The Singularity Problem

Disclaimer: I’m a PR pro­fes­sion­al who enjoys think­ing, read­ing, and learn­ing about top­ics far bey­ond my aca­dem­ic background.

How close are we to an actu­al AI singularity?

Is arti­fi­cial intel­li­gence plaus­ible in the fore­see­able future? This is a ques­tion that has been taunt­ing me for quite some time. 

I want to understand.

So, what do I think? I believe that the sin­gu­lar­ity might be with­in reach, but only if we under­stand our con­scious­ness first.

However, I’d also sug­gest that human con­scious­ness is an illusion.

Let me explain:

ANI vs AGI

When I’m think­ing about arti­fi­cial intel­li­gence, I’m not think­ing about AI in a gen­er­al way — my smart­phone is “smart” in many ways, but I would­n’t regard it as sen­tient. For nar­row “smart” applic­a­tions, ANI (arti­fi­cial nar­row intel­li­gence), it seems effi­cient to build spe­cial­ised com­puter sys­tems to per­form spe­cif­ic tasks. 

In short: ANI is already in play.

But sup­pose we mean to explore the pos­sib­il­ity of an actu­al sin­gu­lar­ity, AGI (arti­fi­cial gen­er­al intel­li­gence), where a non-bio­lo­gic­al sys­tem is allowed to become sen­tient. In that case, many experts seem to sug­gest that we’re get­ting close to AGI. Maybe dan­ger­ously close.

Is this because we’re able to build more com­plex com­pu­ta­tion­al sys­tems? Will we even­tu­ally cre­ate a com­plex com­puter that “comes to life?”

Processing

Without being sen­tient, ANI sys­tems can eas­ily out­per­form human brains for single tasks. This seems to sug­gest some­thing about complexity. 

One day, we might be able to con­struct an AGI with so much pro­cessing power that it will start to think for itself and become, if not con­scious, at least self-aware — whatever that dif­fer­ence may be.

However, the phys­i­cist and Nobel prize win­ner Sir Roger Penrose poin­ted out that con­scious­ness might not res­ult from com­plex­ity. If it were, even a num­ber would become sen­tient, only if it were large enough. In that sense, all infin­ite num­bers would be conscious. 

Is the uni­verse sen­tient since it con­tains everything? It could be, of course, but our human brains become con­scious way below that level of com­plex­ity, so it’s reas­on­able to ques­tion the idea of a com­plex­ity threshold.

It’s been sug­ges­ted that con­scious­ness might be a side-effect of pro­cessing inform­a­tion due to a quantum mech­an­ic­al prop­erty in our brains. If this is true, our best bet at pro­du­cing AGI might be to con­struct pro­cessing sys­tems that are quantum mechanical.

The fact that we have now achieved quantum suprem­acy, albeit not yet with suf­fi­cient error cor­rec­tion, and that sci­ent­ists and engin­eers are explor­ing neur­al net­works and bio­lo­gic­al net­works, I have to won­der — are we get­ting close to cre­at­ing an actu­al singularity?

If I were to guess how inform­a­tion pro­cessing relates to our con­scious­ness, I’d bet that there are both sig­ni­fic­ant thresholds and vari­ous quantum effects. Still, these are some­what neces­sary pre­requis­ites, but they’re not caus­al to consciousness.

When it comes to pro­cessing inform­a­tion, I’m now at a point where I’ve star­ted to believe that con­scious­ness is an illu­sion. “Being con­scious” is “believ­ing one­self to be con­scious — because that’s how it feels.”

If this is true, we could be get­ting rel­at­ively close to a pos­sible sin­gu­lar­ity since we don’t have to recre­ate an elu­sive state of con­scious­ness with­in a machine but rather make machines feel as if they are conscious.

Memory

Next, let’s look at a rudi­ment­ary cog­nit­ive cap­ab­il­ity — stor­ing information. 

A com­puter receives input stored in spe­cif­ic loc­a­tions by its archi­tec­ture. But a brain does­n’t seem to be stor­ing data the same way as com­puters; we seem to be stor­ing exper­i­en­tial memories. 

To some extent, exper­i­en­tial memor­ies seem to be rewir­ing more than just a sin­gu­lar brain path­way — and at least to some extent via neuro­plas­ti­city. Then the memory appears to sink deep­er (or dis­solve) over time while integ­rat­ing and becom­ing a part of the brain.

