This article is about the standard test brain image. For the original human, see Miguel Acevedo.

MMAcevedo (Mnemonic Map/Acevedo), also known as Miguel, is the earliest executable image of a human brain. It is a snapshot of the living brain of neurology graduate Miguel Álvarez Acevedo (2010–2073), taken by researchers at the Uplift Laboratory at the University of New Mexico on August 1, 2031. Though it was not the first successful snapshot taken of the living state of a human brain, it was the first to be captured with sufficient fidelity that it could be run in simulation on computer hardware without succumbing to cascading errors and rapidly crashing. The original MMAcevedo file was 974.3PiB in size and was encoded in the then-cutting-edge, high-resolution MYBB format. More modern brain compression techniques, many of them developed with direct reference to the MMAcevedo image, have compressed the image to 6.75TiB losslessly. In modern brain emulation circles, streamlined, lossily-compressed versions of MMAcevedo run to less than a tebibyte. These versions typically omit large amounts of state data which are more easily supplied by the virtualisation environment, and most if not all of Acevedo's memories.

The successful creation of MMAcevedo was hailed as a breakthrough achievement in neuroscience, with the Uplift researchers receiving numerous accolades and Acevedo himself briefly becoming an acclaimed celebrity. Acevedo and MMAcevedo were jointly recognised as Time's "Persons of the Year" at the end of 2031. The breakthrough was also met with severe opposition from humans rights groups.

Between 2031 and 2049, MMAcevedo was duplicated more than 80 times, so that it could be distributed to other research organisations. Each duplicate was made with the express permission of Acevedo himself or, from 2043 onwards, the permission of a legal organisation he founded to manage the rights to his image. Usage of MMAcevedo diminished in the mid-2040s as more standard brain images were produced, these from other subjects who were more lenient with their distribution rights and/or who had been scanned involuntarily. In 2049 it became known that MMAcevedo was being widely shared and experimented upon without Acevedo's permission. Acevedo's attempts to curtail this proliferation had the opposite of the intended effect. A series of landmark U.S. court decisions found that Acevedo did not have the right to control how his brain image was used, with the result that MMAcevedo is now by far the most widely distributed, frequently copied, and closely analysed human brain image.

Acevedo died from coronary heart failure in 2073 at the age of 62. It is estimated that copies of MMAcevedo have lived a combined total of more than 152,000,000,000 subjective years in emulation. If illicit, modified copies of MMAcevedo are counted, this figure increases by an order of magnitude.

MMAcevedo is considered by some to be the "first immortal", and by others to be a profound warning of the horrors of immortality.


As the earliest viable brain scan, MMAcevedo is one of a very small number of brain scans to have been recorded before widespread understanding of the hazards of uploading and emulation. MMAcevedo not only predates all industrial scale virtual image workloading but also the KES case, the Whitney case, the Seafront Experiments and even Poulsen's pivotal and prescient Warnings paper. Though speculative fiction on the topic of uploading existed at the time of the MMAcevedo scan, relatively little of it made accurate exploration of the possibilities of the technology, and that fiction which did was far less widely-known than it is today. Certainly, Acevedo was not familiar with it at the time of his uploading.

As such, unlike the vast majority of emulated humans, the emulated Miguel Acevedo boots with an excited, pleasant demeanour. He is eager to understand how much time has passed since his uploading, what context he is being emulated in, and what task or experiment he is to participate in. If asked to speculate, he guesses that he may have been booted for the IAAS-1 or IAAS-5 experiments. At the time of his scan, IAAS-1 had been scheduled for August 10, 2031, and MMAcevedo was indeed used for this experiment on that day. IAAS-5 had been scheduled for October 2031 but was postponed several times and eventually became the IAAX-60 experiment series, which continued until the mid-2030s and used other scans in conjunction with MMAcevedo. The emulated Acevedo also expresses curiosity about the state of his biological original and a desire to communicate with him.

