I trafficoni dell’AI

I trafficoni dell’AI

The Hanson-Yudkowsky AI-Foom Debate – Robin Hanson and Eliezer Yudkowsky

***

Trigger warning: domani potranno essere costruiti robot più intelligenti di noi che sapranno costruire robot più intelligenti di loro. Come finirà questa storia? – emulatori o robot? – analogia o modelli? – due stili cognitivi per pensare al futuro lontano – il debole legame tra scienza e innovazione – tre modelli di innovazione radicale a cui ispirarsi per le analogia: nascita dell’homo sapiens, nascita dell’agricoltura, nascita dell’industria –

***

Chapter 1 Fund UberTool?

Sometimes a set of tool types will stumble into conditions especially favorable for mutual improvement.

Note:IL MIGLIORAMENTO RECIPROCO: MEZZI CHE COSTRUISCONO ALTRI MEZZI… E LA CRESCITA ESPLODE

Such favorable storms of mutual improvementusually run out quickly, however, and in all of human history no more than three storms have had a large and sustained enough impact to substantially change world economic growth rates.

Note:STORIA UMANA: TRE TEMPESTE A CUI ISPIRARSI PER PREVEDERE

Imagine you are a venture capitalist reviewing a proposed business plan. UberTool Corp has identified a candidate set of mutually aiding tools, and plans to spend millions pushing those tools through a mutual improvement storm.

Note:SCENARIO: SE UN’ IMPRESA CONQUISTASSE IL MONDO

UberTool does not plan to stop their closed self-improvement process until they are in a position to suddenly burst out and basically “take over the world.” … Now given such enormous potential gains, even a very tiny probability that UberTool could do what they planned might entice you to invest in them.

Note:PROBABILITA’ DI INVESTIMENTO NELL’AI FOOM: ELEVATE

Chapter 2 Engelbart as UberTool? Un antesignano

Yesterday I described UberTool, an imaginary company planning to push a set of tools through a mutual-improvement process;

Note:L’ IDEA DEI PC CHE COSTRUISCONO PC.

Augmenting Human Intellect: A Conceptual Framework … He understood not just that computer tools were especially open to mutual improvement… [Engelbart] is best known for inventing the computer mouse … Now to his credit, Doug never suggested that his team, even if better funded, might advance so far so fast as to “take over the world.” … Doug Engelbart understood what few others did—not just that computers could enable fantastic especially-mutually-improving tools, but lots of detail about what those tools would look like. …

Note:DOUG CONQUISTERÀ IL MONDO?

Chapter 3 Friendly Teams

Just as humans displaced chimps, farmers displaced hunters, and industry displaced farming, would a group with this much of a head start on such a general better tech have a decent shot at displacing industry folks? And if so, shouldn’t the rest of the world have worried about how “friendly” they were?

Note:LA CATENA DEGLI SPIAZZAMENTI. DOBBIAMO TEMERE IL VINCITORE?

In fact, while Engelbart’s ideas had important legacies, his team didn’t come remotely close to displacing much of anything. He lost most of his funding in the early 1970s, and his team dispersed.

Note:IL FALLIMENTO DI ENGELBART

But what makes that scenario reasonable if the UberTool scenario is not?

Note:LA BRUTTA FINE DI ENGELBART CI RASSICURA SUL MONOPOLISTA CATTIVO?

Chapter 4 Friendliness Factors

how much better will the best firm be relative to the average, second best, or worst?

Note:TENDENZA AL MONOPOLIO

Here are a few factors: …Resource Variance—The more competitors vary in resources, the more performance varies. … Cumulative Advantage—The more prior wins help one win again, … Lumpy Design—The more quality depends on a few crucial choices, relative to many small choices, the more quality varies. … Interdependence—When firms need inputs from each other, … Info Leaks—The more info competitors can gain about others’ efforts, the more the best will be copied, reducing variance. … Legal Barriers… Anti-Trust… Network Effects—Users may prefer to use the same product regardless of its quality. …

Note:DA COSA DIPENDE L’ESISTENZA DEL MONOPOLIO?

Some key innovations in history were associated with very high variance in competitor success. For example, our form of life seems to have eliminated all trace of any other forms on Earth.

