09/28/2023 - James Evans

[bhavyapan]

We ask that you:

• Post the questions and comments you have about the suggested readings.
• Upvote (“thumbs up”) at least 5 questions from other people. Upvote questions and recommendations you like or would like to critique.
• Respond with YOUR answer or follow-on engagement to one of those 5 questions.

[natashacarpcast]

Hello, good morning/afternoon/evening everyone! I found the idea of the "Complementary AI" (Sourati, J., Evans, J.A., 2023) very interesting and useful. I see the great utility and benefit in this, mostly in solving urgent issues such as finding cures to diseases, where time is crucial to save people. My question is: do you intend for this tool to be available for public usage? I see that many organizations could benefit from it to find solutions to their issues. How would you distribute it and make it user friendly to those not familiar with computational science?

[secorey]

Hi Dr. Evans,

Thanks for sharing your work. I have a question regarding the Sourati and Evans (2023) paper - if the type of complementary AI described in the article is intended to "avoid the crowds" so that it can produce novel predictions, how did you avoid issues of data availability?

[alejandro-sarria-morales]

The papers propose a model of AI as complement human intelligence. From a developmental viewpoint, how can AI be a complement to human intelligence as it developes during a lifetime and what do you think are the developmental implications growing alongside AI?

[lbitsiko]

Many thanks for presenting your intriguing work in our workshop.

I would like to focus on the "Complementary Artificial Intelligence" piece. Even though, the idea of "complementary" AI is quite catchy, I feel the focus should be more on the design of these systems (an aspect the article only briefly touches upon). Sure, we 'd all aspire to have AI algorithms crafted to collaborate with us. However, a critical question we should be asking is: who's designing these systems? I'm afraid that they're not creations of society for society, but rather systems at the hands of corporations. Speaking of which, I couldn't help but think of "complimentary AI" in the context of this article: https://hannesbajohr.de/en/2023/04/08/whoever-controls-language-models-controls-politics/

It's evident that potential dangers lurk within the concept of AI in general, but, hopefully, my viewpoint doesn't come across as neo-luddite. AI is undeniably here, fundamentally reshaping our societies, much like industrial machines did in the distant past, or nuclear technology in the more recent past (you see, I recently watched Oppenheimer). Reflecting on history, I can't help but wonder whether the world became better or worse after the industrial revolution, or the invention of nuclear technologies. The answer is clearly not a simple "yes" or "no". As a matter of fact, sometimes the answer depends on who we're asking. As far as AI (the technological revolution of our times) is concerned, I believe this serves as an important lesson.

[Marugannwg]

I feel like the short article is trying to propose a new type of more innovative/beneficial approach to AI -- Complementary AI" -- in contrast to those current population/commercially more successful AI products. I'd love to see how this new type of AI can empower people with different expertise and provide unique perspectives that our current society has overlooked. But, since I (and I believe, many of us) do NOT have a technical and geeky background, I need help clarifying those questions:

• What are the differences between a complementary AI and a currently popular AI model (like the GPT language model)? What makes those differences significant? (e.g. a task the former one can solve yet the latter one can NOT for some reason?)

• If I understand correctly, I assume that there are fundamental differences between creating a complementary AI model and a model that just "mimics human intelligence". Can you illustrate the process of how both types of AI are made? (What resources/technologies do we need to craft an AI like those?)

[ksheng-UChicago]

I understand that complementary AI can be helpful for individual humans to understand a piece of specific knowledge or to avoid possible bias as a human coach, collaborator, and coordinator. However, I feel that relying on any particular technology might lead to a decreasing enthusiasm for putting in one's own efforts. For example, if AI can detect and make an unbiased decision during a hearing, the judge (human being) might tend to count on the AI tool, instead of taking full responsibility. I am wondering where the boundary between relying on AI and utilizing AI as a complementary tool draws. What suggestions will you give to the audience on utilizing complementary AI tools?

[zihua-uc]

Hi Dr. Evans,

Do you think the development of complementary AI is feasible in the near future (say 5 years)? Is the current state of AI able to support this research program, and if not, what are the bottlenecks?

[ksheng-UChicago]

Hello, good morning/afternoon/evening everyone! I found the idea of the "Complementary AI" (Sourati, J., Evans, J.A., 2023) very interesting and useful. I see the great utility and benefit in this, mostly in solving urgent issues such as finding cures to diseases, where time is crucial to save people. My question is: do you intend for this tool to be available for public usage? I see that many organizations could benefit from it to find solutions to their issues. How would you distribute it and make it user friendly to those not familiar with computational science?

How can the technology be applied to the public seems like an important factor to foresee whether it is promising in the long run.

[anzhichen1999]

The "wisdom of crowds" principle emphasizes the importance of diverse and independent inputs. Given the progress in AI ensemble techniques and their consistent performance in data competitions, how can we guarantee that the creation of supplementary AI incorporates a wide range of human viewpoints and avoids biases that might unintentionally perpetuate societal inequalities?

[anzhichen1999]

The "wisdom of crowds" principle emphasizes the importance of diverse and independent inputs. Given the progress in AI ensemble techniques and their consistent performance in data competitions, how can we guarantee that the creation of supplementary AI incorporates a wide range of human viewpoints and avoids biases that might unintentionally perpetuate societal inequalities?

So from my perspective, computational social scientists need to ensure a diverse data collection and be able to detect potential bias, and follow ethical guidelines and regulations. I think ChatGPT has followed some types of ethical guidelines, but I am not sure whether it does a good job in diversity. But a diverse data collection might not be ChatGPT's strength.

[Brian-W00]

Since the AI was trained on pre-existing data, to what extent can it truly offer innovative or novel ideas?

[66Alexa]

Hi everyone,
I found it interesting reading this article Complementary Artificial Intelligence, but got some questions: What might be some potential unintended consequences of this shift toward complementary AI? As AI is designed to think differently and expand human capacity, is there a risk of over-reliance or unforeseen negative impacts?

[maxzhuyt]

The short essay proposes a redefinition of AI as complementary to maladaptive human intelligence. I understand that the longer article attempts to examine one such instance of (potentially maladaptive) human intelligence—the psychological tendency that limits how scientists foresee future possibilities of discovery. An AI model of discoverability assists us in finding areas of potential discovery that are unlikely to be explored by human intelligence alone. It seems to me that the paper's major result lies in predicting the patterns of unassisted human endeavor, but I am curious how one could then identify promising hypotheses from the "less discoverable" territory. This step seems to warrant tremendous domain knowledge.

[naivetoad]

For the world nowadays, augmented intelligence looks like a more pragmatic approach than the traditional notion of artificial intelligence, but it seems not to be intended for any creative or innovative purposes which have been deemed more and more important to the future. Therefore, I have two questions for Professor Evans. Which path, the practical or creative one, do you think the artificial intelligence industry should prioritize now? Are there any real-life cases where these two forms of artifical intelligence are working together?

[ecg1331]

Hi Dr. Evans,

These were both really interesting papers to read!

I am especially intrigued by the idea of these complementary or “alien” discoveries that were discussed in “Accelerating science with human-aware Artificial Intelligence” and had a question regarding the current applicability of these discoveries.

Throughout the paper, a common theme seemed to be the discussion around the way the current scientific/research community conducts research and how this creates a “human crowd” of very similar (and already established) research, leaving little room for true “discoveries.”

It was also mentioned that these complementary discoveries were unlikely to be discovered by researchers, or if they were to be discovered by researchers it would be in the distant future. Do you think this would pose a problem, seeing as though the community tends to gravitate more towards researching branches of established theories, or do you think that the community would be ready to look at these alien/complementary inferences now?