Blog

Posted on

Gods, Interns and Cogs

Drew Breunig brings some much-needed clarity to the discussion of AI by suggesting that there are three core use cases.

  1. Gods: Super-intelligent, artificial entities that do things autonomously.
  2. Interns: Supervised copilots that collaborate with experts, focusing on grunt work.
  3. Cogs: Functions optimized to perform a single task extremely well, usually as part of a pipeline or interface.

The Gods use case is where most of the angst is focused. At present, it is mainly hype. Who knows how it will turn out?

Interns are tools desgned for use in particular domains by experts in that domain, for example Github Copilot for programmers, Grammarly for writers, and NotebookLM for researchers. I have been using NotebookLM in recent weeks and can vouch for its rapidly growing usefulness and capability. Drew posits that Interns are providing most of the value obtained from AI at present.

Cogs (aka Agents) are essentially components that perform one function correctly and reliably. They are being used to build systems in ways resonant of the software intensive systems we are already accustomed to. It seems certain that these systems will in due course become a driver of technological change.

Azeem Azhar suggests that the Interns category “can be split into two distinct types of AI assistance: “True Interns” and “Copilots”. True Interns can take on multi-step projects autonomously – imagine telling them “build a guest list for our AI conference” and they’ll research speakers, check social media presence, cross-reference past events, and return with a curated list. This is what Claude’s computer use promises to do. Then there are Copilots – AI that works alongside you in real-time, suggesting code as you type or refining your writing on the fly. They’re more like having a knowledgeable colleague looking over your shoulder than an intern working independently.“

Posted on

AI as a Partner in Learning

Although some see artificial intelligence (AI) as our potential nemesis, there is great scope for it to become a transformative ally. In the field of education, AI is already beginning to reshape roles, enhance capabilities, and pose new challenges for educators, students, administrators and support staff. Arguably, we are heading towards a future in which AI’s role is that of a collaborative partner in the educational journey. How might that play out in a university context?

Continue reading “AI as a Partner in Learning”
Posted on

Long-term memory for Generative AI

Large language models (LLMs) such as ChatGPT have embedded within them and can make use of the huge amount of information they were fed during training. A user is able to access that embedded knowledge by giving the LLM instructions during the course of a conversation with it. At present, however, the LLM has a limited capacity to remember the details of a conversation; that capacity is determined by the size of its context window.

Continue reading “Long-term memory for Generative AI”
Posted on

The growth in ChatGPT’s capability

The capabilities of ChatGPT are increasing at pace. The latest upgrade turns it into a multimodal AI. Instead of being restricted to text-only input and output, ChatGPT can now accept prompts with images or voice as well as text and can output its responses in one of five AI-generated voices. A user can switch seamlessly between text, image and voice prompts within the same conversation.

Browse with Bing enables ChatGPT to search the internet to help answer questions that benefit from recent information.

Advanced Data Analysis (formerly called Code Interpreter) enables ChatGPT to upload and download files, analyze data, do maths, and create and interpret Python code. These are powerful capabilities but there are restrictions which include: no internet access; a limited set of preinstalled packages; maximum upload and runtime limits; state is cleared (along with any generated files or links) when the environment dies.

Open Interpreter is an open source project which seeks to overcome the restrictions of Advanced Data Analysis. Open Interpreter runs in your local computer and interacts with ChatGPT. It has full access to the internet, is not restricted by time or file size, and can utilize any code package or library. Thus Open Interpreter combines the power of GPT-4’s Advanced Data Analysis with the flexibility of your local development environment.

Plugins enable ChatGPT to interact with functionality provided by other systems. Examples are:
Wolfram Plugin for ChatGPT gives it access to powerful computation, accurate maths, curated knowledge, real-time data and visualization through Wolfram|Alpha and Wolfram Language.
Show Me ChatGPT Plugin allows users to create and edit diagrams directly within a conversation in ChatGPT. 
There is a growing number of plugins; some are shown here.

Posted on

Plugins expand ChatGPT’s capability

ChatGPT has the ability to make use of third-party plugins which give it access to external sources of information. This is useful because it enables to AI to apply its impressive language capabilities to information that was not in its training data and, unlike the training data which is now two years old, that information can be current.

ScholarAI is a ChatGPT plugin “designed to provide users with access to a database of peer-reviewed articles and academic research“. In this conversation with ChatGPT, I explore a little of what the AI can do when the ScholarAI plugin has been installed. I found that it was able to search for papers, on a given subject, summarise the content of a paper, and answer questions about that content. I have not yet investigated the quality of the answers provided.

