January 10, 2024
A Gentle Introduction to Artificial Intelligence (AI)
and Large Language Models (LLMs)
Ira Fuchs
President of BITNET, former computer science professor at Princeton University
A Gentle Introduction to Artificial Intelligence (AI)
and Large Language Models (LLMs)
Ira Fuchs
President of BITNET, former computer science professor at Princeton University
Ira Fuchs and Nancy Becker, Introducer
Minutes of the 14th Meeting of the 82nd Year
John Cotton presided, and Julia Coale led the invocation. Ruth Miller read the minutes of the previous meeting. There were 100 members and guests in attendance. The following members introduced their guests:
Nancy Becker introduced Karen Fuchs, wife of the speaker
Ferris Olin introduced Donald Schnur, a candidate for membership
Greg Dobbs introduced Larry Riddle
Steve Kahn introduced Florence Kahn
Lynn Johnston introduced Weezie Stebbins
Marcia Snowden introduced Christina Rountree
Professor Fuchs is an internationally known authority on technology innovation in higher education. He is a co-founder of BITNET and was inducted into the Internet Hall of Fame in 2017. A graduate of Columbia University with a BS and MS in Computer Science and Electrical Engineering, he served as the first Executive Director of the Computer Center of CUNY. He oversaw the connection of millions of users, from more than 1400 higher education, government, and other institutions around the world. He continued his career at Princeton University from 1985-2000, followed by additional academic and prestigious assignments in his field.
His presentation included seven segments:
AI in everyday life;
The Evolution of AI;
Introduction to Large Language Models (LLM’s);
Practical Uses of AI and LLM’s
Limitations of AI and LLM’s
Future of AI and LLM’s
A Demonstration of their use
“If you have a smartphone, you’re a user of AI.” Besides phones, AI is used in appliances such as refrigerators and robotic vacuums; it powers “recommender” systems such as when Amazon tells you what additional products “people like you” have purchased; as part of autonomous vehicle operations; it’s used in banking to calculate loan worthiness; in court for establishing sentencing guidelines; for surveillance such as facial recognition as used in China; and in healthcare. The speaker suggested to any radiologists or pathologists in the audience, that now might be an excellent time to retire.
The quest for AI began in the 1950’s with Allen Turing, who created “The Imitation Game,” where a human and a computer would talk to each other behind a curtain. If a human observer could not tell which participant was the human and which was the computer, then AI would have been achieved. However, this failed. And failure initiates long periods of “AI winter,” whether interest in and funding of, dries up. Conversely, we are now in a period of aggressive AI development.
In 2022, Sam Altman introduced Chat-GPT, an open-source version of Artificial Intelligence. In two months, he had 100 million users, a level that took Facebook four years to achieve.
AI is becoming multimedia, able to produce video, realistic human voice audio, as well as text. And AI is an excellent one-on-one tutor, as developed by Sal Kahn Academy which can tutor you on any subject you wish.
Among other important practical uses of this technology are writing computer code, writing fictional stories, often in the style of famous writers, poems, and scripts. Hollywood scriptwriters, the radiologists and pathologists are crying over there.
And idea generation is another function now possible.
So, are plumbers really safe from being replaced by AI? Not ALL of them, but many of them. If an AI powered robot can perform delicate surgery, do you think they can’t connect a pipe, link electrical wiring, or do many of the tasks that we don’t yet realize can be done by AI-powered robots. There may always be plumbers and programmers, but far fewer of them due to the speed and quality with which AI can perform these tasks.
There are limitations to AI. First, it deals in “facts,” which are just repetitive patterns. But these “facts” may or may not be true, just repeated a lot. In fact, AI is capable of “hallucinations,” where it basically “makes stuff up.” A recent case involved an attorney who asked ChatGPT to write a brief. The AI made up all the case citations, which the blundering lawyer failed to verify, revealing one of the problems the speaker cited, namely over reliance by unthinking, lazy humans.
Other limitations are that AI has limited “reasoning” ability, lacks what we call common sense, is subject to bias and disinformation.
The speaker recounted the “Paperclip Parable” to illustrate how things might go wrong. An autonomous AI, when tasked with improving the economics of making paper clips, might decide the best course of action would be to procure all of the world’s necessary raw materials, leaving behind a world with nothing but paperclips. If tasked with “protecting the USA from Russian missile attack,” AI might conclude that the way to do this is to bomb Russia first. And do it!
The speaker says he is not a “doomer,” but 5% in a recent poll of experts felt uncontrolled AI could be an “extinction event” for the human race.
Another issue is the energy necessary to run AI. An AI search requires ten times the energy of a Google search. If all the Google searches were done by AI, the energy required would equal all the energy used by the Republic of Ireland.
