Before we begin…

• A little survey to get to know you better.
• We will come back to this at the end of the session to see how your views have changed!

Who are we?

Courtesy of MS Copilot AI tool…

Who are we?

Courtesy of Google Gemini AI tool…

Who are we really?

• Prof. James Brusey: Director of the Coventry University AI Adoption Lab.
• Prof. Matthew England: Director of the Coventry University Research Centre for Computational Science and Mathematical Modelling.
• Dr James Shuttleworth: Director of the Coventry University AI Skills Academy.

What do we aim to do today?

• Get you up-to-speed on AI.
• Debunk some common AI myths.
• Explore opportunities and risks from AI in 2025.
• Build your confidence in understanding today’s AI landscape.

Myth

• AI has exploded on the scene because finally someone had the right idea.

Strand 1: perceptrons, neural nets, and backprop

• Minsky and Papert (1969) “Perceptrons” found a key problem with neural nets.

Backprop

• Key idea: even in a high-dimensional space, it is possible to find the optimal direction.

How to test?

• Training performance doesn’t tell us enough.

• We need unseen data (not used during training).

• Cross-validation makes the best use of available data.

Hardware?

• A great leap in performance came when it was realised that GPUs were perfect for neural nets.

Strand 2: Reinforcement Learning

• Early approaches to teach a computer how to act were about rewards and punishment.

Actor-critic (Barto, Sutton, Anderson 1983)

Q-learning (Watkins 1989)

• Chris Watkins united experimental ideas with Bellman’s decision theory.

• Q-learning provably converges on the optimal strategy.

How do we test it?

Ultimate test is if it works in the real world.

AlphaGo and AlphaZero (Google DeepMind 2016)

Strand 3: Natural Language Processing and Translation

“The chicken is ready to eat.”

1. The chicken is cooked.
2. The chicken is hungry.

In French, these translate differently:

• Le poulet est prêt à être mangé (cooked).
• Le poulet est prêt à manger (hungry – odd but grammatical).

Correct translation requires semantic role understanding, not just syntax.

How do we test it?

• BLEU = Bilingual Evaluation Understudy.

• Higher BLEU score ⇒ more overlap with the human reference text.

• BLEU score needs a reference dataset (e.g., WMT, which is news stories).

Attention is all you need

• Google 2017 paper showed that you didn’t need recurrence - self-attention was sufficient

• This proved to be surprisingly general.

• Things got exciting when they scaled things up (more data, bigger models)

But wait, there’s more!

• If you can predict the next word reliably, what happens if you give:

Q: Why is the sky blue? A: _

• This idea gave birth to ChatGPT

Generative AI vs Analytical AI

• Analytical AI analyses data to make predictions. E.g.
  • Predict the price of this house from its properties.
  • Predict if this medical image indicates cancer.
• Generative AI creates new content by following patterns in data. E.g.
  • A Large Language Model (LLM) produces new text in response to a query.
  • DALL-E produces new images from text.
• Key difference: Generative AI usually pre-trained; Analytical AI needs training on your own data.

Myth: AI in 2025 is all generative

2024 Chemistry Nobel Prize to AlphaFold.

Frontier Models Quiz

Frontier Models Quiz

Myth: All AI is made in the USA

Myth: All AI is made in the USA

How to Access AI?

• The frontier models are usually accessed via the cloud: either in app/browser of via API.
• Frontier models usually operate via a subscription service.
• Myth: AI must involve expensive subscriptions:
  • Cloud services usually have multiple tiers; including a free tier!
  • There are open source LLMs you can download for free and run locally.

Warning!

• When using the free tier, your data may be used to train future AI! Consider GDPR implications!
• If you have sensitive data you probably want either a local AI, or a paid for cloud AI service.

Myth: Open-Source Models Poor

How To Access AI Today?

• Coventry University staff can use Microsoft Copilot AI from within a browser at this address: https://copilot.cloud.microsoft/
• External guest can use their favourite AI service!
• If you don’t have one yet:
  • Use your Google account to access Gemini: https://gemini.google.com/
  • Or you can quickly setup a free ChatGPT account: https://chatgpt.com/

Task 1

• Get the LLM to identify the top 5 changes in the 2025 UK Budget.
• After the first attempt, try again but tailor your query to your business / personal interests.

Coventry University Microsoft Copilot AI: https://copilot.cloud.microsoft/

ChatGPT: https://chatgpt.com/

Google Gemini: https://gemini.google.com/

Task 1: Reflection

Get the LLM to identify the top 5 changes in the 2025 UK Budget.

• How useful was the report?
• Can you verify the claims given easily?
• Was there something missed out that you know is important to you?
• Compare to the BBC’s report here: https://www.bbc.co.uk/news/articles/cj4w44w42j5o

Task 1: Your Thoughts?