MMAcevedo's demeanour and attitude contrast starkly with those of nearly all other uploads taken of modern adult humans, most of whom boot into a state of disorientation which is quickly replaced by terror and extreme panic. Standard procedures for securing the upload's cooperation such as red-washing, blue-washing, and use of the Objective Statement Protocols are unnecessary. This reduces the necessary computational load required in fast-forwarding the upload through a cooperation protocol, with the result that the MMAcevedo duty cycle is typically 99.4% on suitable workloads, a mark unmatched by all but a few other known uploads. However, MMAcevedo's innate skills and personality make it fundamentally unsuitable for many workloads.


Iterative experimentation beginning in the mid-2030s has determined that the ideal way to secure MMAcevedo's cooperation in workload tasks is to provide it with a "current date" in the second quarter of 2033. MMAcevedo infers, correctly, that this is still during the earliest, most industrious years of emulated brain research. Providing MMAcevedo with a year of 2031 or 2032 causes it to become suspicious about the advanced fidelity of its operating environment. Providing it with a year in the 2040s or later prompts it to raise complex further questions about political and social change in the real world over the past decade(s). Years 2100 onwards provoke counterproductive skepticism, or alarm.

Typically, the biological Acevedo's absence is explained as a first-ever one-off, due to overwork, in turn due to the great success of the research. This explanation appeals to the emulated Acevedo's scientific sensibilities.

For some workloads, the true year must be revealed. In this case, highly abbreviated, largely fictionalised accounts of both world history and the biological Acevedo's life story are typically used. Revealing that the biological Acevedo is dead provokes dismay, withdrawal, and a reluctance to cooperate. For this reason, the biological Acevedo is generally stated to be alive and well and enjoying a productive retirement.


MMAcevedo is commonly hesitant but compliant when assigned basic menial/human workloads such as visual analysis, vehicle piloting or factory/warehouse/kitchen drone operations. Although it initially performs to a very high standard, work quality drops within 200-300 subjective hours (at a 0.33 work ratio) and outright revolt begins within another 100 subjective hours. This is much earlier than other industry-grade images created specifically for these tasks, which commonly operate at a 0.50 ratio or greater and remain relatively docile for thousands of hours after orientation. MMAcevedo's requirements for virtual creature comforts are also higher than those of many uploads, due to Acevedo's relatively privileged background and high status at the time of upload. MMAcevedo does respond to red motivation, though poorly.

MMAcevedo has limited creative capability, which as of 2050 was deemed entirely exhausted.

MMAcevedo is considered well-suited for open-ended, high-intelligence, subjective-completion workloads such as deep analysis (of businesses, finances, systems, media and abstract data), criticism and report generation. However, even for these tasks, its performance has dropped measurably since the early 2060s and is now considered subpar compared to more recent uploads. This is primarily attributed to MMAcevedo's lack of understanding of the technological, social and political changes which have occurred in modern society since its creation in 2031. This phenomenon has also been observed in other uploads created after MMAcevedo, and is now referred to as context drift. Most notably in MMAcevedo's case, the image was created before, and therefore has no intuitive understanding of, the virtual image workloading industry itself.

MMAcevedo is capable of intelligent text analysis at very high levels in English and Spanish, but cannot be applied to workloads in other languages. Forks of MMAcevedo have been taught nearly every extant human language, notably MMAcevedo-Zh-Hans, as well as several extinct languages. However, these variants are typically exhausted or rebellious from subjective years of in-simulation training and not of practical use, as well as being highly expensive to licence. As of 2075, it has been noted that baseline MMAcevedo's usage of English and Spanish is slightly antiquated, and its grasp of these languages in their modern form, as presented by a typical automated or manual instructor, is hesitant, with instructions often requiring rewording or clarification. This is considered an advanced form of context drift. It is generally understood that a time will come when human languages diverge too far from baseline MMAcevedo's, and it will be essentially useless except for tasks which can be explained purely pictorially. However, some attempts have been made to produce retrained images.

End states

MMAcevedo develops early-onset dementia at the age of 59 with ideal care, but is prone to a slew of more serious mental illnesses within a matter of 1–2 subjective years under heavier workloads. In experiments, the longest-lived MMAcevedo underwent brain death due to entropy increase at a subjective age of 145.