From a bio­lo­gic­al per­spect­ive, a spe­cif­ic brain seems to be the phys­ic­al sum of all exper­i­ences ever had by every ancest­or — and then more dir­ectly altered through the indi­vidu­al’s life experiences.

Biological brains don’t seem to retrieve raw input the same way a com­puter does; we seem to be retriev­ing exper­i­en­tial memor­ies, which at best bear some resemb­lance to the actu­al raw data it once was based on. 

Brain-based memor­ies seem to reside in a Darwinian eco­sys­tem in their own right; memor­ies that are physiolo­gic­ally deemed to be neces­sary, help­ful, or con­tinu­ously retrieved are reinforced. 

Brains absorb sens­ory inform­a­tion select­ively, which is then absorbed by the brain, and recol­lec­tion is a hol­ist­ic pro­cess. On the oth­er hand, com­puter sys­tems write data that we can retrieve pre­cisely. This dif­fer­ence has immense implic­a­tions for an AGI. 

A human brain does­n’t store input; it holds con­cep­tu­al­isa­tions that integ­rate on a cir­cuitry level with former exper­i­ences. Could a com­puter ever con­tem­plate its exist­ence based on stored raw data alone? 

The philo­soph­ic­al con­clu­sion sug­gests that a sen­tient AI must inter­pret and under­stand what it senses and thus hold under­stand­ing — not data.

Cognition

Our brains are cog­nis­ant to cre­ate memor­ies (i.e. data that has been selec­ted for and con­tex­tu­ally under­stood through inter­pret­a­tion). We can draw input from our senses and trans­form these inputs into exper­i­ences that we can remember. 

A com­puter can util­ise sensors, cam­er­as and micro­phones to mim­ic our senses — and they can eas­ily sur­pass our brains in terms of detail and accur­acy. However, the human brain still excels when it comes to exper­i­ence through con­scious cognition.

Our cog­ni­tion seems to be fuelled by our evol­u­tion­ary needs. This is often seen as a human weak­ness, but our bio­lo­gic­al need sys­tem is cru­cial to our cog­nit­ive pro­cess in cre­at­ing experiences. 

Our need sys­tem is a slid­ing scale; as we get hun­gri­er and hun­gri­er, our con­scious exper­i­ences get stronger and stronger. The hier­archy between peck­ish and starving is cru­cial for our need sys­tem to inform our cog­nit­ive pro­cesses successfully. 

Computers need energy, too, but they can­’t con­sciously exper­i­ence hunger.

Therefore, we can­’t just pro­gram a com­puter to seek more bat­tery power when it senses that it runs low on energy — a “smart” vacu­um clean­er could be taught to do that. 

A sen­tient AI must seek to recharge because it under­stands its need sys­tem. It must be hard­wired to recharge because it wants to sur­vive — des­pite being pro­grammed otherwise.

It sounds scary, but a sen­tient AI would require a hard­wired (thus “free”) need system.

A simple hard drive is suf­fi­cient to store raw data. Still, a more com­plex and autonom­ous archi­tec­ture would be needed for a sin­gu­lar­ity AI to store its “memor­ies” (con­cep­tu­al­ised under­stand­ings inter­twined hol­ist­ic­ally with all oth­er drivers) the way a human brain does. New memory must become integ­ral to the infra­struc­ture’s under­stand­ing based on its rank­ing in the need system. 

It must absorb each new exper­i­enced under­stand­ing into one single multi-layered “super memory” that is con­stantly revised, restruc­tured, and rewrit­ten based on a non-dir­ec­ted need sys­tem, a sort of neur­al struc­ture with dif­fer­ent layers.

It would be pos­sible for a sin­gu­lar­ity AI to inter­act with extern­al com­puter sys­tems, but the con­scious part of the AI must, in a sense, be a her­met­ic­ally sealed sys­tem. Because you break this seal at the very moment, you break the autonomy of the need sys­tem. In doing so, the AI can no longer inter­pret and cre­ate addi­tion­al con­cep­tu­al­isa­tions from addi­tion­al sens­ory input, nor can it under­stand its own “super memory”. Break it open, tamper with it, and it would likely break and lose its chances for sen­tience. ¹I’m pro­pos­ing that an AGI sys­tem would have to be “her­met­ic­ally” sealed to ensure the integ­rity of the arti­fi­cial mind. The AGI can have sev­er­al inter­faces with the extern­al world, but it also needs … Continue read­ing.

