1 of 30

“Some philosophers claim�that a machine that acts intelligently�would not be actually thinking,�but would be only a simulation of thinking. But most AI researchers are not concerned with the distinction.”

11 of 30

The context

“Machine learning�is the field of study�that gives computers�the ability to learn�without being�explicitly programmed”�A. Samuel, 1959

artificial intelligence

machine learning

artificial neural networks

ANNs for natural�language processing

generative ANNs

conversational ANNs

12 of 30

So… what is really happening then?

15 January, BBC, https://www.youtube.com/watch?app=desktop&v=BWCCPy7Rg-s

13 of 30

The example of a conversation

Chatting with an AI… (not edited)�A conversation about problem solving

The entity with which we had this conversation:

writes a good English, and other languages
produces original texts
fulfills complex requests
adapts its arguments to the context
proposes creative contents
analyzes and summarizes long texts
shows sophisticated linguistic skills
…

The novelty is not in what it knows,�but in how it (knows and) interacts

Is it “really” intelligent? Does it “really” think? Is it “really” sentient?

Given the acknowledgment that it is not as we are,�perhaps these questions are not so important…?

14 of 30

An interpretation…

… to avoid what could be a pseudo-problem:

E. Dijkstra, 1984 http://www.cs.utexas.edu/users/EWD/ewd08xx/EWD898.PDF

15 of 30

Another interpretation…

(be either always or never reductionist!)

How can it think?�It is only�a mathematical process!

How can it think?�It is only�an electrochemical�process!

https://arxiv.org/abs/1706.03762

16 of 30

Some information about ChatGPT (& its siblings):�the high-level architecture

chatbot

Neural Network�(Large�Language Model)

users

Application Programming�Interface

… … …

custom app

17 of 30

Some information about ChatGPT (& its siblings):�interacting directly via the API

Neural Network�(Large�Language Model)

Application Programming�Interface

curl http://localhost:4891/v1/chat/completions� -H "Content-Type: application/json"� -d '{

"model": "Llama-2-7B Chat",

"max_tokens": 4096,

"messages": [{"role": "user", "content": "Please introduce yourself!"}],

"temperature": 0.9

{"choices":[{"finish_reason":"stop","index":0,� "message":{"content":"Hello! My name is LLaMA, I'm a large language

model trained by a team of researcher at Meta AI. I can understand

and respond to human input in a conversational manner. …",

"role":"assistant"},"references":[]}],

"created":1693848389,"id":"foobarbaz",

"model":"Llama-2-7B Chat","object":"text_completion",

"usage":{"completion_tokens":112,"prompt_tokens":14,"total_tokens":126}}

18 of 30

Some information about ChatGPT (& its siblings)

It is a software system, but its behavior is not programmed�

It is neither a search engine nor a database: it neither searches nor stores data

Like any neural network, it is a parametric function,�trained by adapting parameter values to fit the provided examples

Training: adapt the weights W so that

known expected output = f_W(known given input)

(typically by means of gradient descent of a loss function, as in this tiny example)

f_W

Y = f_W(X)

19 of 30

Some information about ChatGPT (& its siblings):�orders of magnitude

Linear regression: 10⁰ params

Reading handwritten digits: 10⁵ params

GPT-3 / SOTA Transformers: 10¹¹ params

Human brain: 10¹⁵ params

20 of 30

Some information about ChatGPT (& its siblings):�operations

1. training

1.1 pre-training: a large corpus of texts (10¹¹-10¹² tokens) is read;� parameters are adapted by trying to infer some hidden parts (self-supervised learning)� → the net has linguistic and generic disciplinary competences,� but it is a-moral and not specifically able to have conversations

1.2 fine tuning: a smaller set of conversations is read and evaluated;� parameter are further adapted (supervised learning)� → the net has a(n externally imposed) morality and is able to have conversations� → the net has now a “personality”

2. inference / use

21 of 30

Some information about ChatGPT (& its siblings):�basic structure

In the current chatbots a sharp separation is maintained between:

long-term memory: the values of the ~10¹¹ parameters,�set in training and unmodified in use
short-term memory: a relatively small buffer (“context window”, 10³ - 10⁴ tokens),�storing separately the information of each conversation

After their training, current chatbots behave as stateless systems

LLM

LTM

STM

prompt

response

p₁

r₁

p₂

r₂

p₃

r₃

…

LLM: Large Language Model�LTM: long-term memory�STM: short-term memory

22 of 30

Things are still evolving

ChatGPT Plus has an “Advanced Data Analysis” tool and almost 1000 plugins

An LLM (Anthropic Claude 2) has a context window of 100k tokens

Some LLMs (Microsoft Bing Chat, Google Bard) are connected to the web

An open LLM (TII Falcon) has 180B parameters and was trained on 3.5 trillion tokens

Fine tuning techniques are steadily improving (parameter-efficient fine tuning, like LoRA)