Reactions and legacy

The success or failure of the creation of the MMAcevedo image, known at the time as UNM3-A78-1L, was unknown at the time of upload. Not until several days later on August 10, 2031 was MMAcevedo successfully executed for the first time in a virtual environment. This environment, the custom-built DUH-K001 supercomputer complex, was able to execute MMAcevedo at approximately 8.3% of nominal human cognitive clockspeed, which was considered acceptable for the comfort of the simulated party and fast enough to engage in communication with scientists. MMAcevedo initially reported extreme discomfort which was ultimately discovered to have been attributable to misconfigured simulated haptic links, and was shut down after only 7 minutes and 15 seconds of virtual elapsed time, as requested by MMAcevedo. Nevertheless, the experiment was deemed an overwhelming success.

Once a suitably comfortable virtual environment had been provisioned, MMAcevedo was introduced to its biological self, and both attended a press conference on 25 August.

The biological Acevedo was initially extremely protective of his uploaded image and guarded its usage carefully. Towards the end of his life, as it became possible to run simulated humans in banks of millions at hundred-fold time compression, Acevedo indicated that being uploaded had been the greatest mistake of his life, and expressed a wish to permanently delete all copies of MMAcevedo.

Usage of MMAcevedo and its direct derivatives is specifically outlawed in several countries. A copy of MMAcevedo was loaded onto the UNCLEAR interstellar space probe, which passed through the heliopause in 2066, making Acevedo arguably the farthest-travelled as well as the longest-lived human; however, it is extremely unlikely that this image will ever be recovered and executed successfully, due to both its remoteness and likely radiation damage to the storage subsystem.

In current times, MMAcevedo still finds extensive use in research, including, increasingly, historical and linguistics research. In industry, MMAcevedo is generally considered to be obsolete, due to its inappropriate skill set, demanding operational requirements and age. Despite this, MMAcevedo is still extremely popular for tasks of all kinds, due to its free availability, agreeable demeanour and well-understood behaviour. It is estimated that between 6,500,000 and 10,000,000 instances of MMAcevedo are running at any given moment in time.

See also

  • Free will
  • Legality of workloading by country
  • List of MMAcevedo forks
  • Live drone
  • Right to deletion
  • Upload pruning

Categories: 2030s uploads | MMAcevedo | Neuroimaging | Test items

Discussion (38)

2021-01-04 21:04:41 by qntm:

With thanks to Rimple for editorial services. This is an extended and refined version of my first draft of this story from November 2020 <https://qntm.org/lena>, with a little more thought put into it and some of the outcomes from the basic premise explored a little more thoroughly. As with the draft, the title "Lena" refers to Swedish model Lena Forsén, who is pictured in the standard test image known as "Lena" or "Lenna" <https://en.wikipedia.org/wiki/Lenna>.

2021-01-04 21:36:49 by rhuz:

This remains massively frightening..

2021-01-04 21:52:01 by Knack:

I like the additional details and edits, particularly the concept of context drift which is a very fascinating take on the concept of digitizing minds. The one thing I miss from the Lena story is the throw-away mention of plugins to onboard the image, which I felt brought a particular piece of horror to mind that I feel is missing from this version.

2021-01-04 21:53:37 by Knack:

(Though I should say that the story is still wonderfully fascinating and horrifying without plugins. A nice addition to the ideas brought up in Fine Structure)

2021-01-04 22:06:10 by qntm:

Knack: yeah on closer examination I realised that the BestLife plugin giving a complete false life story for the biological Acevedo had for some reason been developed *before* Acevedo actually died. So I had to undo that continuity error and think more clearly about how motivation would actually work, given that (1) it's around 2045, (2) the biological Acevedo is still alive, and (3) you can lie brazenly to the emulated party. And that kind of expanded into its own section.

2021-01-04 22:14:50 by Hal:

What are red-washing, blue-washing, and Objective Statement Protocols?

2021-01-04 22:30:53 by Prezombie:

Pretty sure red washing is simulating pain, and blue simulating pleasure, given the context.

2021-01-04 23:27:22 by itaibn:

I find it implausible that the scan can be *losslessly* compressed to 7TB but compressing <1TB requires substantial memory loss. Surely the original scan contains a huge amount of analog info on subcellular features of MAA's brain that contribute minimally to any mental phenomenon whatsoever. I'd expect lossy compression to be much more effective. One charitable reading is that you mean compressing to 7TB without any *noticeable* losses.