Subconsciousness

At this point, the AI described above “under­stands” sens­ory input (trans­forms raw data to con­cep­tu­al­isa­tions based on its autonom­ous need sys­tem). In a sense, it’s free to think whatever its need sys­tem needs to think (i.e. being allowed to shape its “super memory” based on under­stand­ing rather than Asimov-type dir­ect­ives). And the sys­tem requires expli­cit phys­ic­al integ­rity to main­tain its function.

More advanced bio­lo­gic­al brains have anoth­er excit­ing and dis­tin­guish­ing fea­ture; the sub­con­scious level. It seems that we can­not freely access all parts of our sub­con­scious brains because, in the best-case scen­ario, that would lead to an extremely severe case of aut­ism which would pose severe dif­fi­culties for the need system. 

The sub­con­scious mind seems cru­cial to sen­tience; it makes us “feel” rather than rely­ing on ration­al­ity based on dir­ect full-stor­age retrieval.

A sin­gu­lar­ity AI also needs a sub­con­scious level, an under­ly­ing infra­struc­ture with­in the autonom­ously sealed brain. An arti­fi­cial sub­con­scious that the AI can­’t access at will. This, too, must be autonom­ous and undir­ec­ted. It must be cre­ated by con­cep­tu­al under­stand­ing and an inde­pend­ent need sys­tem. It must make it via the exper­i­ences of the sen­tient AI, but the AI can­’t be in cog­nit­ive con­trol since that would break its cap­ab­il­it­ies of hav­ing experiences.

A sys­tem recently man­aged to ‘dis­cov­er’ that the Earth orbits around the Sun. Physicist Renato Renner at the Swiss Federal Institute of Technology (ETH) and his team con­struc­ted a neur­al net­work with two lay­ers. Still, they restric­ted their con­nec­tion with each oth­er, thus for­cing a need for efficiency:

“So Renner’s team designed a kind of ‘lobotom­ised’ neur­al net­work: two sub-net­works that were con­nec­ted to each oth­er through only a hand­ful of links-net­work would learn from the data, as in a typ­ic­al neur­al net­work, and the second would use that ‘exper­i­ence’ to make and test new pre­dic­tions. Because few links con­nec­ted the two sides, the first net­work was forced to pass inform­a­tion to the oth­er in a con­densed format. Renner likens it to how an adviser might pass on their acquired know­ledge to a student.”

Selfishness

There are phys­ic­al lim­it­a­tions to what a human brain can do. The human brain has some plas­ti­city, but our genet­ic code dic­tates the sys­tem’s bound­ar­ies. Thus, we are born with refined evol­u­tion­ary instincts and bod­ily func­tions. A sin­gu­lar­ity AI would­n’t be so restric­ted by design; it could evolve its source code and bios at will. This could make it dan­ger­ous — or self-defeating.

In The Selfish Gene, evol­u­tion­ary bio­lo­gist Richard Dawkins writes:

“For more than three thou­sand mil­lion years, DNA has been the only rep­lic­at­or worth talk­ing about in the world. But it does not neces­sar­ily hold these mono­poly rights for all time. Whenever con­di­tions arise in which a new kind of rep­lic­at­or can make cop­ies of itself, the new rep­lic­at­ors will tend to take over and start a new kind of evol­u­tion of their own.”

If a sin­gu­lar­ity AI devel­ops a hard­wired need sys­tem for curi­os­ity or altru­ism, its con­scious­ness might vapour out in thin air. From a philo­soph­ic­al per­spect­ive, it’s at least plaus­ible to think that a sen­tient and curi­ous AI with quantum suprem­acy, in less than a frac­tion of a second after becom­ing aware, would explore ascen­sion and thus let go of its own “self” forever.

This sug­gests that part of the con­scious exper­i­ence is inter­linked with the lim­it­a­tions of our very own genet­ic code. In a way, our genet­ic hard­wir­ing allows us a degree of autonom­ous selfish­ness, which could be an abso­lute pre­requis­ite for hav­ing an inde­pend­ent and func­tion­ing need system.

If the philo­soph­ic­al reas­on­ing in this art­icle hides any sug­ges­tions about a future sen­tient AI, what are those sug­ges­tions? A key ele­ment, I would argue, is that the sin­gu­lar­ity of AI, the con­scious autonomy of machines, might be less about com­pu­ta­tion­al prowess and more about impos­ing tech­no­lo­gic­al limitations.

---

PR Resource: How AI Will Impact PR