I like the qualifier "machine" to distinguish AI from human intelligence or animal intelligence, because what goes on inside that black-box is different than what goes on in our brains, in ways we don't fully understand (and should try to understand better by getting to interpretability and explainability).
Totally agree about the need to consider the purpose--and *test* rigorously to ensure that purpose is achieved without spiraling negative second- and third-order effects. The initial Bing Chat rollout definitely seemed like a "drunk driving" scenario.
I see your point about interpretability and explainability. We should expect it as much as possible. But I think it is kind of unfair. After all, we don’t ask people to explain “how” they came up with a brilliant idea all the time. Sometimes they just do. Just like the human mind, many of the outputs of deep neural networks are brilliant but not very explainable.
I think we could try the triangulation approach as a proxy in some cases. Try to get the same conclusion from 3 different independent sources.
Humans learn in many different ways. We should not expect machines to be much different.
+1 to triangulation, that sounds like a good idea. Re: brilliant ideas, "I was thinking about X and then I thought about Y" seems like something a human could feasibly say and a machine also might be able to offer.
This is purely speculation on my part, but since we can use MRI in humans to see which regions of the brain light up as we think about different things, is there some analogous process that would allow developers to see which pathways to focus on for troubleshooting/tracing?
Wow love the MRI idea to learn about how machines think. I don’t know the details but I think people are using techniques in the algorithm to figure this out. FDA will require explainability for AI driven medical software. So I am thinking that this must be a hot area of research.
I am so glad to have found your writing Stephanie as dialog like this is just GREAT. It would be a very good undergrad essay question for students to explain the difference between AI and ML and how they complement each other. Automatic zero if using GPT. I believe the newly announced deep integration of AI with Google Application Suite and Gmail will be fun to watch.
Sorry this is a little long but one of my favorite topics.
I enjoy your writing about risk as it relates to AI. I especially enjoy the analogies to real life. I continue to believe most of the fanfare around AI at this point is sorely misplaced. I believe there are some a-grade challenges ahead though. The focus continues on consumer-facing "write a school essay" or "write a thank you letter". These happen to be the edge of the shallow water where the inquisitive are dipping their toes and reporting on twitter what happened.
While far from sexy, the closest think to ML in the consumer space is the now well recognized capacity of Gmail to block spam. I think when people migrate to GMail they are struck by its innate capabiities to manage an Inbox far beyond what a person can do on their own. It is a weird thing we all take for granted. The amount of compute, insight, and ML required to manage about 40% of the world's email traffic is a feat much more impressive than writing a middle school essay. This in a nutshell is why Apple finds a large subset of their users on their "superior platform" not being able to get along with all of these ML-charged products like Gmail and Gmaps for example. The converse is not true, there is no one beating a path to their door to please let me use Apple Mail
Yeah, "write a school essay" is just scraping the surface! Interesting comment about Gmail's filtering, I've viewed it as unthreatening and almost an unmitigated good, probably because it's so narrow. You raise an interesting point about where that boundary is between effectiveness and intelligence.
I think pure effectiveness and ease of use/convenience is what makes Apple (imho) the best OS right now and Gmail the best email and so on. I don't want to have to think about these things, and whichever product allows me to just use it and do the things I want to do with it, that's the best one in my view. AI will probably take this to new levels in ways we can't foresee (like, it would be amazing if a billion senior citizens could make their own websites by just saying what they want it to say and adjusting the design, and I think that is possible *today*).
The migration (and generalized agreement) to LINUX is a GREAT thing. iOS, Azure, AWS and Google are all there. Some work hard to avoid the OpenSource compliance, others share. It is definitely a great thing everyone has settled on the same basis. Because they have forked and exploit the EULA, it is inevitable they will all end up creting walled gardens with taller and taller walls. My favored observation is how hard some of them foster the illusion they invented something. It is part of their marketing story.
Hmm, I guess it depends how we define "idea". "Idea" can mean "let's go to the grocery store" or "I think this person really meant XYZ when they said ABC" or "it would be great to have an app that lets senior citizens make their own web pages by just saying what they want".
The first involves taking initiative on entirely familiar stuff, the second involves imputing meaning that wasn't overtly stated, and the third involves imagining new stuff that doesn't exist yet. To what degree can AI draw on past material/data for new inspiration? That's one of the big questions of the future.
Thank you for sharing your insights.
To me, the qualifier “artificial” in AI is inappropriate. AI is intelligence, just like any other intelligence.
Therefore the more relevant question in my mind is what we want to use this intelligence for?
If we deploy AI without thinking about our purpose, it is no different than letting a drunk person drive a car.