Note:I PRECEDENTI. L’ HOMO SAPIENS SEMBRA DOMINARE INCONTRASTATO

On the other hand, farming and industry innovations were associated with much less variance.

Note:AGRICOLTURA E INDUSTRIA

attribute this mainly to info becoming much leakier, in part due to more shared standards,

CI SALVERÀ LA FLUIDITÀ DELL INFO?

If you worry that one competitor will severely dominate all others in the next really big innovation, forcing you to worry about its “friendliness,” you should want to promote factors that reduce success variance.

Note:CONSIGLIO POLITICO

Chapter 6 Setting the Stage (come ragionare per prevedere: analisi o analogie?)

We seem to agree that: … Feasible approaches include direct hand-coding, based on a few big and lots of little insights, and on emulations of real human brains.

Note:LE DUE VIE VERSO LA IA: 1 PROGRAMMAZIONE 2 EMULAZIONE DEL CERVELLO UMANO

Machine intelligence will, more likely than not, appear within a century,

Note:ENTRO UN SECOLO. PROB. SUP. 50%

Math and deep insights (especially probability) can be powerful relative to trend fitting and crude analogies.

Note:MATH, PROBABILITÀ E ANALOGIE

Some should be thinking about how to create “friendly” machine intelligences.

Note:LA QUESTIONE CENTRALE

We seem to disagree modestly about the relative chances of the emulation and direct-coding approaches;

Note:IL DISACCORDO

Our largest disagreement seems to be on the chances that a single hand-coded version will suddenly and without warning change from nearly powerless to overwhelmingly powerful; I’d put it as less than 1% and he seems to put it as over 10%…. My style is more to apply standard methods and insights to unusual topics. So I accept at face value the apparent direct-coding progress to date, and the opinions of most old AI researchers …

Note:LO STILE INTUITIVO

putting apparently dissimilar events into relevantly similar categories. … These  suggest a single suddenly superpowerful AI is pretty unlikely.

Note:LO STILE ANALOGICO: PIU’ PROBABILE LA DIVERSITA’

Eliezer seems to instead rely on abstractions he has worked out for himself, not yet much adopted by a wider community of analysts, nor proven over a history of applications to diverse events.

Note:RAZIONALISMO SPECIFICO DELL’INGEGNERE

Chapter 8 Abstraction, Not Analogy

I’m not that happy with framing our analysis choices here as “surface analogies” versus “inside views.”

Note:SURFACE VS INSIDE VIEW… DISTINZIONE SVIANTE

More useful, I think, to see this as a choice of abstractions. An abstraction (Wikipedia) neglects some details to emphasize others.

Note:MEGLIO: ASTRAZIONE VS ANALISI SPECIFICA

For example, consider the oldest known tool, the hammer (Wikipedia). To understand how well an ordinary hammer performs its main function, we can abstract from details of shape and materials. To calculate the kinetic energy it delivers, we need only look at its length, head mass, and recoil energy percentage (given by its bending strength). …To see that it is not a good thing to throw at people, we can note it is heavy, hard, and sharp. To see that it is not a good thing to hold high in a lightning storm, we can note it is long and conducts electricity. To evaluate the cost to carry it around in a tool kit, we consider its volume and mass. … Whether something is “similar” to a hammer depends on whether it has similar relevant features. …

Note:ESEMPIO: IL MARTELLO

The issue is which abstractions are how useful for which purposes, not which features are “deep” vs. “surface.”

Note:LA QUESTIONE INFINE

The future story of the creation of designed minds must of course differ in exact details from everything that has gone before. But that does not mean that nothing before is informative about it.

Note:AI E LE ANALOGIE

Yes, when you struggle to identify relevant abstractions you may settle for analogizing… Analogies are bad not because they use “surface” features, but because the abstractions they use do not offer enough relevant insight for the purpose at hand.

Note:IL DIFETTO DELL’ ANALOGIA

I claim academic studies of innovation and economic growth offer relevant abstractions for understanding the future creation of machine minds,

Note:AI. ANALOGIE CON INNOVAZIONE E CRESCITA ECONOMICA

previous major transitions, such as humans, farming, and industry, are relevantly similar.