The demonstration portion of the lecture suffered from an intermittent lack of internet connection. The speaker quoted Steve Jobs, who once told Professor Fuchs that the probability of a live demonstration failing is proportional to the number of people in the room. And we had 100 members and guests!
When the connection was good, the audience saw how quickly ChatGPT could respond to a well-formed question. And you can ask for answers with conditions, like “answer at the level of a 10-year-old,” or “answer in the format of a Shakesperean sonnet.”
OK, I couldn’t resist asking ChatGPT to write these minutes. How did it turn out?
ChatGPT started by saying it had no information in its training data, but wrote a general summary based on common topics covered in such presentations. This was way too general and contained no specifics. I tried again, by providing the date of the talk.
GPT then tried to upgrade me to a higher level of service for $20/month.
Finally, I uploaded my actual written minutes at 850-word length, and ChatGPT wrote a quite acceptable 600-word version of my minutes. Bottom Line: The Old Guard will need “real” minute takers for the foreseeable future.
See below for ChatGPT's 600-word summary.
Respectfully submitted,
Owen Leach
Nancy Becker introduced Karen Fuchs, wife of the speaker
Ferris Olin introduced Donald Schnur, a candidate for membership
Greg Dobbs introduced Larry Riddle
Steve Kahn introduced Florence Kahn
Lynn Johnston introduced Weezie Stebbins
Marcia Snowden introduced Christina Rountree
Professor Fuchs is an internationally known authority on technology innovation in higher education. He is a co-founder of BITNET and was inducted into the Internet Hall of Fame in 2017. A graduate of Columbia University with a BS and MS in Computer Science and Electrical Engineering, he served as the first Executive Director of the Computer Center of CUNY. He oversaw the connection of millions of users, from more than 1400 higher education, government, and other institutions around the world. He continued his career at Princeton University from 1985-2000, followed by additional academic and prestigious assignments in his field.
His presentation included seven segments:
AI in everyday life;
The Evolution of AI;
Introduction to Large Language Models (LLM’s);
Practical Uses of AI and LLM’s
Limitations of AI and LLM’s
Future of AI and LLM’s
A Demonstration of their use
“If you have a smartphone, you’re a user of AI.” Besides phones, AI is used in appliances such as refrigerators and robotic vacuums; it powers “recommender” systems such as when Amazon tells you what additional products “people like you” have purchased; as part of autonomous vehicle operations; it’s used in banking to calculate loan worthiness; in court for establishing sentencing guidelines; for surveillance such as facial recognition as used in China; and in healthcare. The speaker suggested to any radiologists or pathologists in the audience, that now might be an excellent time to retire.
The quest for AI began in the 1950’s with Allen Turing, who created “The Imitation Game,” where a human and a computer would talk to each other behind a curtain. If a human observer could not tell which participant was the human and which was the computer, then AI would have been achieved. However, this failed. And failure initiates long periods of “AI winter,” whether interest in and funding of, dries up. Conversely, we are now in a period of aggressive AI development.
In 2022, Sam Altman introduced Chat-GPT, an open-source version of Artificial Intelligence. In two months, he had 100 million users, a level that took Facebook four years to achieve.
AI is becoming multimedia, able to produce video, realistic human voice audio, as well as text. And AI is an excellent one-on-one tutor, as developed by Sal Kahn Academy which can tutor you on any subject you wish.
Among other important practical uses of this technology are writing computer code, writing fictional stories, often in the style of famous writers, poems, and scripts. Hollywood scriptwriters, the radiologists and pathologists are crying over there.
And idea generation is another function now possible.
So, are plumbers really safe from being replaced by AI? Not ALL of them, but many of them. If an AI powered robot can perform delicate surgery, do you think they can’t connect a pipe, link electrical wiring, or do many of the tasks that we don’t yet realize can be done by AI-powered robots. There may always be plumbers and programmers, but far fewer of them due to the speed and quality with which AI can perform these tasks.
There are limitations to AI. First, it deals in “facts,” which are just repetitive patterns. But these “facts” may or may not be true, just repeated a lot. In fact, AI is capable of “hallucinations,” where it basically “makes stuff up.” A recent case involved an attorney who asked ChatGPT to write a brief. The AI made up all the case citations, which the blundering lawyer failed to verify, revealing one of the problems the speaker cited, namely over reliance by unthinking, lazy humans.
Other limitations are that AI has limited “reasoning” ability, lacks what we call common sense, is subject to bias and disinformation.
