on fundamental confusion. Engineering is one of the major drivers of science; failed attempts to build amplifiers drove our first decent understanding of semiconductors, for instance. But this is a very different path-to-impact than directly iterating on “alignment”, and it makes sense to optimize our efforts differently if the path-to-impact is through fundamental understanding. Just take some confusing concept which is fundamental to agency and alignment (like abstraction, or optimization, or knowledge, or .) and try to engineer anything which can robustly do something with that concept. For instance, a lot of my own work is driven by the vision of a “thermometer of abstraction”, a device capab...

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 Plan, published by johnswentworth on December 10, 2021 on LessWrong.

This is a high-level overview of the reasoning behind my research priorities, written as a Q&A.

What’s your plan for AI alignment?

Step 1: sort out our fundamental confusions about agency

Step 2: ambitious value learning (i.e. build an AI which correctly learns human values and optimizes for them)

Step 3: .

Step 4: profit!

. and do all that before AGI kills us all.

That sounds. awfully optimistic. Do you actually think that’s viable?

Better than a 50/50 chance of working in time.

Do you just have really long timelines?

No. My median is maybe 10-15 years, though that’s more a gut estimate based on how surprised I was over the past decade rather than a carefully-considered analysis. (I wouldn’t be shocked by another AI winter, especially on an inside view, but on an outside view the models generating that prediction have lost an awful lot of Bayes Points over the past few years.)

Mostly timelines just aren’t that relevant; they’d have to get down to around 18-24 months before I think it’s time to shift strategy a lot.

. Wat. Not relevant until we’re down to two years?!?

To be clear, I don’t expect to solve the whole problem in the next two years. Rather, I expect that even the incremental gains from partial progress on fundamental understanding will be worth far more than marginal time/effort on anything else, at least given our current state.

At this point, I think we’re mostly just fundamentally confused about agency and alignment. I expect approximately-all of the gains-to-be-had come from becoming less confused. So the optimal strategy is basically to spend as much time as possible sorting out as much of that general confusion as possible, and if the timer starts to run out, then slap something together based on the best understanding we have.

18-24 months is about how long I expect it to take to slap something together based on the best understanding we have. (Well, really I expect it to take <12 months, but planning fallacy and safety margins and time to iterate a little and all that.)

But iterative engineering is important!

In order for iterative engineering to be useful, we first need to have a strong enough understanding of what we even want to achieve in order to recognize when an iteration has brought us a step closer to the goal. No amount of A/B testing changes to our website will make our company profitable if we’re measuring the wrong metrics. I claim that, for alignment, we do not yet have a strong enough understanding for iteration to produce meaningful progress.

When I say “we’re just fundamentally confused about agency and alignment”, that’s the sort of thing I’m talking about.

To be clear: we can absolutely come up with proxy measures of alignment. The problem is that I don’t expect iteration under those proxy measures to get us meaningfully closer to aligned AGI. No reasonable amount of iterating on gliders’ flight-range will get one to the moon.

But engineering is important for advancing understanding too!

I do still expect some amount of engineering to be central for making progress on fundamental confusion. Engineering is one of the major drivers of science; failed attempts to build amplifiers drove our first decent understanding of semiconductors, for instance. But this is a very different path-to-impact than directly iterating on “alignment”, and it makes sense to optimize our efforts differently if the path-to-impact is through fundamental understanding. Just take some confusing concept which is fundamental to agency and alignment (like abstraction, or optimization, or knowledge, or .) and try to engineer anything which can robustly do something with that concept. For instance, a lot of my own work is driven by the vision of a “thermometer of abstraction”, a device capab...

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: There is essentially one best-validated theory of cognition., published by abramdemski on December 10, 2021 on LessWrong.

There are many theories of cognition. But if you want to work within a framework with the following properties:

Explains the major cognitive phenomena we know about.

Fits experimental data well, down to human reaction times, in a wide variety of psychological experiments.

Has a relatively complete story about functional neuroanatomy.

Well, then, I'm not aware of any theories which fit the bill as well as ACT-R theory.

You might also be interested in the common model of cognition (initially named standard model of the mind), which is consistent with the ACT-R picture, but also consistent with several competing theories. Think of it as the high-certainty subset of ACT-R.

References for ACT-R

I am no expert on ACT-R, so unfortunatery I can't tell you the best place to get started! However, here are some references.

Books by John R Anderson

John R Anderson is the primary researcher behind ACT-R theory. I have not read all of the following.

Learning and memory: an integrated approach. This book is not about ACT-R. It's basically a well-written textbook on everything we know about learning and memory. I think of it as John Anderson's account of all the empirical psychological phenomena which he would like to explain with ACT-R.