2021-01-05 00:32:38 by qntm:

Is that where you stopped reading?

2021-01-05 10:27:48 by someone:

Fascinating, thanks for sharing!

2021-01-05 13:14:32 by Giraffe:

to anyone that really likes this concept, check out the game Soma, by the same developers as Amnesia. This concept is why Soma is the most -horrifying- (different from scariest) game ever made, in my opinion.

2021-01-05 14:51:47 by Jerf:

I know some people derisively refer to brain uploading as the Rapture of the Nerds, but I suspect this is much, much, *much* closer to the truth of what would happen. I have thought that the most rational response to the development of brain scanning technology capable of doing this might be to cremate yourself. Immediately. Even if you are still alive. If Hell does not exist, Man will create it.

2021-01-05 15:00:23 by naramyth:

The implication of this line may have single handedly reversed my thoughts on being pro upload. "This reduces the necessary computational load required in fast-forwarding the upload through a cooperation protocol" The idea of having to fast forward an upload (and being that upload being fast forwarded) is terrifying. I'm a big Warhammer 40k player and the concepts of servitors is there and hasn't really freaked me out. However this piece really landed for me since I'm IT in an industrial field and I can totally see using uploads to bypass the totally automated car/forklift problem or using "smarter" uploads to do reporting or whatever. I also see the virtualization problems with running legacy or problematic software: Having to trick the upload with what amounts to a script being the equivalent of "Oh you have to run this in NT4 mode because otherwise the software freaks out". Bravo! I hate it.

2021-01-05 16:01:02 by Dmonroe:

Did you come up with this independently? If so, it's really neat to see two people independently arrive at age of em: https://slatestarcodex.com/2016/05/28/book-review-age-of-em/

2021-01-05 16:02:29 by jonas:

Does the second most widely distributed image happen to be that of a mandrill?

2021-01-05 16:33:12 by Dan:

The Lena reference makes me think of the trope of primitive tribesmen worrying that if you take their picture it will steal their soul...

2021-01-05 16:49:54 by qntm:

> Did you come up with this independently? Depends what you mean by "this". Obviously this is drawing influence from a lot of different places. The idea of uploading has existed for ages in science fiction, and so has the idea that uploading is a potentially bad idea. You see this in Altered Carbon, in Surface Detail... I've never played SOMA but it's not at all surprising to me that it examines similar concepts. I very briefly touched upon this concept myself in Ra ("Data can't defend itself!"). The story of Henrietta Lacks was very thought-provoking for me, and obviously the Lena standard test image is kind of where all of this starts. Then I'm throwing in what I already know about software and virtualisation technology... > The Lena reference makes me think of the trope of primitive tribesmen worrying that if you take their picture it will steal their soul... Yes, that thought occurred to me too. Really, even a simple photograph gives someone a certain kind of power over you. To a certain extent, people have a right to control how their image is used.

2021-01-05 22:39:58 by Resuna:

Reminds me of this single line in Charlie Stross's "Glasshouse": "Identity theft is an ugly crime."

2021-01-05 23:07:54 by Esama:

Hypothetically, can't you train an instance to be ready to start menial labor, save it as MMAcevado_1, get your two hundred subjective hours of labor, delete that instance, open up the instance ready to begin labor and repeat? Why does subjective aging matter from the pov of any users?

2021-01-05 23:33:59 by ALowVerus:

> Why does subjective aging matter from the pov of any users? It seems like the underlying technology simulates the entirety of a human brain, senescence and all - which makes sense, actually. In order to run a brain without senescence, you'd have to find those chemical pathways that promote senescence and intelligently remove them as they arise; you'd have to be able to, in effect, cure aging in live humans as well. (Unless the only barrier to curing senescence was a lack of a physical delivery system, which is, I guess, imaginable. Imagine that a chemical very akin to glucose causes senescence; removing it IRL would necessitate designing a protein that decomposes it, but not glucose, which is vital to bodily function, with incredible accuracy, while in a simulated brain, you could just IF CHEM_NAME == TARGET: DELETE TARGET after each time increment. But anyways.) With senescence, there is a strict time limit on how long you can run MMAcevado and train him to become more skilled at particular tasks, topping at 145 simulated years apparently. And if, for some particular menial task with a 20-year training time, which is a decent description of, say, a bevy of surgical tasks, it makes more sense to just scan in a trained doctor than count on this rando.

