Amazon AGI Team Say Their AI is Showing "Emergent Properties"

Noor Al-Sibai

Futurism.com

Originally posted 15 Feb 24

A new Amazon AI model, according to the researchers who built it, is exhibiting language abilities that it wasn't trained on.

In a not-yet-peer-reviewed academic paper, the team at Amazon AGI — which stands for "artificial general intelligence," or human-level AI — say their large language model (LLM) is exhibiting "state-of-the-art naturalness" at conversational text. Per the examples shared in the paper, the model does seem sophisticated.

As the paper indicates, the model was able to come up with all sorts of sentences that, according to criteria crafted with the help of an "expert linguist," showed it was making the types of language leaps that are natural in human language learners but have been difficult to obtain in AI.

Named "Big Adaptive Streamable TTS with Emergent abilities" or BASE TTS, the initial model was trained on 100,000 hours of "public domain speech data," 90 percent in English, to teach it how Americans talk. To test out how large models would need to be to show "emergent abilities," or abilities they were not trained on, the Amazon AGI team trained two smaller models, one on 1,000 hours of speech data and another on 10,000, to see which of the three — if any — exhibited the type of language naturalness they were looking for.

The info is here.

My overall conclusion from the paper linked in the article:

BASE TTS (Text To Speech) represents a significant leap forward in TTS technology, offering superior naturalness, efficiency, and potential for real-world applications like voicing LLM outputs. While limitations exist, the research paves the way for future advancements in multilingual, data-efficient, and context-aware TTS models.

Rational learners and parochial norms

Partington, S. Nichols, S., & Kushnir, T.

Cognition

Volume 233, April 2023, 105366

Abstract

Parochial norms are narrow in social scope, meaning they apply to certain groups but not to others. Accounts of norm acquisition typically invoke tribal biases: from an early age, people assume a group's behavioral regularities are prescribed and bounded by mere group membership. However, another possibility is rational learning: given the available evidence, people infer the social scope of norms in statistically appropriate ways. With this paper, we introduce a rational learning account of parochial norm acquisition and test a unique prediction that it makes. In one study with adults (N = 480) and one study with children ages 5- to 8-years-old (N = 120), participants viewed violations of a novel rule sampled from one of two unfamiliar social groups. We found that adults judgments of social scope – whether the rule applied only to the sampled group (parochial scope), or other groups (inclusive scope) – were appropriately sensitive to the relevant features of their statistical evidence (Study 1). In children (Study 2) we found an age difference: 7- to 8-year-olds used statistical evidence to infer that norms were parochial or inclusive, whereas 5- to 6-year olds were overall inclusive regardless of statistical evidence. A Bayesian analysis shows a possible inclusivity bias: adults and children inferred inclusive rules more frequently than predicted by a naïve Bayesian model with unbiased priors. This work highlights that tribalist biases in social cognition are not necessary to explain the acquisition of parochial norms.

From the General discussion

The widespread prevalence of parochial norms across history and cultures have led some to suggest parochialism is itself a human universal (Clark et al., 2019; Greene, 2013) in part owing to evolved, group-based biases in social norm acquisition (Chalik & Rhodes, 2020; Chudek & Henrich, 2011; Roberts et al., 2017). In this paper, we investigated whether a rational learning process can also explain this phenomenon. In Study 1, we found that adults can acquire distinctions of social scope in a statistically appropriate manner, and this finding was robust across two forms of measurement (rule judgments and open response). In Study 2, older children displayed the adult-like statistical sensitivity in their rule judgments, and even younger children did so in their open responses. Computational analyses suggests that rule judgments were inclusively biased: compared to an unbiased Bayesian learner, children tended to assume that novel rules apply to everyone in a candidate population. Adults also displayed an inclusive bias, albeit to a lesser extent than children.

Broadly, these findings suggest that rational learning processes can indeed explain the acquisition of parochial norms and highlight an important sense in which children's norm learning can be biased in the opposite direction of tribalism. At the least, the finding that children and adults are inclusively biased serves as an existence proof that deep-rooted tribal biases in social learning are not necessary to explain the acquisition of parochial norms. Rather, if children and adults are rational learners, they can acquire a parochial norm when presented with evidence that is consistent with parochialism. However, tribalism can still play a role in norm acquisition, for example, by influencing the sort of evidence that adults seek out, or the evidence to which children are exposed.

The research is here.

Parents’ Political Ideology Predicts How Their Children Punish

Leshin, R. A., Yudkin, D. A., Van Bavel, J. J., 

Kunkel, L., & Rhodes, M. (2022). 