How can the human mind occur in the physical universe? This is a book about ACT-R. In this book, John R Anderson seeks to spell out "how the gears clank and how the pistons go and all the rest of that detail" -- which is to say, the inner workings of the brain.

John R Anderson also has several other books, which I haven't looked at very much; so maybe a different one is actually a better starting place.

Other References

Here are some other potentially useful starting places.

References for the Common Model of Cognition

A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics. This is the original paper for the common model of cognition. It rests on the observation that many cognitive architectures seem to be converging to common principles, much more than they were even 10 years earlier. To codify this consensus, leading researchers behind three of these theories got together and made a close comparison. (One of the three is my phd advisor, which is how I became aware of this.)

Empirical Evidence from Neuroimaging Data for a Standard Model of the Mind. What it says on the tin.

If you search "common model of cognition" in Google Scholar, you will find a number of other papers discussing it.

How should we evaluate all this?

It doesn't scratch the same itch that Solomonoff Induction and other ideas more commonly discussed in this community do. Like, at all. The ACT-R interpretation of the question "how can the human mind occur in the physical universe?" has little to do with embedded agency questions, and much more to do with computational psychology.

Although it is supposed to be a computational model of cognition, ACT-R is not an AGI. I would say it's not even a proto-AGI. The ACT-R community cares about closely matching human performance in psychological tests. It turns out this task is a lot different from, say, building a high-performance machine learning system. ACT-R is definitely not the latter.

As such, despite ACT-R's success at replicating human learning behavior in the context of psychology tests, we might broadly conclude that it's missing something which modern machine learning is not missing.

I'm also not claiming that you should believe ACT-R. If you take existing theories of cognition and compare them in Bayesian terms, I think ACT-R comes out on top. But ACT-R has a lot of pieces, and I don't think all of them are necessarily...

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: There is essentially one best-validated theory of cognition., published by abramdemski on December 10, 2021 on LessWrong.

There are many theories of cognition. But if you want to work within a framework with the following properties:

Explains the major cognitive phenomena we know about.

Fits experimental data well, down to human reaction times, in a wide variety of psychological experiments.

Has a relatively complete story about functional neuroanatomy.

Well, then, I'm not aware of any theories which fit the bill as well as ACT-R theory.

You might also be interested in the common model of cognition (initially named standard model of the mind), which is consistent with the ACT-R picture, but also consistent with several competing theories. Think of it as the high-certainty subset of ACT-R.

References for ACT-R

I am no expert on ACT-R, so unfortunatery I can't tell you the best place to get started! However, here are some references.

Books by John R Anderson

John R Anderson is the primary researcher behind ACT-R theory. I have not read all of the following.

Learning and memory: an integrated approach. This book is not about ACT-R. It's basically a well-written textbook on everything we know about learning and memory. I think of it as John Anderson's account of all the empirical psychological phenomena which he would like to explain with ACT-R.

How can the human mind occur in the physical universe? This is a book about ACT-R. In this book, John R Anderson seeks to spell out "how the gears clank and how the pistons go and all the rest of that detail" -- which is to say, the inner workings of the brain.

John R Anderson also has several other books, which I haven't looked at very much; so maybe a different one is actually a better starting place.

Other References

Here are some other potentially useful starting places.

References for the Common Model of Cognition

A Standard Model of the Mind: Toward a Common Computational Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics. This is the original paper for the common model of cognition. It rests on the observation that many cognitive architectures seem to be converging to common principles, much more than they were even 10 years earlier. To codify this consensus, leading researchers behind three of these theories got together and made a close comparison. (One of the three is my phd advisor, which is how I became aware of this.)

Empirical Evidence from Neuroimaging Data for a Standard Model of the Mind. What it says on the tin.

If you search "common model of cognition" in Google Scholar, you will find a number of other papers discussing it.

How should we evaluate all this?

It doesn't scratch the same itch that Solomonoff Induction and other ideas more commonly discussed in this community do. Like, at all. The ACT-R interpretation of the question "how can the human mind occur in the physical universe?" has little to do with embedded agency questions, and much more to do with computational psychology.

Although it is supposed to be a computational model of cognition, ACT-R is not an AGI. I would say it's not even a proto-AGI. The ACT-R community cares about closely matching human performance in psychological tests. It turns out this task is a lot different from, say, building a high-performance machine learning system. ACT-R is definitely not the latter.

As such, despite ACT-R's success at replicating human learning behavior in the context of psychology tests, we might broadly conclude that it's missing something which modern machine learning is not missing.

I'm also not claiming that you should believe ACT-R. If you take existing theories of cognition and compare them in Bayesian terms, I think ACT-R comes out on top. But ACT-R has a lot of pieces, and I don't think all of them are necessarily...