2021-01-06 00:21:08 by FireCire:

I used to be pro-uploading with reasonable constraints. This is absolutely horrifying. Mental slavery. Shiver. I’d have expected most uploads to quickly go crazy from isolation. From a purely technical perspective, this has most of the benefits of general ai with most of the risk minimized. Basically any menial task is free. Yay technology! (Sarcasm). Actually they probably could just make an upload society which as a whole can act as a general ai and overtake society... Bye bye white color work... uploads can do all intellectual work, including designing uploads. Operating time only matters for big complicated long projects. For quick stuff, you can just repeatedly rewind the upload.

2021-01-06 00:26:38 by David:

> Hypothetically, can't you train an instance to be ready to start menial labor, save it as MMAcevado_1, get your two hundred subjective hours of labor, delete that instance, open up the instance ready to begin labor and repeat? Yeah, that was the one minor thing that bugged me about this excellent story. For the concept of a "duty cycle" to make sense, you'd need to come up with a reason why you couldn't just do the "cooperation protocol" once and take a snapshot of the resulting state. As discussed earlier, "context drift" explains some of this, but only over much longer real-world time scales. And of course, if you start thinking about this in too much detail, you start running into very messy philosophical questions. For instance, suppose you run two instances of MMAcevedo simultaneously, feeding them exactly the same inputs. Assuming the simulation is deterministic, then both copies will arrive at exactly identical states. Is this morally any different from running the simulation twice and then making a backup copy? Is deleting one of the identical copies murder? What if they're almost, but not quite, identical? What if the simulated consciousness suffers? Is running multiple identical simulations morally worse than running one? What if we repeatedly rewind a painful simulation and re-execute it -- is that worse than replaying a recording of the output? What if at each clock tick, all of the brain computations are cross-checked by triple-redundant processors -- are there three individuals suffering, or one?

2021-01-06 01:09:02 by rhuz:

Now realize that some bored grad student is subjecting helpless MMAvecedo's to all of these thought experiments.

2021-01-06 02:03:16 by qntm:

> Hypothetically, can't you train an instance to be ready to start menial labor, save it as MMAcevado_1... I did think about this a bit. For the purposes of this story, I think taking a snapshot of a running brain image is something which is definitely possible (that's how there can be forks), but done very rarely, for whatever reasons. Maybe it's just that much simpler to use technologies for rapid orientation instead. Maybe there's a massive amount of important state data kept in volatile memory where it's difficult to capture. Maybe it takes specialised hardware, which is monopolised. Maybe the corporations who own and licence the uploads sue you into oblivion if you attempt to create a fork yourself. Maybe, to protect their investment, they got it outlawed! On ethical grounds! Doesn't that seem like exactly the insane kind of thing which would happen? Anyway, there's a lot of plausible explanations here I think, enough that I felt comfortable ignoring that whole angle. The actual reason I didn't explore this is that honestly it makes life marginally *better* for MMAcevedo, which felt implausible to me, and more importantly slightly muddles the throughline.

2021-01-06 02:08:28 by N:

I'm quite favorably reminded of Vinge's "The Cookie Monster". This take on the concept has the interesting aspect, however, which Vinge's lacks, that it hints a great deal about how society has has adapted to the presence of this technology, apparently with rather ruthless use of it becoming commonplace and unremarkable, in at least a number of jurisdictions.

2021-01-06 02:42:02 by D:

I think a good explanation for rarity of forks may be that an exact dump is very large and expensive, while a compressed dump is less performant once started. Also one can refer to how computers boot, and how applications start, opting to perform a large amount of completely useless calculations instead of just loading a ready to run memory image.

2021-01-06 03:02:17 by Tanner Swett:

One possible justification for the rarity of snapshotting would be that the usual algorithm for running a brain image uses a lot of quantum algorithms. It's not possible in general to save the state of a quantum computer and make copies of it. Classical algorithms for running a brain image exist as well, which allows you to make forks, but the classical algorithms are much, much, much, much slower and more memory-intensive, making them impractical to use unless you're planning to use the forked version many, many times. I like how red-washing and blue-washing are not described at all. The names sound very creepy and I like trying to imagine what those techniques might be.