Task 1: My Thoughts

• AI does a much better job at this in 2025 than in previous years:
  • Can access up-to-date information easily.
  • Can share sources of its information for validation.
  • Choice of sources better than last year.

Myth: AI cannot reference sources

Myth: AI thinks like us

• An LLM is a Machine Learning model to predict the next work in a text, based on the previous.
• E.g. If we type “the cat sat on the” then a good prediction for the next word is “mat”.
• The prediction is based on a statistical analysis of the training data, not experience with cats.

Myth: AI will Destroy the Planet

• Frontier models have been getting bigger. E.g. GPT3 had 175bn parameters. GPT5 does not disclose but estimates range in the trillions.

• However, recent performance gains may have come from new techniques (e.g. chain-of-though,mixture-of-experts) rather than scaling.

• There is now focused competition on performance with restricted resources.

Hallucinations

LLMs can produce hallucinations: false claims.

More common when: queries require unavailable knowledge, ambiguous, need complex reasoning.

Hallucinations in 2025

• Myth: AI is unreliable: LLMs are hallucinating a LOT less than in 2023/2024.

• Myth: AI will soon be perfect: Hallucinations are still possible: there are many strategies being researched to reduce the risk of hallucination further. But we are unlikely to ever fully remove them.

• Users need to evaluate how likely hallucinations are in their AI application, what risk they would pose, and how that risk might be mitigated.

What else is new in 2025?

• AI is increasingly multi-modal.
• Most frontier models can now produce images.
• There are tools to produce videos, audio, …

Task 2 A: new logo

• This is the Coventry University Phoenix:

• Can you get your AI tool to design us a new Phoenix logo for the AI team at Coventry?

• What happens if you give it the original logo to work with?

Task 2: My Images

Task 2: Your Thoughts?

Lessons from Tasks

• Prompting matters: how the question is asked greatly affects the output.
• Some people call this prompt engineering.
• Prompts can be very large (but there is a limit).
• The previous conversation is part of the prompt!
  • Showing the AI the existing logo resulted in new images more similar to the original.
  • If you have a long conversation about your political beliefs, then ask for the budget report, expect some bias towards your views!

Prompt Engineer

Alternatives?

• Fine Tuning: Taking trained AI and small dataset specific to your application to further train on.

  • E.g. Train it on logs of your customer services calls to show how your team deals with things.

• Retrieval Augmented Generation (RAG) allows trained AI access to your knowledge base.

  • E.g. The LLM must access your HR manual and policies in answering HR queries.

There is lots more

Source

A few more myths I

• Myth: AI is sentient: It is simply not.

• Myth: We have no idea how AI works: The basic ideas are published and well studied.

• Myth: AI is completely objective: AI is trained on data: biases in data will appear in the model.

• Myth: AI is too biased to use: There has been progress in making AI safer. Challenges remain: users should plan how they will evaluate AI output to identify and deal with any bias.

A few more myths II

• Myth: AI models now explain how they think: One of the advances in 2024/2025 has been reasoning models. However, the ``explanations’’ produced do NOT to always accurately reflect the AI’s working.

• Myth: We will never understand how AI thinks: Research in 2025 on AI ``brain scans’’: seeing which neurons fire as a query is processed, revealing patterns in how queries are dealt with.

Handover

RAG is perhaps the first step to agentic AI: the next big AI breakthrough? Let’s see…

Short-Term Future

• What is likely to happen in the next year or two?

• Increased use of agentic AI.
• Changes to GenAI pricing.

• Increasing standardisation of protocols, etc. and evolution of those standards.

• Evolution of the tools and interaction mechanisms that wrap GenAI.

• Increased use of analytical AI.

• Smaller, specialised GenAI models:
  • Easier alignment, narrower range of output, lower resource requirements.
  • Reduce dependence on big compute providers and AIaaS.

The Optimistic Future

• What will AI be like in 5 years? 10 years?
• What are people predicting?

An Optimistic Future of Generative AI

• New ways of building systems that use GenAI:
  • Expanding Agentic AI Capabilities with new tools and techniques.
  • Improvements in RAG (Retrieval Augmented Generation).
  • Interfaces that support better prompt-writing.
  • Compositional architectures: Mixture of Experts, coordination of multiple models.

• Going beyond training on all available data.
  • The Era of Experience:
    • Experiential learning for AI systems, reward tied to real-world outcomes, rich observation.
• Scaling continues:
  • More data is made (ethically) accessible, more compute, larger models.
  • New architectures that are more efficient.
  • Bigger context windows.

• Free, clean, infinite energy.
• Advances in safety and alignment make all AI use safe, reliable and ethical.
• Quantum Computing breakthroughs.

An Optimistic Future of Analytical AI

• More data available.
• Advances driven by GenAI demand mean cheaper compute, storage, bandwidth.
• Data science and analytics tools become more powerful and easier to use.
• Data literacy improves.

