Generative AI: A Brief History
If AI is defined as a non-human way of mimicking human intelligence, then the 1967 Caltech handheld calculator must be considered the ultimate generative AI. For example, entering 2*2 will return the human readable answer “= 4”. This is called rule-based AI. It’s not rule-based machine learning because the calculator isn’t learning, it’s following rules from an expert’s intellect. In fact, you don’t want to program your pocket calculator with learning algorithms or user his feedback. This is because it can lead to erroneous results. Each generating algorithm has its own purpose and should adhere to the standards of the end user, not the programmer.
Rule-based machine learning, as the term suggests, ingests large amounts of data to allow algorithms to better learn or infer large numbers of rules that are not easily documented. For example, you can replace each word in the sentence “George Washington was the first President of the United States” with a grace tag (e.g. proper noun = NP0, VVB = verb, ADJ = adjective, etc.). = “NP0 NPO VBD ATO AJ0 NN1 PRF ATO AJ0 NN2.”
This so-called natural language processing (NLP) can infer grammatical patterns that humans are likely to understand. The more frequent the pattern, the more likely it will be understood by the end user. With NLP, you can split sentences by verbs (VBD = “was”), extract knowledge of “right hand” and “left hand”, and distinguish between subject and object English grammar, as humans do It can also reflect rules. IBM Watson, introduced in 2011, is an example of an AI system using NLP.
Scraping billions of sentences from digitized sources, this method provides a simple, irrational, knowledge-oriented chatbot where one sentence generates two chunks of knowledge.
Q: Who was America’s first president? (search to the right)
A: George Washington (return left side)
and
Q: Who is George Washington? (search left)
A: First President of the United States (back to right)
Note that in both cases the text exists in nature and has been plagiarized from some source, but only fragments are returned. A simple knowledge bot doesn’t need a neural network. Good old NLP is good enough.
Now imagine replacing the text with “synonyms” that can be used for speculation and scientific reasoning.
Aspirin suppresses headaches in chimpanzees.
become
<分子> teeth <ゲノム配列または種> of <症状> impede
A single sentence can represent years of research and millions of dollars in crystallization (eg, “The cure for malaria is XYZ”). Because knowledge is in the public domain, computers can waste billions of dollars of investment reflected in petabytes of text. You can now apply inference and predict what scientists might infer. cool.
Problems of modern chatbots
Neural networks are useful when pre-generated transformers or the now-famous GPT are one class and the exact rules are not well documented to help the programmer achieve the desired result. It is also useful when the rules and strategies for implementing them are not well understood or too costly to program compared to cognitive approaches. Deep generative models and underlying algorithms confidently make what look like educated guesses. However, even when the rules are fully understood, they may not follow known rules. that’s a problem.
If the user knows the rules and the AI doesn’t follow them, the output could be interpreted as ‘disinformation’, ‘deception’, ‘lie’, ‘propaganda’ or even worse ‘social manipulation’. I have.If you are a programmer Add rule-based logic above the “black box” Then the concern over manipulation is amplified. Bad advice or direction given without attribution or warning can kill someone or start a riot.
Compare two didactic or educational limericks generated from different methodologies. But use the same prompt “love” (e.g. write a didactic limerick about love). One uses black box deep learning and the other uses rules. The “didactic” imperative is to emphasize that the poem should teach the reader the meaning or definition of love. Limericks are his five-line poems, usually humorous and often rude or vulgar. anapestrimeterThere is a strict rhyming scheme of AABBA, with lines 1, 2 and 5 rhyming, but lines 3 and 4 being shorter and sharing different rhymes.
Poem #1: ChatGPT
Once upon a time there was such a lovely love
shining like the bright sunshine
brought joy to every day
and made everything shine
It was love intended, not a plan
Poem #2: TotoGEO
There was a pigeon girl there.
And she wanted words of love.
Like a meat injection?
No, it’s affection!
The literate gal from The Dove.
Which poem do you like? This may be a matter of taste. Which poem is the better limerick? The answer is obvious: Poetry #2. Why? Because it’s limericks and it defines love as affection. Younger readers may speculate that verse #1 is limericks and be forever misunderstood, and his second verse, which follows known rules, is based on the idea that the rules themselves are known, as cognitive AI approaches usually do. It can never be “wrong” as long as it is programmed with
The GPT approach has a wide range of capabilities, but can be expensive to implement and subject to guesswork. recent media reportfor example, concluded that “According to the poet, ChatGPT is pretty bad at poetry.”
In contrast, rule-oriented approaches are built on a large general linguistic graph applied to a narrow formulaic writing style. Limerick’s algorithm is not the same as the one described by Sonnet, even though it shares the same backend data graph and other code. The method in poem #2 in the example above is much less expensive to implement than GPT, and like calculators, the expected result is never wrong.visit totopoetry.com Type a word and get some examples.
The algorithmic rules and backend methodology collectively referred to as “TotoPoetry” (see above) have achieved several milestones. The first complete dictionary consisting of verse-written definitions for about 20 genres of poetry. And the largest collection of poems written by the same source.
Graph-generated text can be viewed as a hybrid generation methodology, as the data fed into TotoPoetry is itself partially generated from a neural network. TotoPoetry creates millions of poems, so a second algorithm acts as an editor and selects the best poems to post.
real fight
What’s really happening here is the expectation that the output generated by AI is correct, reliable, unbiased, or at least not completely wrong. The autopilot shouldn’t be trying to fly the plane through the mountains. Similarly, generative AI platforms should not be misleading. At the very least, you should demonstrate that you may be ignorant on topics on which you have no training. He asked ChatGPT to write a Wikipedia article on Tinwarrole, and got the following (note the bot’s assertive tone, not due to rules imposed on the language model): It seems).
Chat GPT: Tinwarrole is a fictional substance and has no real-world basis. It does not exist in reality, so it has no known properties or uses. It is not a recognized term in science or medicine and there is no information about it. As a result, I can’t write Wikipedia-style articles about Tinwarrole. Because there is no reliable and verifiable information to include in such articles.
Our lab leverages strict rules to provide factual (non-fictitious) answers derived from ASCII files. indexing Google About 10 years ago:
TotoGEO: Tinwarrole was the name of an African slave aboard the NS de Montserrat. Tinwaror was enslaved in 1818, he was registered as a 44-year-old man. NS de Montserrat was a Spanish schooner captained by Idriso de Pascual. This ship had a mortality rate of 1% for her. If there are any descendants of Tinwaror, he was freed from slavery in the United States following the Emancipation Proclamation in 1863.
Generative AI is just getting started. Things will improve over time. Asking to suspend research in this area seems naive in its lack of specificity. Companies selling this technology should be careful not to release too early or too broadly. Minimum quality standards should be encouraged. Many of the platforms developed in the INSEAD TotoGEO AI Lab are in the hands of hand-picked curators who verify authenticity and provenance. Our work includes the use of generative AI to exacting standards in agriculture and research.
We can expect to see more hybrid AI methodologies that blend deep learning with cognitive, symbolic, control, and/or various rule-based approaches in the near future. In either case, engineers can and do introduce strategic bias into the systems created, requiring transparency, provenance, and accountability. If a plane crashes due to a faulty autopilot system, the company that makes or knowingly uses such a faulty system could be held liable.
Future articles in this series will discuss how these methodologies can take business consulting, education, journalism, and search engines to a whole new level.