Plugins can also provide ChatGPT with additional functionality. In an earlier post, I mentioned the prospect of the AI interfacing with Wolfram Alpha. The Wolfram Alpha plugin is one instance of that, and it enables ChatGPT to give correct answers to prompts that require computation. See below for an example. We can be confident that answers obtained from Wolfram Alpha are of high quality.

There are many plugins to choose from. Websites such as whatplugin.ai can help us to find the ones we need.

Continue reading “Plugins expand ChatGPT’s capability”
Posted on

How are neural networks implemented?

What can ChatGPT tell us about its inner workings? I instructed it to take the role of a professor in the faculty of engineering at University College London and said that I had the role of a third year undergraduate student. A full transcript of the conversation is here.

To begin, I asked the professor for a high level explanation of network theory, telling him to express the ideas in British English and without using jargon, to take it step by step, and to begin by getting my agreement for what would be the steps in the explanation. After orientating this outline more towards AI networks we experimented with the professor’s ability to reflect on what he had produced – ie to re-evaluate his previous outline. This resulted in a notably better structure for the explanation I sought – see below.

Professor ChatGPT then led me through the subject one step at a time. There were many diversions as I asked him to clarify things. We got as far as Step 2.2 before one of these diversions took me down a path to a good breakpoint: agreement on an understandable explanation of how the concept of an AI’s neuron – a node in its neural network – is implemented in software and hardware and how very large numbers of these nodes can execute simultaneously.

Continue reading “How are neural networks implemented?”
Posted on

ChatGPT’s temperature

In a previous post I mentioned a parameter called ‘temperature’ which ChatGPT uses to vary its answers from being more focused and deterministic to being more diverse and unpredictable; a kind of creativity, perhaps. This is how the AI explains its ‘temperature.

Question: Consider the questions I have asked you in this conversation. If I now ask you the same questions in the same sequence, would your answers be different and, if so, might they contradict the answers given first time around?

ChatGPT

As an AI, my responses are generated based on patterns learned from a large dataset during training, and are influenced by the specific wording and context of the prompts given to me. If you ask the exact same questions in the same sequence, you should generally expect to receive the same or very similar answers, assuming the underlying model and its parameters remain unchanged.

However, GPT-4, like previous versions of the model, includes an element of randomness in generating responses. This is intended to make the output more diverse and to avoid always giving the exact same response to the same input. As a result, the specific wording and phrasing of my responses may vary slightly each time you ask the same question.

Continue reading “ChatGPT’s temperature”
Posted on

ChatGPT is not intelligent

As I understand things, ChatGPT uses probabilistic methods applied to a massive amount of data (currently the language content of the internet up to September 2021) to predict what word is most appropriate to follow the preceding words in its output. To kick-start that process you give it an instruction which guides the form of its response. So it’s answers are a re-packaging of the previously-written material on which is has been trained; it does not create new ideas. There is a parameter called ‘temperature’, however, which can vary its answers from being more focused and deterministic to being more diverse and unpredictable; a kind of creativity, perhaps. 

At present, therefore, we are the intelligent agents and ChatGPT is simply our assistant. Our assistant can retrieve information quickly and package it in ways that can help us think through the ideas we are pursuing. Also, we can ask it to do things that will help us analyse the matter at hand, for example by taking large stacks of data from several sources, combining them and charting certain characteristics. And when we ask it to identify connections between things, it will sometimes find connections we would not have thought of ourselves. 

Continue reading “ChatGPT is not intelligent”
Posted on

Synthetic Interactive Persona Agents

In this article by Sunil Ramlochan, a Synthetic Interactive Persona Agent (SIPA) is predicated as “offering realistic and effective dialogues that align with real-world human tendencies“. That is, it can take the place of human participants in online conversations. To illustrate what can be achieved, Sunil takes us through the creation of a fictional character: Jack Thompson is from Texas and identifies as a Republican. The task is to:
“incorporate the socio-political context of Texas, Republican values and beliefs, and personal characteristics in developing John’s persona;
simulate how John would think, speak, and act in various scenarios, ensuring his actions reflect his political identity;
make the depiction … respectful and nuanced, avoiding stereotypes while highlighting the individuality and complexity of John’s persona.“
Jack is then asked questions and his answers shown. At face value, Jack’s answers seem convincing.

Sunil suggests eleven use cases out of many possibilities. Most seem reasonable, for example the creation of synthetic data for use in the training of customer service agents and chatbot agents. Others give pause for thought, for example to simulate the political viewpoints of different demographics for use in political polling. Hence the article ends with thoughts about limitations and ethics and urges us “not to misrepresent synthetic data as actual human responses in any form of reporting or decision-making“. That will prove quite a challenge.