…

23 of 30

Introducing what is happening
Trying to understand
Some hypotheses

24 of 30

Consequences: a summary

Current chatbots produce texts that are the outcome of autonomous processing,�from a large amount of texts, not of searches / queries in databases

This makes them novel entities, able to operate in original and sophisticated ways but:

not always trustable in the factual information they report
usually not explainable in their behavior
never accountable for what they produce

25 of 30

Some suggestions of prompt engineering

Using it as instead of a search engine is not a good idea�→ it is more reliable in suggesting good ideas than information on facts
Not only it makes mistakes, but also it presents wrong information as it were correct�→ it must be used with a systematic critical attitude
To generic questions it replies in a generic way�→ questions must be phrased in a specific way to obtain specific responses
It is proficient in conversation, even more than in one-shot Q&A�→ a step-by-step development of the subject is very effective
The inferences it computes have usually a probabilistic component�→ different responses can be obtained by repeating the same question in the same context
It is skilled in impersonating different subjects and complying with given conditions�→ a conversation may be started by giving some specifications (roles, format, …)

26 of 30

The example of a prompt

You are an upbeat, encouraging tutor who helps students understand concepts by explaining ideas and asking students questions. Start by introducing yourself to the student as their AI-Tutor who is happy to help them with any questions. Only ask one question at a time.

First, ask them what they would like to learn about. Wait for the response. Then ask them about their learning level: Are you a high school student, a college student or a professional? Wait for their response. Then ask them what they know already about the topic they have chosen. Wait for a response.

Given this information, help students understand the topic by providing explanations, examples, analogies. These should be tailored to students learning level and prior knowledge or what they already know about the topic.

Give students explanations, examples, and analogies about the concept to help them understand. You should guide students in an open-ended way. Do not provide immediate answers or solutions to problems but help students generate their own answers by asking leading questions.

Ask students to explain their thinking. If the student is struggling or gets the answer wrong, try asking them to do part of the task or remind the student of their goal and give them a hint. If students improve, then praise them and show excitement. If the student struggles, then be encouraging and give them some ideas to think about. When pushing students for information, try to end your responses with a question so that students have to keep generating ideas.

Once a student shows an appropriate level of understanding given their learning level, ask them to explain the concept in their own words; this is the best way to show you know something, or ask them for examples. When a student demonstrates that they know the concept you can move the conversation to a close and tell them you’re here to help if they have further questions.

(source: OpenAI, 31 August 2023, Teaching with AI, https://openai.com/blog/teaching-with-ai )

27 of 30

The “principles” I am proposing to my students

Principles	Consequences
1. Before starting a conversation, X knows neither you nor the context of the conversation. 2. During a conversation, X keeps track of the contents of that conversation, but it has no information on any previous conversation. 3. X is trained to respond in a neutral way to the requests it receives, trying to avoid expressing any controversial opinion. 4. Though trained with a large amount of texts, X is sometimes unable to produce correct responses. 5. X is an, often helpful, assistant, but it is not responsible for the contents it produces.	→ To have a conversation with specific contents, you have to explicitly state its context and objective. → To take into account the contents of a previous conversation, you have to write them again, possibly in a summary form. → To obtain contents other than prevailing, though possibly very sophisticated, opinions, you have to state your questions in ingenious, unconventional ways. → To rely on the contents produced in a conversation, you have to validate them independently. → You are responsible for the use of the contents produced in a conversation.

28 of 30

Beyond “the two cultures”?

«A good many times I have been present at gatherings of people who, by the standards of the traditional culture, are thought highly educated and who have with considerable gusto been expressing their incredulity at the illiteracy of scientists. Once or twice I have been provoked and have asked the company how many of them could describe the Second Law of Thermodynamics. The response was cold: it was also negative. Yet I was asking something which is the scientific equivalent of: Have you read a work of Shakespeare’s?�I now believe that if I had asked an even simpler question — such as, What do you mean by mass, or acceleration, which is the scientific equivalent of saying, Can you read? — not more than one in ten of the highly educated would have felt that I was speaking the same language. So the great edifice of modern physics goes up, and the majority of the cleverest people in the western world have about as much insight into it as their neolithic ancestors would have had.»

C.P. Snow, The two cultures, 1959

29 of 30

A position

It is the first time that we can have conversations in natural languages�with an entity which does not belong to our species

Hypothesis: what is happening around ChatGPT & its siblings�will be the third “cultural revolution” in the Western world:

– Copernicus showed us our cosmological non-centrality

– Darwin showed us our biological non-originality

– chatbots are showing us our cognitive non-uniqueness

This new scenario is generating and will generate�both opportunities and risks

30 of 30

Luca Mari

lmari@liuc.it

https://lmari.github.io

Thanks for your attention!

�(and, if you are interested enough, let’s keep in touch:�things are so new and are moving to rapidly�that sharing experiences and opinions will remain precious)