The speaker recounted the “Paperclip Parable” to illustrate how things might go wrong. An autonomous AI, when tasked with improving the economics of making paper clips, might decide the best course of action would be to procure all of the world’s necessary raw materials, leaving behind a world with nothing but paperclips. If tasked with “protecting the USA from Russian missile attack,” AI might conclude that the way to do this is to bomb Russia first. And do it!
The speaker says he is not a “doomer,” but 5% in a recent poll of experts felt uncontrolled AI could be an “extinction event” for the human race.
Another issue is the energy necessary to run AI. An AI search requires ten times the energy of a Google search. If all the Google searches were done by AI, the energy required would equal all the energy used by the Republic of Ireland.
The demonstration portion of the lecture suffered from an intermittent lack of internet connection. The speaker quoted Steve Jobs, who once told Professor Fuchs that the probability of a live demonstration failing is proportional to the number of people in the room. And we had 100 members and guests!
When the connection was good, the audience saw how quickly ChatGPT could respond to a well-formed question. And you can ask for answers with conditions, like “answer at the level of a 10-year-old,” or “answer in the format of a Shakesperean sonnet.”
OK, I couldn’t resist asking ChatGPT to write these minutes. How did it turn out?
ChatGPT started by saying it had no information in its training data, but wrote a general summary based on common topics covered in such presentations. This was way too general and contained no specifics. I tried again, by providing the date of the talk.
GPT then tried to upgrade me to a higher level of service for $20/month.
Finally, I uploaded my actual written minutes at 850-word length, and ChatGPT wrote a quite acceptable 600-word version of my minutes. Bottom Line: The Old Guard will need “real” minute takers for the foreseeable future.
See below for ChatGPT's 600-word summary.
Respectfully submitted,
Owen Leach
|
ChatGPT Summary:
Ira Fuchs' presentation on "A Gentle Introduction to Artificial Intelligence (AI) and Large Language Models (LLMs)" delivered on January 10, 2024, to the Old Guard of Princeton covered a wide range of topics, providing insights into the current state, practical applications, limitations, and the future of AI and LLMs The presentation was organized into seven segments, beginning with an exploration of AI in everyday life. Fuchs highlighted that anyone with a smartphone is a user of AI, emphasizing its presence in various aspects of daily routines, from autonomous vehicles to recommender systems and healthcare applications. He underlined the ubiquity of AI, even in household appliances like refrigerators and robotic vacuums. The Evolution of AI was another key segment, tracing the quest for AI back to the 1950s with Alan Turing's "The Imitation Game." Fuchs discussed periods of "AI winter," where interest and funding waned, but contrasted this with the current era marked by aggressive AI development. He mentioned Sam Altman's introduction of Chat-GPT in 2022, an open-source AI that rapidly gained 100 million users in just two months. The third segment delved into an Introduction to Large Language Models (LLMs). Fuchs explained how LLMs, like Chat-GPT, have evolved to become multimedia, capable of producing video, realistic human voice audio, and text. He emphasized the effectiveness of AI as a one-on-one tutor, citing the example of the Khan Academy. Practical Uses of AI and LLMs were discussed extensively, covering diverse applications such as writing computer code, generating fictional stories, poems, and scripts. Fuchs highlighted the impact on professions like Hollywood scriptwriters and even suggested that certain tasks traditionally performed by professionals, including plumbers, could be automated by AI-powered robots. However, Fuchs didn't shy away from addressing the Limitations of AI and LLMs. He pointed out that AI deals in "facts," which are essentially repetitive patterns and may not always be accurate. The potential for "hallucinations" was highlighted, where AI could generate information that is not true. Other limitations included a lack of reasoning ability, common sense, susceptibility to bias, and misinformation. The speaker illustrated the potential risks with the "Paperclip Parable," emphasizing how an autonomous AI, tasked with optimizing paperclip production, might inadvertently lead to catastrophic consequences by exhausting global resources. Fuchs acknowledged concerns about uncontrolled AI, with a poll indicating 5% of experts viewing it as a potential "extinction event" for humanity. Energy consumption emerged as a critical issue in AI, with Fuchs noting the significant energy requirements for AI searches compared to traditional searches. He provided a striking comparison, stating that if all Google searches were conducted by AI, the energy consumption would equal that of the entire Republic of Ireland. The presentation concluded with a demonstration of ChatGPT, showcasing its ability to respond quickly to well-formed questions. Despite intermittent internet connection issues, the audience witnessed the model's versatility in generating responses based on different conditions, such as responding at the level of a 10-year-old or in the format of a Shakespearean sonnet. Overall, Ira Fuchs' presentation offered the Old Guard of Princeton a comprehensive overview of AI and LLMs, blending historical context, practical applications, limitations, and thought-provoking considerations about the future impact of these technologies. |