1. Introduction
A digital mind is a sense that runs on a computer. One kind of digital sense is a mentality upload. This hypothetical sense was initially human but moved into a digital format and is being run as a software program on a computer. Another type of digital mind is that of artificial general intelligence (AGI). While uploads stand established on bringing existing human senses and near repeating them in software, AGI may be built on computer science principles and have little or no resemblance to the human psyche. Either type of digital mind might be created within decades to centuries. A recent roadmap charting the technological needs for making uploads proposes they may be possible by mid-century (Sandberg & Bostrom, 2008). Sotala and Valpola (2012) state that research into prostheses reproducing the parts of the hippocampus and the cerebellum is well underway and suggest that a feasible future development would be “excorticates,” implants that can be connected to human brains, which gradually take over cortical brain function.
Interest in AGI research is also growing, with an increasing number of special sessions, workshops, and conferences explicitly devoted to topics such as AGI that have been held in recent years (Baum et al., 2011). In a professional examination survey performed at the Artificial General Intelligence 2009 (AGI-09) gathering, the median assessments for when there would be a 10%, 50%, or 90% chance of holding an AGI capable of handling the third stage were in 2020, 2030 and 2075, respectively (Baum et al., 2011). In an informal survey conducted at the 2011 Winter Intelligence Conference, the median estimates for when there would be a 10%, 50%, or 90% probability of designing human-level machine intelligence existed in 2028, 2050, and 2150 (Sandberg & Bostrom, 2011).
Some previous work has focused on examining the consequences of creating digital intelligence. Focusing specifically on uploads, researchers have read some of the economic implications of the ability to copy minds (Hanson 1994, 2008) and the improved coordination capacity stemming from being capable of reproducing, deleting, and restoring senses (Shulman, 2010). Hanson (1998) and Kaas et al. (2010) look more generally at the economic effects of digital reasons that can be copied. Sotala and Valpola (2012) research how mind uploading may guide “mind coalescence,” the capacity to connect formerly different minds.
Other researchers have claimed that once we have AGI, it will exceed the capabilities of humans in multiple domains at an extreme speed (see, e.g., Bostrom [2003], Vinge [1993], Yudkowsky [2008a], and Chalmers [2010]). A “hard launch” (Yudkowsky, 2001; Bugaj & Goertzel, 2007; Hall, 2008; Vinge, 2008) applies an AGI getting a fact that permits it to quickly accumulate different advantages and influence, becoming a predominant power before humans have the time to respond appropriately. Should existing human preparation be inadequate for such a drastic event, there could be severe consequences, up to and including human extinction (Bostrom, 2002; Yudkowsky, 2008a; Chalmers, 2010). This paper attempts to study the effects of creating a digital mind in terms of the advantages that they may enjoy over humans.
1.1. Intelligence, Optimization Power, and Advantages
Various definitions have been offered for the word “intelligence.” A report summarising the most essential content is that intelligence measures an agent’s ability to achieve goals in various environments (Legg & Hutter, 2007). A mind has plans that it tries to achieve, and more intelligent reasons are better at finding, inventing, and evaluating various ways of achieving their goals. A generalization of the concept of intelligence is the notion of optimization power (Yudkowsky, 2008a; Muehlhauser & Helm, 2012), an agent’s general ability to achieve its goals. While intelligence is derived from what is generally considered “mental” faculties, an agent’s optimization power is also a factor of its allies and resources and its ability to obtain more of them.
The crucial risk involved in creating digital minds is the possibility of creating reasons whose goals are very different from humanity’s and who end up possessing more optimization power than society. Recent human intentions and desires seem to be very difficult and not well-understood. There is a mighty chance that only a minimal subset of all possible plans will, if victorious, lead to consequences that humans would consider favourable (Yudkowsky, 2008a; Muehlhauser & Helm, 2012). If digital intelligence is created and has (as a group) more optimization power than humanity, its goals are very different from humanity’s goals. The results are likely to be regarded as very bad by most humans.