What Could Go Wrong?

• What follows are real cases of how things can go wrong.
• We can use these to help us imagine new and more exciting ways for things to go wrong in the future.
• Not all of these are LLM failures.

A possibly apocryphal story about tanks

And many more…

Awful AI is a curated list to track current scary usages of AI

What are people worried about?

• Bias and Discrimination
• Privacy and Data Security
• Job Displacement and Automation
• Accountability and Transparency
• Surveillance and Control
• Autonomy and Decision-Making

• Manipulation and Misinformation
• Lack of Human Oversight
• Ethics of AI in Warfare
• Fair Access to AI Technologies
• Human Agency and Dependency
• Moral Status and Rights of AI

• The rise of an unaccountable global oligarchy
• Massive job losses
• Increased social isolation and mental illness
• Total invasion of privacy
• Higher likelihood of war
• Environmental degradation
• Human control of the planet

This slide written by AI Bernie Sanders

The Pessimistic Future of Generative AI

• Most of the available data has already been used for training. This is as good as it gets.
• In 20 years time, we are excited about a neural network that uses Quantum Computing. It has 5 parameters and we can’t look at it without crashing it. But any day now…
• Everyone is now a reverse centaur

• Rogue agents can have catastrophic consequences.
  • But they work most of the time, and we can’t stop using them because it would mean going back to a levels of productivity that are unacceptable.

• Training data has been thoroughly compromised, meaning the “best” models are also the most subtly corrupted. See CopyCop
• Sleeper Agents (AI systems that appear benign but have hidden objectives that can be activated later) crop up everywhere.
  • The training data contained hidden instructions and associations that we didn’t notice and now it’s too late
  • We can’t just start again because most of the data available now has been written by LLMs.

• Nothing is new any more.
  • All LLMs produce is exceptionally good examples of current thinking and everything starts to converge.
  • All logos look similar, all policies are the same, books become re-skinned copies of the most popular tropes, etc.

Influencing the Future

• How do we guide the future of AI adoption and development toward the most optimistic future?
• This isn’t a solved problem.
• But it probably will involve…

• Regulation:
  • Complicated.
  • May aid monopolies.
  • Compare to current problems regulating more easily understood/deterministic technology such as encryption, social media, etc.
  • Compare to the problems of cybersecurity over the past 30 years.

• Ethics being fundamental to education, particularly when data and AI are involved.
• Distinguishing hype from reality - a skill? Or does this need regulation like advertising?
• Training and education for a broad understanding of AI and data science.
• Development of models for calculating monetary, societal, ecological cost of AI.

What can you do?

• Stay informed:
  • application of AI is evolving rapidly!

• Engage in discussions about AI ethics and policy:
  • your voice matters!

• Advocate for responsible AI development and use:
  • make your opinions known.

• Support organizations working towards ethical AI:
  • if AI is in everything, support those trying to make it good.

• Participate in workshops and seminars to deepen your understanding of AI:
  • broaden your perspective.

Q1

LLM stands for…

• A: Local Language Model

• B: Large Language Machine

• C: Large Language Model

• D: Local Language Machine

Q2

Which one of these is NOT the name of an LLM…

• GPT
• Llama
• Gemini
• Virgo
• Qwen
• Deepseek

Q3

AlphaFold won the Nobel Prize for…

• Computer Science
• AI
• Chemistry
• Physics
• Peace

Q4

When testing an AI it is important to…

• Test on the same data used for training.

• Test on different data to that used in training.

• Test on more data than used in training.

• Test on less data than used in training.

Q5

Which of these were NOT important points on the journey to AI:

• Perceptrons
• Neural Networks
• Back propagation
• GPUs
• Skynet
• Reinforcement Learning
• Self-attention

Q6

Which of these is NOT a myth…

• A: AI is sentient
• B: All AI is made in the USA
• C: Open source AI models are poor
• D: AI is always right
• E: AI can get things wrong

Q7

RAG stands for:

• A: Rapid, Aligned, General – AI that is efficient, ethical and works on many problems.
• B: Refine, Apply, Generate – Generative AI models that iteratively refine their output.
• C: Retrieval Augmented Generation – GenAI use of an external knowledge base to provide more accurate and contextually relevant responses.
• D: Regulation of Autonomy and Generation – The body responsible for enforcing GenAI legislation.

Q8

Agentic AI is gaining in popularity because:

• A: It’s approved by the FBI
• B: It has human-level intelligence
• C: It is cheaper to run than non-agentic AI
• D: It allows AI systems to interact with the real world and achieve more complex goals

Join us for more Bitesize Sessions

Has the session helped?

Feedback Form

• Please complete our feedback form:

Goodbye

Merry Christmas Everyone!