Psychological Science

https://doi.org/10.1177/09567976221117154

Abstract

From an early age, children are willing to pay a personal cost to punish others for violations that do not affect them directly. Various motivations underlie such “costly punishment”: People may punish to enforce cooperative norms (amplifying punishment of in-groups) or to express anger at perpetrators (amplifying punishment of out-groups). Thus, group-related values and attitudes (e.g., how much one values fairness or feels out-group hostility) likely shape the development of group-related punishment. The present experiments (N = 269, ages 3−8 from across the United States) tested whether children’s punishment varies according to their parents’ political ideology—a possible proxy for the value systems transmitted to children intergenerationally. As hypothesized, parents’ self-reported political ideology predicted variation in the punishment behavior of their children. Specifically, parental conservatism was associated with children’s punishment of out-group members, and parental liberalism was associated with children’s punishment of in-group members. These findings demonstrate how differences in group-related ideologies shape punishment across generations.

Conclusion

The present findings suggest that political ideology shapes punishment across development. Counter to previous findings among adults (King & Maruga, 2009), parental conservatism (vs. liberalism) was not related to increased punishment overall. And counter to previous developmental research on belief transmission (Gelman et al., 2004), our patterns did not strengthen with age. Rather, we found that across development, the link between ideology and punishment hinged on group membership. Parental conservatism was associated with children’s punishment of out-groups, whereas parental liberalism was associated with children’s punishment of in-groups. Our findings add rich insights to our understanding of how costly punishment functions in group contexts and provide new evidence of the powerful transmission of belief systems across generations.

A pdf is here.

Socially evaluative contexts facilitate mentalizing

Woo, B. M., Tan, E., Yuen, F. L, & Hamlin, J. K.

Trends in Cognitive Sciences, Month 2022, 

Vol. xx, No. xx

Abstract

Our ability to understand others’ minds stands at the foundation of human learning, communication, cooperation, and social life more broadly. Although humans’ ability to mentalize has been well-studied throughout the cognitive sciences, little attention has been paid to whether and how mentalizing differs across contexts. Classic developmental studies have examined mentalizing within minimally social contexts, in which a single agent seeks a neutral inanimate object. Such object-directed acts may be common, but they are typically consequential only to the object-seeking agent themselves. Here, we review a host of indirect evidence suggesting that contexts providing the opportunity to evaluate prospective social partners may facilitate mentalizing across development. Our article calls on cognitive scientists to study mentalizing in contexts where it counts.

Highlights

Cognitive scientists have long studied the origins of our ability to mentalize. Remarkably little is known, however, about whether there are particular contexts where humans are more likely to mentalize.

We propose that mentalizing is facilitated in contexts where others’ actions shed light on their status as a good or bad social partner. Mentalizing within socially evaluative contexts supports effective partner choice.

Our proposal is based on three lines of evidence. First, infants leverage their understanding of others’ mental states to evaluate others’ social actions. Second, infants, children, and adults demonstrate enhanced mentalizing within socially evaluative contexts. Third, infants, children, and adults are especially likely to mentalize when agents cause negative outcomes.  Direct tests of this proposal will contribute to a more comprehensive understanding of human mentalizing.

Concluding remarks

Mental state reasoning is not only used for social evaluation, but may be facilitated, and even overactivated, when humans engage in social evaluation. Human infants begin mentalizing in socially evaluative contexts as soon as they do so in nonevaluative contexts, if not earlier, and mental state representations across human development may be stronger in socially evaluative contexts, particularly when there are negative outcomes. This opinion article supports the possibility that mentalizing is privileged within socially evaluative contexts, perhaps due to its key role in facilitating the selection of appropriate cooperative partners. Effective partner choice may provide a strong foundation upon which humans’ intensely interdependent and cooperative nature can flourish.

The work cited herein is highly suggestive, and more work is clearly needed to further explore this possibility (see Outstanding questions). We have mostly reviewed and compared data across experiments that have studied mentalizing in either socially evaluative or nonevaluative contexts, pulling from a wide range of ages and methods; to our knowledge, no research has directly compared both socially evaluative and nonevaluative contexts within the same experiment.  Experiments using stringent minimal contrast designs would provide stronger tests of our central claims. In addition to such experiments, in the same way that meta-analyses have explored other predictors of mentalizing, we call on future researchers to conduct meta-analyses of findings that come from socially evaluative and nonevaluative contexts. We look forward to such research, which together will move us towards a more comprehensive understanding of humans’ early mentalizing.

A pdf is here.