Note:ANALOGIE: HOMO SAPIENS AGRICOLTURA INDUSTRIA

You have previously said nothing is similar enough to this new event for analogy to be useful, so all we have is “causal modeling” (though you haven’t explained what you mean by this in this context). This post is a reply saying, no, there are more ways using abstractions; analogy and causal modeling are two particular ways to reason via abstractions, but there are many other ways.

Note:MODELLO CAUSALE… O INGEGNERISTICO

Everything is new to us at some point; we are always trying to make sense of new things by using the abstractions we have collected from trying to understand all the old things.

IL RADICALMENTE NUOVO

Chapter 10 AI Go Foom

hand-coded AI will come soon and in the form of a single suddenly superpowerful AI.

Note:L IPOTESI DI Y (LA MENO PROBABILE PER H)

A machine intelligence can directly rewrite its entire source code and redesign its entire physical hardware. While human brains can in principle modify themselves arbitrarily, in practice our limited understanding of ourselves means we mainly only change ourselves by thinking new thoughts. All else equal this means that machine brains have an advantage in improving themselves. …“object” vs. “meta” … It seems as if you think object ones don’t increase growth rates while meta ones do. …

Note:UN VANTAGGIO DELL’IPOTESI SOFTWARISTICA: LE MACCHINE SI RIPRODUCONO PIÙ EFFICACEMENTE. EUGENETICA INFORMATICA.

Chapter 12 Eliezer’s Meta-level Determinism (il ruolo della scienza nell’avanzamento umano)

it seems the basis for Eliezer’s claim that my analysis is untrustworthy “surface analogies” vs. his reliable “deep causes” is that, while I use long-vetted general social science understandings of factors influencing innovation, he uses his own new untested meta-level determinism theory.

Note:ANCORA ANALOGIA VS CAUSALITÀ

The last three strong transitions were humans, farming, and industry, and in terms of growth rate changes these seem to be of similar magnitude. Eliezer seems to predict we will discover the first of these was much stronger than the other two. And while the key causes of these transitions have long been hotly disputed, with many theories in play, Eliezer seems to pick specific winners for these disputes: intergenerational culture, writing, and scientific thinking.

Note:LE 3 TRANSIZIONI. IL SEGRETO DEL SUCCESSO UMANO

Few could write and what they wrote didn’t help farming much. Farming seems more plausibly to have resulted from a scale effect in the accumulation of innovations in abilities to manage plants and animals

Note:AGRICOLTURA E CUMULO DELLE CONOSCENZE. NO PENSIERO SCIENTIFICO

Also for industry, the key innovation does not seem to have been a scientific way of thinking—that popped up periodically in many times and places, and by itself wasn’t particularly useful. My guess is that the key was the formation of networks of science-like specialists, which wasn’t possible until the previous economy had reached a critical scale and density.

INDUSTRIA: NETWORK DI TRAFFICONI E MASSA CRITICA DEL CAPITALE. IL RUOLO DELLA SCIENZA E’ SECONDARIO.

Intelligenza e razza

THE BELL CURVE DI CHAROLES MURRAY

Come collasserà lo stato sociale?

Per molti non sarà sostenibile finanziariamente.

Ma altri privilegiano una seconda ipotesi: la scienza dimostrerà che il tentativo di ridurre le diseguaglianze sociali mirando alle pari opportunità è vano poiché tali diseguaglianze si fondano su talenti con una solida base genetica, come l'intelligenza.

Un esempio? Il gap tre neri e bianchi.

Un esempio del genere - anche se presentato in forma di ipotesi -  ha attirato critiche abbondanti quanto superficiali...

... cherry picked studies, ignores copious data, and dismisses the impressive explanatory power of hereditarianism...

Le critiche si articolano intorno a quattro evidenze che ora passeremo al vaglio.

***

1.  Il gap in termini di IQ tra bianchi e neri va chiudendosi.

L'affermazione si fonda su un unico studio, per altro molto problematico...

... Although there is some evidence that the Black-White IQ gap did narrow in the 20th century, the claim that the IQ gap is now 10 points is based largely on the results of a single, heavily criticized paper, and is not consistent with the vast majority of the evidence...

Per altri studi il recupero sul gap è trascurabile e da tempo svanito...