This paper analyzes the consequences of creating a digital mind from the perspective of the optimization power they might accumulate. Factors that may lead to digital reasons getting more optimization power than humans are called advantages; factors that may lead to digital minds earning less optimization power than humans are called disadvantages.
- Hardware Advantages
A digital mind running on a computer can upgrade the system to utilize more powerful hardware, while biological humans cannot drastically upgrade their brains. Suppose that some minimum hardware configuration provides a digital mind with roughly the same processing power and memory as a human brain. Any increase in hardware resources past this point is a hardware advantage favouring the digital sense.
1.1. Superior Processing Power
The processing power required to run a digital mind is still being determined.[1]For uploads, Sandberg and Bostrom (2008) place 1018 to 1025 FLOPS as the most likely amount required to run one in real time. Current movements would open these levels for purchase at 1 million dollars near 2019 to 2044.
The goal of uploading is to accurately replicate the functions of the entire human brain, including all necessary details. In contrast, AGI designers can use any working algorithm, regardless of its biological plausibility. The human brain has evolved to operate within the constraints of biology, which may differ greatly from what would run efficiently on a computer. This leaves the possibility that an AGI would require much less processing power than an uploaded brain. There are varying estimates on the amount of processing power needed to power the human mind. These assessments run from current-day computers (Hall, 2007) to 1011 FLOPS (Moravec, 1998) and 1014 FLOPS (Bostrom, 1997).
1.1.1. Superior Serial Power
Humans sense the world on a particular expected time hierarchy. A mind being executed on a system with significantly superior serial power could run on a faster timescale than we do. For instance, a sense with twice the serial control of the human brain might experience the equivalent of two seconds passing for each second that we did, thinking twice the amount of thoughts simultaneously. This benefit would be precisely observable in time-critical decision-making.
Even a tiny advantage would accumulate, given enough time. Over a year, a 10% difference in speed would give the faster mind more than an extra month. This would allow it to outcompete any reason with equal skills and resources without the speed advantage.
In the “speed explosion” scenario (Solomonoff, 1985; Yudkowsky, 1996; Chalmers, 2010), digital researchers, running at an accelerated speed, work to develop faster computers. If the minds doing the research could take advantage of the quicker hardware they produced, the time required to create the next generation of hardware could keep getting shorter as the researchers would be getting more done simultaneously. This could continue until some bottleneck, such as the time needed to build the computers or a fundamental physical barrier was reached.
1.1.2. Increased Parallel Power, Increased Memory
Current advancements in computing capacity have been increasingly similar rather than serial. If the trend continues, a future computer may not use its superior processing power to increase speed over the human brain if the tasks in question do not parallelize well. Amdahl’s law states that if a fraction f of a program’s performance can be parallelized, then the speedup given by n processors instead of one is 1/((1−f)+ f/n) (Amdahl, 1967). A difference of several charges of importance in computing power might translate to a much more modest modification in speed. Gustafson (1988) notes that, in practice, the parallelizable part of a problem grows as data is added, and the serial part remains constant. Even if increasing the number of processors did not allow a problem to be solved in less time, it can enable a more significant issue to be solved in better detail.
As the human brain works massively parallel, at least some highly parallel algorithms must be involved with general intelligence. Extra similar power might not allow for a direct improvement in speed, but it could provide something like a greater working memory equivalent. More trains of thought could be pursued at once, and more things could be considered when considering a decision. Brain size correlates with intelligence within rats (Anderson, 1993), humans (McDaniel, 2005), and across species (Deaner et al., 2007), suggesting that increased parallel power could make a mind generally more intelligent. Some sources estimate MIPS (Millions of Instructions Per Second), while others use FLOPS (Floating-Point Operations per Second). These are not directly similar, and there is no dependable way to transform between the two. For this paper, we have used the rough estimate in Sandberg and Bostrom (2008) that FLOPS increase as MIPS to the force of 0.8. The authors alert that this tendency may reverse, with the booster may be evolving bigger than 1.