I like the qualifier "machine" to distinguish AI from human intelligence or animal intelligence, because what goes on inside that black-box is different than what goes on in our brains, in ways we don't fully understand (and should try to understand better by getting to interpretability and explainability).
Totally agree about the need to consider the purpose--and *test* rigorously to ensure that purpose is achieved without spiraling negative second- and third-order effects. The initial Bing Chat rollout definitely seemed like a "drunk driving" scenario.
I see your point about interpretability and explainability. We should expect it as much as possible. But I think it is kind of unfair. After all, we don’t ask people to explain “how” they came up with a brilliant idea all the time. Sometimes they just do. Just like the human mind, many of the outputs of deep neural networks are brilliant but not very explainable.
I think we could try the triangulation approach as a proxy in some cases. Try to get the same conclusion from 3 different independent sources.
Humans learn in many different ways. We should not expect machines to be much different.
+1 to triangulation, that sounds like a good idea. Re: brilliant ideas, "I was thinking about X and then I thought about Y" seems like something a human could feasibly say and a machine also might be able to offer.
This is purely speculation on my part, but since we can use MRI in humans to see which regions of the brain light up as we think about different things, is there some analogous process that would allow developers to see which pathways to focus on for troubleshooting/tracing?
Love your last statement, agree.
Wow love the MRI idea to learn about how machines think. I don’t know the details but I think people are using techniques in the algorithm to figure this out. FDA will require explainability for AI driven medical software. So I am thinking that this must be a hot area of research.
I am so glad to have found your writing Stephanie as dialog like this is just GREAT. It would be a very good undergrad essay question for students to explain the difference between AI and ML and how they complement each other. Automatic zero if using GPT. I believe the newly announced deep integration of AI with Google Application Suite and Gmail will be fun to watch.
Sorry this is a little long but one of my favorite topics.
I enjoy your writing about risk as it relates to AI. I especially enjoy the analogies to real life. I continue to believe most of the fanfare around AI at this point is sorely misplaced. I believe there are some a-grade challenges ahead though. The focus continues on consumer-facing "write a school essay" or "write a thank you letter". These happen to be the edge of the shallow water where the inquisitive are dipping their toes and reporting on twitter what happened.
While far from sexy, the closest think to ML in the consumer space is the now well recognized capacity of Gmail to block spam. I think when people migrate to GMail they are struck by its innate capabiities to manage an Inbox far beyond what a person can do on their own. It is a weird thing we all take for granted. The amount of compute, insight, and ML required to manage about 40% of the world's email traffic is a feat much more impressive than writing a middle school essay. This in a nutshell is why Apple finds a large subset of their users on their "superior platform" not being able to get along with all of these ML-charged products like Gmail and Gmaps for example. The converse is not true, there is no one beating a path to their door to please let me use Apple Mail
Yeah, "write a school essay" is just scraping the surface! Interesting comment about Gmail's filtering, I've viewed it as unthreatening and almost an unmitigated good, probably because it's so narrow. You raise an interesting point about where that boundary is between effectiveness and intelligence.
I think pure effectiveness and ease of use/convenience is what makes Apple (imho) the best OS right now and Gmail the best email and so on. I don't want to have to think about these things, and whichever product allows me to just use it and do the things I want to do with it, that's the best one in my view. AI will probably take this to new levels in ways we can't foresee (like, it would be amazing if a billion senior citizens could make their own websites by just saying what they want it to say and adjusting the design, and I think that is possible *today*).
The migration (and generalized agreement) to LINUX is a GREAT thing. iOS, Azure, AWS and Google are all there. Some work hard to avoid the OpenSource compliance, others share. It is definitely a great thing everyone has settled on the same basis. Because they have forked and exploit the EULA, it is inevitable they will all end up creting walled gardens with taller and taller walls. My favored observation is how hard some of them foster the illusion they invented something. It is part of their marketing story.
Does AI have ideas?, are human ideas only based on information?, or creativity can be a real source of new ideas.
Or creative ideas involved in art.
Wiki for “idea” tells:
The capacity to create and understand the meaning of ideas is considered to be an essential and defining feature of human beings.
Hmm, I guess it depends how we define "idea". "Idea" can mean "let's go to the grocery store" or "I think this person really meant XYZ when they said ABC" or "it would be great to have an app that lets senior citizens make their own web pages by just saying what they want".
The first involves taking initiative on entirely familiar stuff, the second involves imputing meaning that wasn't overtly stated, and the third involves imagining new stuff that doesn't exist yet. To what degree can AI draw on past material/data for new inspiration? That's one of the big questions of the future.