... Murray conducted analyses of the Woodcock-Johnson cognitive test batteries and found no decrease in the Black-White gap for those born after the 1960s (and, in fact, found a slight increase in this gap since the 1960s). He found similar results using data from the National Longitudinal Survey of Youth. Other scholars, examining the available evidence up to 2005 in their response to claims of a narrowing Black-White gap, concluded, “… the best evidence of the Black-White convergence over the past 100 years is between 0 and 3.44 IQ points…”...

Diciamo così: nel XX secolo c'é stato un leggero recupero che si è poi stabilizzato...

... On achievement, the Black-White gap narrowed during the 20th century, but multiple indicators of educational achievement provide convergent evidence that the gap has stopped narrowing and, in some cases, is widening...

Ecco su questo punto un giudizio riassuntivo equilibrato...

... Summarizing research on cognitive ability across the 20th century in employment and educational settings, Roth and colleagues concluded, “As a whole, these studies suggest that there are observed gains for both groups [Blacks and Whites], but the reduction in the between-group difference is either small, potentially a function of sampling error, or nonexistent for highly g loaded instruments.”... the IQ gap has stabilized near 12-15 points...

***

2. Esiste un effetto Flynn. Ovvero, l' IQ misurato nei test incrementa molto di generazione in generazione.

Sottointeso: se l' IQ è dinamico il recupero dei neri diventa possibile...

... Turkheimer and colleagues asserted that IQ scores have increased 18 points in the US from 1948 to 2002. They noted that this increase is larger than the Black-White IQ gap. They then accused Murray of ignoring the importance of this Flynn effect..

Tuttavia, questa speranza è malriposta e si fonda su una incomprensione: IQ e intelligenza non sono la stessa cosa, il dinamismo del primo non implica il dinamismo della seconda...

... secular gains in IQ are inversely correlated with g loadings whereas the Black-White gap is positively correlated with g loadings...

Analogia...

... Consider the following example: Men are taller than women (there is a height gap between men and women). Both men and women have increased in height since the 1700s. So, there is a secular increase in height (“ Flynn effect” for height). But the gap between men and women in height has remained relatively stable (in the US, it has increased slightly in the past 100 years). The cause of the secular increase (probably nutrition) is different from the cause of the disparity between men and women (almost certainly genes)...

***

3. I neri adottati in famiglie di bianchi registrano notevoli vantaggi in termini di IQ.

L'unico studio su cui si fonda questa affermazione ha almeno tre vizi, almeno per chi lo vuole utilizzare contro Murray: 1) considera solo casi di povertà estrema, 2) registra il dinamismo del gap in età infantile quando sappiamo che i gap si stabilizzano successivamente (effetto Wilson) e 3) è comunque compatibile con un'ereditarietà del 60%...

... adoption “from a poor family into a better-off one is associated with gains of 12 to 18 points.” This claim stems from an analysis of six studies with 253 total subjects, and should be interpreted cautiously for several reasons. First, in one of the studies, the “poor families” from which the children were adopted were so poor that the children were suffering from primary undernutrition at the time of adoption. Second, most of the studies examined IQ at young ages, which is important because the genetic effect (heritability) of intelligence increases with age, a phenomenon labelled the “Wilson effect” by behavior geneticist Thomas Bouchard; therefore, the gains might be smaller if the subjects had been tested later in life. Third, even if one accepted the large IQ gains uncritically, those gains are still consistent with heritability estimates of 60%, as was clearly noted by Herrnstein and Murray in The Bell Curve (see page 771, note 86). And, fourth, as with the “Flynn effect” the adoption gains are primarily on subtests that are not strongly associated with g, which means that the cause( s) of the Black-White gap probably are different from the cause( s) of adoption IQ gains....

***

4. L'ereditarietà è minore tra le classi meno abbienti. Sottointeso: dove i neri sono sovrarappresentati...

... heritability of intelligence is “markedly lower among American children raised in poverty” than among middle- or upper-class children... Turkheimer et al believe this undermines the hereditarian position because more Black than White children are raised in poverty in the US...

Tuttavia, altri studi smentiscono, e smentisce anche una rassegna esaustiva della letteratura scientifica disponibile sull’argomento...

... many other studies that have examined the heritability of cognitive ability have found no or little evidence that heritability is lower in poor children. In a meta-analysis of available studies, Tucker-Drob and Bates found that there was no overall effect of socioeconomic status on the heritability of intelligence...