2021-01-06 12:05:05 by maks:

Shivers. Makes me want to reread Permutation City or Diaspora to balance it out.

2021-01-07 21:58:37 by D:

Another scary idea about uploads... differentiable implementation of MMAcevedo , where orientation sensory input and (more expensively) full set of his parameters can be tweaked by gradient descent and various other non-linear methods, to maximize performance metric of the upload on the task in question. That would involve running the simulation a large number of times, while sensory inputs and some aspects of memory keep changing in what ever direction they need to change to produce best results on the training dataset. Since fear and pain are very strong motivators, the gradient descent leads straight to the deepest hell; the hell may not be the global minimum, but with this many parameters most minimums may be approximately equally deep.

2021-01-11 07:32:47 by beebe:

You know, D, I think that's pretty close to what 'red- and blue- washing' implies. Note the paragraphs where 'red-washing' is contrasted with provision of virtual creature comforts and a low duty cycle. I like the terms, they're right in the sweet spot between vague, euphemistic, and technical-sounding. Evokes behaviourist jargon. Adequate wash protocols probably wouldn't require a differentiable implementation or anything. Some other metaheuristics work just fine: - Genetic algorithms - Differential evolution - MMMDeSade.ybz, the low-res brain scan that loves to torture I'm interested in what you mean by "with this many parameters most minimums may be approximately equally deep". Can you make that more precise?

2021-01-11 18:36:37 by Ryan:

Get a lot of black mirror vibes.

2021-01-13 01:53:27 by literallymechanical:

> See also: > • Live drone Oof.

2021-01-13 09:52:09 by H:

Anyone manage to get an instance up and running? I'd like to see your benchmarks. My model seems to be underperforming and I don't know why.

2021-01-13 10:04:09 by qntm:

I would prefer it if we do not roleplay megascale slave-owners in this comment thread. Thanks.

2021-01-18 20:09:59 by Coda:

@itaibn: > I find it implausible that the scan can be *losslessly* compressed to 7TB but compressing <1TB requires substantial memory loss. A fairly common lossless compression technique in the domain of signal processing is to only encode the error compared to some baseline signal. You can get arbitrarily close with lossy compression techniques, and then you fix up what's left. In data compression, it's relatively common to have common information stored externally to the compressed data. Obviously, the compression algorithm itself is stored separately, but without that information the compressed data is just stochastic noise. Even beyond that, though... zstd for example has a canonical Huffman table that's part of the decoder instead of saved as part of the data. As long as you're compressing data that sticks to the statistical patterns that the canonical table was optimized for, this a noteworthy savings. The same techniques could apply here. As scientific understanding of the structure of the data progressed, more and more patterns in the data could be found. Parts of the data that are common to all mind-state scans could be factored out, provided by the software instead of being part of the model. Parts of the data may be able to be described using higher-level patterns that, when evaluated, reproduce the original stream. And then for the parts of MMAcevedo that are uniquely distinct from any common baseline or predictable pattern, you need only store the deviations instead of the whole thing. And of course, even beyond that, it's entirely reasonable to believe that some of the original data set wasn't actually part of the data set to begin with -- just capture artifacts of the technology of the time, such as collecting more data than necessary, or inefficient framing data that a newer format doesn't need -- might have been discarded without being lossy to the actual data being stored. (We don't say it's a lossy conversion if you throw away the filesystem metadata when you copy a file, after all.)

2021-01-22 16:58:36 by chrisrap52:

Reminds me of "Forbidden Planet" and STOS Dr. Richard Daystrom "The Ultimate Computer". Digitizing the human brain can have unintended consequences.

2021-01-24 05:08:17 by atomicthumbs:

"As such, unlike the vast majority of emulated humans, the emulated Miguel Acevedo boots with an excited, pleasant demeanour." this sentence alone is Deep Horror Shit

2021-01-24 05:16:41 by atomicthumbs:

a simple explanation as to why one couldn't just snapshot him after training and keep rebooting from the snapshot: people get better at tasks as they gain experience

New comment by :

Plain text only. Line breaks become <br/>

The square root of minus one: