When will uploaded minds be a reality?

Question

The Premise

Setting a near-ish future world, where people live a full life and then upload their minds into different hardware and continue on.

Setting a timeline for this, when (what year) are we looking at? Note that the development is an ongoing process so initially it becomes technically possible^※, then the rich can do it, and eventually white-collar workers or retiree’s can do it.

Hardware

There are several issues. First is the hardware. The needed hardware is discussed in this (sadly under-rated) answer. Taking as a baseline the spiking network implementation of a neocortex, add some overhead to handle porting an existing mind. That is, it will need flexibility to adjust the cortical columns to match what had grown in the brain, and there may be additional ad-hoc mechanisms needed to make the feedback and control work exactly like the brain rather than a clean designed implementation. But it would still be more efficient than simulating individual brain cells, or at least require much less memory.

So figure 10¹⁹ flops, but only 4 to 40 Tbytes of RAM to hold the state of all the pattern matcher complexes (the figure is sensitive to the degree of fan-out among the pattern matcher units) and double that for the loose individual nurons and ad-hoc stuff that doesn’t fit the model.

The Costs

The new post human will need to have this measure of compute resources dedicated to him, whether owned, leased, or rented; so the cost of that comes down to him. Certainly it’s worth taking out a mortgage for, or saving up for. But consider the equipment lifetime and the earning potential of the post human using it. Although the price of an upgrade will drop years later, he needs to sustainably afford to keep hardware.

Next, something that I don’t see considered in this kind of fiction! The cost of energy to run the brain! The first usable exobrains will use much more power than the old brain, so grocery money isn’t enough to keep it going.

The growing population of post humans may have a larger energy demand than they did in their biological life, and this is an interesting societal aspect to explore. But in this question in particular, the post human needs to afford to “eat”. This is an ongoing expense that goes with his uptime.

People in this state might be earning based on saved capital, but as it works down to more ordinary people, we need his earnings as a post-human to sustain his expenses.

misc

Meanwhile… fiction often portrays AI and uploaded minds as being enormously faster than life. But I think they will initially be slower! If the upload runs at a fraction of the speed of his biological counterparts, it will affect his ability to work job and amplify the electricity cost per consciousness-hour.

I suppose that virtualization of environments and connectivity will be cheap enough locally (within the post-human living center) that it’s included in the power cost. Connectivity beyond his home center would be the same as any network access costs.

Finally, there is the one-time expense of getting scanned and uploaded. That is something that can be dragged out as long as needed since it is a one-time and not reocurring expense. As long as the post-human can earn money (whether by working or ownership) this won’t be a limiting factor.

---

The Question

So, what future timeline are we looking at for the phasing in of post-humanism? When will it be possible for the rich-enough, and when for middle class first-world citizens?

Consider cost of the exobrain and ongoing running costs, compared with people’s ability to afford that before and after porting themselves.

If the idea of mind uploading is so distracting that you can’t help but to react to that instead, try to forget that and consider the question to be “When will people (of various economic) classes be able to afford to finance and run a computer of the stated power (when they won’t have normal expenses of housing and food, either)?”

---

Footnote

※ The existence of destructive scanning technology can be assumed as part of the story. It will simply be there once there are computers capable of handling the data.

Likewise, getting potential customers to accept uploading is a different topic from this question.

Clarification

Please note what the question asks. (why do I need to say this?!)

It is not: propose a different story, repudiate the premise, riff on what posthumanism is like, repeat the linked hardware discussion, …,

Read what’s already been discussed if you have a comment.

Answer 1

Interesting set of questions there, let's see how far I can get…

Assuming "10¹⁹ flops, but only 4 to 40 Tbytes of RAM":

The current fastest supercomputer on Earth is ~ 10¹⁷ flops (Sunway TaihuLight, 93.015-125 Petaflops depending who you ask), at a cost of US$ 273 million, and drawing a power of 15 MW.

To reach 10¹⁹ today would cost ~US\$ 27 billion and consume 1.5 GW. At US\$ 0.05 / kWh, that electricity bill will cost US\$ 660 million per year. Only the richest twenty people alive today could afford that, and only the top seven would have enough left over for their investments to cover the energy bill afterwards.

However, assuming Moore's Law continues indefinitely, doubling performance every 18 months is roughly a factor of ten every 5 years:

• In ten years, the world's fastest supercomputer will be able to simulate one mind in real-time, assuming that computer still costs US\$ 273 million to buy, and still uses 15 MW (at a cost of US\$ 6.6 million/year) to run. The exact performance of the world's fastest supercomputer depends on how much governments want to spend on supercomputers as much as it does on the technology.

• In twenty years, an equivalent computer will cost US\$ 2.73 million to buy, and will use 150 kW at a cost of US\$ 66,000 per year to run.

• In 25 years, it will cost US\$ 273,000 to buy, so upper-middle-class families might start selling their homes to do this. They will use 15 kW at a cost of US\$ 6,600 per year to run, a combination which may make business invest in using uploaded minds as slave labour they don't have to pay real money to (once you have one uploaded mind, ctrl-C, ctrl-V).

• In 30 years, that computing power will cost US\$ 27,300 to buy, and will use 1.5 kW at a cost of US\$ 660 per year (US\$ 1.80 per day) to run, making the digital-slave issue practically universal — zero Americans live on this income or lower, and in 2013, less than 10.7% of the world population lived on US\$ 1.90 per day or less.

Alternatively, if the "upload" part is perfected right now and you fix the construction and energy costs at those of the Sunway TaihuLight supercomputer, you get speed-factors compared to reality of:

• Today: 0.01 (1% real-time)
• 2027: 1 (real-time)
• 2037: 100 (every year of real time is a century of subjective time)
• 2042 (when upper-middle-class families can afford real-time simulation): 1,000
• 2047 (when the electricity cost of the real-time version is close to the cost of subsistence food for human slaves): 10,000
• etc.

Answer 2

The brain's software is unknown.

You talk about hardware, but you don't talk about software. That is a problem.

How does the human brain actually work? How does the human brain add 2+2? I know how a computer does it. I doubt anyone knows how the brain does it. Its not like some bits are flipped here and there and the answer just pops out.

Computers and computer software work by building functionality on top of a simple binary mathematics. The brain does not function this way. It has many specially evolved functions which can be combined in various ways to produce output. How will we generate the software functions to perform everything the brain does, when we don't really understand how the brain works in the first place?

I do not see any proposed way of modeling a complex, non-binary system completely with binary electronic computing. Throwing teraflops at this problem fundamentally does not solve it.

The brain's source of creativity is unknown.

Cutting-edge AIs are pretty smart and can do a lot of things. But they are still far from being a general intelligence. The fundamental thing that is lacking is creativity. While computers are now powerful enough to out-think humans in limited problem sets (chess 20 years ago, go just recently), the ability to create something new ex nihilo does not exist for a computer.

A computer can look at its entire information space, which could potentially be huge, like all medical literature ever, and try things until it gets the best answer. But the computer will never be able to look beyond its information space. Humans can do that. How can you replicate this ability with a machine?

Until these things are known, we cannot upload a person into a machine

How long until these things are known? I don't think that we can extrapolate from the information we have on current trends. There is very little we really know about the human brain right now. We have generally identified some areas that are responsible for various things, but cutting edge neuro-science is simply not reliable. Too many bold new findings have not been properly replicated. I'm not saying neuroscience is wrong, it is just searching for answers in the dark, and finding few.

What we do know about the brain is the ability to manipulate it, to a certain extent, with drugs. Until we can and do perform drug-approval style tests of other brain functions (ability to navigate, detect shapes and movement, use language, do math, etc) such that we know how to turn on and turn off these features in a brain, we are nowhere close to understanding how they work.

Conclusion

I wanted to work out a timeline estimate to answer the question, but that would just be irresponsible. We don't know what we don't know about the brain, and until we know more, we don't even know what we need to simulate a human brain.

AI will continue to develop, and computers will continue to move into fields formerly reserved for humans. AI car drivers, medical diagnosers, research assistants, and bank auditors may not be far away. Yet while computers may exceed human capabilities in all these fields soon, there is a still a massive divide between the abilities of a special AI, and the general intelligence of the human brain.

Final note, in the spirit of full disclosure, I desperately do not want it to be possible for a human to be uploaded into a computer. That will be a socially disruptive event orders of magnitudes beyond any previous change, and I cannot imagine humanity surviving it. Maybe some other form of (higher?) life will come out the other end, but whatever they are, they won't be human any more. I like humans, being one, and I want to see us all survive.

Answer 3

The trouble with this question is that we don't know at what rate we are progressing as a species towards these goals. It's not like building a house where we know that we have the foundations right and the walls up and we just need to add a roof. We think we might be going down the right path but we could be on a wild goose chase, in the 1930's they thought that Space Travel would be everywhere by the 21st century (I don't know about you but I'm still waiting on my hoverboard!)

When it comes to hard dates, I would forget it, maybe we'll reach transhumanism this decade or this century but I doubt it currently. I think a better way of approaching it would be to just make the date ineffable in the story.

For how long it takes from the ultra-rich to the average joe to get technology, I would look @ mobile phones for an example. The first handheld was produced in 1973 and it only became much more widespread in the 90's.

However the world is much more connected now with technology much more prevalent, I would halve this as a realistic timescale. Unless you implement plot devices to lengthen it further (by not announcing to the public for many years).

Answer 4

20 Years

...provided you started right now. Existing compute architectures are insufficient in terms of compute power and portability to duplicate what the human bio-brain does now. The OP and many other answers state very high numbers for CPU, RAM and power requirements. These are basically large data centers that require plenty of highly skilled (and very expensive) staff to run. Only the richest of people would be able to afford the costs of running their mind on standard Intel hardware. Even if it were possible, most people rich enough to afford this kind of compute environment would only use it if it were a last grasp effort at survival. There are too many trade-offs.

New Compute Architecture Required

Modern CPUs are general purpose compute engines. They are extremely flexible in how they operate and the kinds of operations they can perform. This flexibility comes with trade-offs. CPUs are ill-adapted for doing the calculations required for high speed graphics. GPUs were invented to handle the specialized and highly parallelized calculations required for 3D graphics. Cryptography and Bitcoin mining are two other examples of where CPUs are at significant disadvantages to specialized hardware.

A new architecture is required to meet the specialized requirements of simulating/duplicating neural networks. Starting with field-programmable gate array (FPGA), program a FPGA to duplicate the behavior of neuron models for simple animals like mice. Once feature parity is achieved, move up to more complicated life forms. At some point, custom silicon will be required.

As a way to make this commercially viable, any hardware that facilitates machine learning will be of intense interest to certain companies starting with G, F, A and M. Use the investment from those companies to design higher performance silicon. It may end up that some parts of human cognition are actually better done on normal CPUs or GPUs, in which case a hybrid system would work best.

A primary characteristic of meat-brains is their massively parallel nature. All billion or so neurons can operate at the same time. In contrast, silicon computers usually only do a few things at the same time. GPUs are an extremely good at parallel operations but they usually handle mere thousands of parallel ops. Duplicating this kind of massive parallelism in silicon will be tricky at best.

Timeline

• 2 to 5 years of work in FPGAs
• 2 to 5 years to translate FPGA discoveries to custom silicon
• 10 years to refine custom silicon and adapt it to the needs of the market as well as the requirements of hosting a human mind.

Conclusion

Uploading a person's mind will be feasible when the compute hardware and exoskeleton to house the compute hardware are sufficiently to approximate human thinking and human physical activity. No one will trade-across to a metal-brain if that metal brain isn't at least as good as their meat-brain. Likewise, they won't trade to a metal-body if it isn't at least as good as their meat-body.

---

Secondary Considerations

You Require More Minerals (for batteries)

Robotists are currently looking for three things: Better power sources, better muscle/actuators and better control schemes. Sticking a human mind (and related metal-based compute hardware) into a robot solves the better control scheme. With the descending price/ascending power of lithium-ion batteries in ten years or so, the power storage problem for robotics should be "solved". (Alternatively, make a small fusion reactor and don't worry about batteries again.) It's going to take some interesting chemistry to get decent muscles although recent research into fishing line muscles might prove a fruitful avenue.

Answer 5

You already have set a ballpark figure for the computing power needed, so all thats left is to figure out a ballpark numbers for the power efficicency per watt and what you deem "affordable" energy cost for the uploaded mind.

The trend for computing power is that flops per watt do currently increase by approx. one order of magnitude every few years: https://en.wikipedia.org/wiki/Performance_per_watt#FLOPS_per_watt Very roughly estimating, we can get about 10¹⁰ flops per watt today with the most energy efficient technology. That means at 1 gigawatt consumption, we could get the job done today.

The cost of energy can be as low as \$50 per mega watt hour and while I haven't found a nice historical graph how the cost has been changing over time, again the trend is that electricity gets cheaper as technology matures: https://en.wikipedia.org/wiki/Cost_of_electricity_by_source#Renewables The aforementioned gigawatt/hour could cost approx. \$50,000 per hour if we could secure us a nice deal with a power plant (or rather, we built a plant for our exclusive use... it seems about right at $438m per year). Not affordable for the average person, but already feasible for really rich people.

You need to settle on what you think the average "mind" can afford as cost of "conciousness" per hour, then extrapolate the flops/w trend to the point where that cost can be covered by the extrapolated energy price.

If you take these trends very optimistically, 10¹⁹ flops could cost as little as $50/h in about 15 years.

But lets take into account that energy production may not keep pace with the demand and therefore energy prices might even go up because the huge demand of the uploaded minds exceeds production. Based on this graph http://data.worldbank.org/indicator/EG.USE.ELEC.KH.PC about 3 Mwh are consumed per year per person. That comes down to about 0.35kw/h per person today. Scaling our infrastructure up to support billions of uploaded minds consuming much more electricity would drive the cost of electricity upwards, someone has to pay for all the extra power plants we would need.

So the world electricity production may put a hard cap on how many uploaded minds can be supported, regardless of the actual monetary cost of electricity.

I personally think mass uploaded minds would only be feasible if a mind can fit into a regular server rack, with comparable electricity consumption (lets say 1kw max.). That could be met in about 30 years, if trend for energy efficiency holds

Answer 6

The answer is now, if what you propose is an artificial intelligence that simulates a human mind. The only issue is how accurate the emulation must be to satisfy the criteria and what limits are acceptable as regards scenarios where the simulation takes place. The question talks about "neocortex" and "matching" the brain but a mechanical simulation of a biological thing usually looks and works nothing like the biological thing: consider a dialysis machine and a kidney. The AI must respond to stimuli in the same way that the original biological did. It must take initiative and do things spontaneously or via its own internal cues in a way that mimicks the original. Many very sophisticated statistical predictive models already exist - for example the ones that govern stock trading. They mimick smart traders (probably an amalgam of several individuals). These are artificial intelligences that predict and execute the behaviors of an individual given a set of external circumstances and stimuli.