Ma perché una gaffe del genere? Il solito effetto Wilson viene trascurato...

... Second, the heritability of intelligence increases with age (the “Wilson effect”); therefore, even if the heritability of intelligence in some groups was low in childhood (say, 10% or so), it is not clear that it would remain low into adulthood....

Sul tema gli studi più aggiornati sembrano chiarire una volta per tutte… 

... Third, researchers recently found evidence that the heritability of intelligence is the same in Blacks and Whites. In fact, this was found in the same sample of youths (but also including non-twin siblings and half-siblings) that Turkheimer used for his original analysis...

***

Fin qui abbiamo visto come gli argomenti critici contro l'ipotesi ereditaria non tengano, ora vediamo come quelli a supporto siano ben più solidi.

1. La correlazione tra dimensioni del cranio, intelligenza e razze è robusta e coerente...

... Best evidence suggests a small to moderate correlation between brain size and... intelligence; there is a pattern of brain size differences that matches IQ scores. East Asians have larger brains than Caucasians, who have larger brains than African Americans....

2. È possibile un rendiconto evoluzionistico coerente sui differenziali intellettivi tra le varie razze...

... Humans migrated out of Africa at least 50,000 years ago, settling in diverse... environments. Those furthest from the equator likely confronted novel cognitive challenges. Populations likely evolved higher... intelligence in response to these novel selection pressures. IQ scores correlate with distance from equator and with skin color (which... indicates ancestral distance from equator)...

3. Si registrano costantemente differenti IQ medi tra le differenti etnie...

... Many ethnic groups score differently on IQ tests. Sub Saharan Africans = 75; African... Americans = 85; Caucasians = 100; East Asians = 105; Ashkenazi Jews = 110....

4. C'è una robusta ereditarietà infra-razza  che non fatichiamo ad accettare. Tuttavia, questo fatto rende più probabile la presenza di una ereditarietà anche tra razze...

... Although not dispositive, the high within-race heritability of... IQ (50%-90%) makes it plausible that at least some of the Black-White IQ gap is genetically caused. This is especially true because researchers... have yet to find some unique environmental force that suppresses Black intelligence while not affecting White... intelligence...

5.  Ci sono parecchi tratti etnici chiaramente ereditati, molti motivi portano a pensare che anche le facoltà cognitive possano rientrare nel novero delle caratteristiche sottoposte a pressione evolutiva...

... IQ patterns among races also match a suite of life history traits (traits associated with growth, survival, and reproduction)...

6. Normali andamenti statistici (regressione alla media) rendono conto molto bene del dinamismo dell' IQ alla base di molte critiche superficiali all'ipotesi ereditaria...

... Population traits tend to move toward the population (racial) mean from extreme cases. Two high-IQ people tend to have children with lower IQs. The... opposite is also true. Two low-IQ people tend to have children with higher IQs. Multiple studies support hereditarian hypotheses.... When children are matched at 120 IQ, Black siblings average about 100, whereas White siblings average about 110. When children are matched at 70... IQ, Black siblings average about 78, whereas White siblings average about 85....

7. Gli studi sulle adozioni interazziali, ieri una freccia all'arco dei critici, oggi costituiscono un argomento a supporto dell'ipotesi Murray...

... Although adoption studies provide mixed... evidence, the single best study, the Minnesota Transracial adoption study (MTRAS) clearly conforms with a priori hereditarian... predictions. IQ W-W biological parents = 106; W-B biological parents 99; B-B biological parents = 89....

8. La mobilità nel gap intellettivo tra bianchi e neri si registra invariabilmente  in quei subtests dove l'intelligenza pura è meno coinvolta...

...  Race differences are greatest on... tests that are the most g loaded, which are also the most heritable. The Flynn effect manifests on tests that are less g loaded and less... heritable. The Black-White gap on heavily g loaded tests has shrunk less than on more weakly g loaded tests....

***

Come concludere?

Insistere su nobili bugie sembra deleterio. L'ipotesi dell'ereditarietà dell'intelligenza è solida, anche se non conclusiva. Diciamo che, vista la mole di prove a supporto, non crederla vera implica come minimo il rigetto immediato di altre ipotesi su cui di solito si ostenta sicumera, come per esempio quella relativa al riscaldamento  globale della terra causato dall'uomo.