Of course there is no "porting". It would be code. One would need to put the rules into such an AI as rules. But one could make an AI that wrote like Mark Twain by uploading a lot of his writings together with rules on how to use them. One could program the AI with rules of behavior to mimic the individual in question or program the AI to learn the behaviors or both. Perhaps the individual has long interactions with the AI whereby it learns behavior. If done today this would be via keyboard. There are already attempts to simulate specific individuals via the Turing test. I am nearly certain that there are AI participants in various online environments. An emulated individuals might not necessarily know she was being emulated.

I can imagine an interaction via text or computer with an AI programmed to simulate a dead friend, and that interaction making me happy because it was true to life. The recorded hacker from Gibson's Neuromancer would be the ultimate step: a static recording capable of the same sorts of interactions (within a circumscribed sphere) as the individual.

I think the real question here might be "when will there be immortality". If the ultimate goal is a recording such that there is uninterrupted subjective experience of self from the biological brain to the AI, that is a lofty goal. Given that subjective experience of self is hard to quantify for ones own self, much less for another entity I am not sure how one would determine if that endeavor were ever successful.

Answer 7

I'm reading a book called "Superintelligence: Paths, Dangers, Strategies" by Nick Bostromright now and I believe it would be super beneficial as a knowledge base for writing on this topic. The book is not about uploading a specific mind, but rather superhuman artificial intelligence. However, modeling human intelligence is a step towards Superintelligence in AI (which surpasses human intelligence) and many of the issues you are trying to get an answer to also need to be tackled on the way towards Superintelligence. I havn't finished it yet though, but I'll still try giving you the fractions of what I've learned here.

Timeline: Expert’s opinions vary greatly, however most guesses for the existence of Superintelligence lie somewhere around 90% certainty by 2100. Depending on how driven the research for creating artificial replicas of people's personality (I'll call it AP from now on) is, it would probably lie somewhere around that time frame that it would have become a more solid technology.

Financing: I'd go with the assumption that super powered efficient computing (quantum computing maybe?) is readily available in the previously described timeline. Therefore I believe the main issue would actually be the running costs. But seeing as we are talking about a person's mind that is probably still able to interact with its environment, at least in a digital manner (messages, E-Mail, etc.) this artificial mind could actually still be part of the active work force! If a live mind was able to do an accountants job, formatting excel, sending e-mails, etc. then an AP could do the exact same job! APs could thereby simply continue working in order to maintain their existence. This of course only works for intellectual work, not manual work. However a continuing automation of manual labor (substitution through robots, etc.) also means that more and more people would be forced to go into academic jobs that do not include any manual work to a point, where the majority of jobs in society consist of them. Having APs as an addition to the job market would of course lead to a massive over saturation and lead to all kinds of interesting issues. Are APs paid the same although they require no food or material goods? Should live humans be preferred over APs? You'd have great potential for a segregation of society.

I hope you found this helpful to some extent. Read the book, it's fantastic!

Regarding Footnote: When explaining how the scanning technology works keep in mind that there are two major problems in creating AP:

1. Modelling the known connections into a meaningful and functional logical model. The Roundworm's (a super simple organism) neural network has been know since the 80s and is still not fully modeled! Check out Open Worm Worm for more Info.

2. Gaining knowledge of all the human connections present in the brain is an enormous task. Even the brain of the common fruit fly has only just been mapped and it's tiny compared to the human one. Automation of that process is essential in order to achieve AP in a meaningful and quick manner (read this article for more info).

Also this woman is working on cretaing AP and making pretty impressive headway. Watch her TED Talk for more info.

Answer 8

Never.

The problem is not being able to simulate a human brain; that one I consider possible, and probably even achievable on affordable hardware at some time in the future. The problem is taking an existing brain and reading out the full state. Since the properties of the person are not just in the electrical activity, but in the very structure of the brain, you cannot determine the relevant parameters without destroying the brain, and you cannot read all of it at once. But if you do it piecewise, you will destroy parts of the brain, which will cause other parts of the brain to malfunction before you get around to reading them. So what you get will be a severely damaged copy of the brain, and since you can't know the pre-damage state, you cannot restore it.