LinkedIn

Table of contents

  • Year autopilot was first introduced
1912

Year autopilot was first introduced

~90%

of a modern commercial flight runs on autopilot

Yet we don't see industry getting excited about this technology that we literaly entrust our lives to….

…Instead we see industry getting excited about a technology that finds it difficult to do primary school tasks.

sure this was 2024 and things should have gotten better by now…

…Or maybe not!

Deterministic AI vs Probabilistic AI

Desiring "Artificial" Intelligence


The mechanical turk was the first chess playing automaton.

…except it wasn't

Q: How do we really put a "brain" into a machine?

Encoding human intelligence?




"Your 8 year child wakes up on Monday morning and says she has a stomach ache"

How would you diagnose and troubleshoot this situation?

             Child says:
         "My stomach hurts."
                 │
                 ▼
       Does the child have a fever?
         ├──────────────┐
        Yes             No
         │              │
         ▼              ▼
 Possible illness     Is today Monday morning?
 Keep home / doctor      ├──────────────┐
                        Yes             No
                         │              │
                         ▼              ▼
                Did the child ask      Ate something
                to stay home?          unusual?
                  ├───────┐             ├───────┐
                 Yes      No           Yes      No
                  │        │            │        │
                  ▼        ▼            ▼        ▼
          Possible school  Go to     Observe   Mild stomach
             avoidance     school    at home      upset

"Expert Systems": Encoding intelligence at scale


Symptom Checker with Body from WebMD - Check Your Medical Symptoms

Uses: Sequential Decision Making and plans it from a specified model (If this….then that)

Limitations of "Expert System"

Fairly Narrow Scope

No Learning

We cannot make an expert of tasks we do not know how to do.

Moving beyond "if- then" encoding and making the machine actually learn




Enter: Machine "Learning"

Machine learning is a method of training computers to find patterns in data and make predictions or decisions

without being explicitly programmed for every rule.

It improves as it sees more data and feedback.





Examples → Training → Model → Prediction!

Now, We You are going to train a

Machine Learning algorithm,

test it, try and fool it,

and then decide when you should trust it.

teachablemachine.withgoogle.com

Teachable Machine

Train a computer to recognize your own images, sounds, & poses. A fast, easy way to create machine learning models for your sites, apps, and more – no expertise or coding required.

What worked and what didn't work?

Questions for Leaders:

  • Data and labels: What did the model learn from, and who labelled it?
  • Bias and testing: Who does the model work for, and who does it fail for?
  • No-code governance: Who is allowed to build, deploy, and rely on this model?


Airport Security AI Challenge

You are an AI architect

Your airport has received €1,000,000 to build one Machine Learning system that will improve passenger security while reducing waiting times.

You have 10 minutes to prepare a proposal for the airport board.

Your goal is not to build the AI.

Your goal is to ....

Design Your AI

  1. What business problem are you trying to solve?
  1. What exactly should the AI predict or classify?
    (What is the output?)
  1. What data would you need to train it?
    (Images, video, sensors, historical records, etc.)
  1. How would you obtain and label the training data?
    (Who decides what is "correct"?)
  1. How will you know your AI is successful?
    (Accuracy? Faster queues? Fewer false alarms? Lower cost?)


Reinforced Learning an underlying algorithm to how machines learn (origins in 1950s, blossomed in 2013)

Emergent strategy, minimal rules, maximizing rewards and learning through repeated self play



Random Bounce of a Roomba

CNET

Why your Roomba takes a weird path to keep your floors clean

Robot vacuums use many methods to navigate and tackle the world. Some rely on lasers, others electronic eyes or even simple bumpers, and these tactics affect how well they clean. Our test room helps us show the difference.

Reinforcement Learning Playground: Why your spotify, keeps playing the same songs!

Demystifying ML: What Machine Learning can really do!

The Brain: Neural Networks

We've seen that ML can Classify, Cluster, or Act/Decide. But how does the machine actually compute all of this at scale? The answer is Neural Networks — the computational engine that powers all three learning strategies.




The Black Box: Inputs & Outputs

Neural networks process various inputs to generate specific outputs. While the inputs and outputs are clear, the intricate connections and transformations within the hidden layers make the internal decision-making process incredibly complex and often inexplicable. This characteristic is often referred to as the "black box" problem, highlighting the challenge of understanding how the network arrives at its conclusions.

Suppose Waze wants to predict

"How many minutes will it take to drive this road?"

Inputs might include

  • Time of day , Day of week. Weather. Current speed
  • Historical traffic, Holidays, Nearby events

Output

  • Expected travel time

playground.tensorflow.org

Tensorflow — Neural Network Playground

Tinker with a real neural network right here in your browser.

The Power of Deep Learning

AlphaGo Computing Power:

The Economist reported that it used 1,920 CPUs and 280 GPUs:

(about 2000Terraflops or 1000 iphone16s)

More on AlphaGo:

Sage Journals

Deep new: The shifting narratives of artificial intelligence from Deep Blue to AlphaGo - Paolo Bory, 2019

The article compares two key events that marked the narratives around the emergence of artificial intelligence (AI) in two different time frames: the game serie...

WIRED

Inside the Epic Go Tournament Where Google’s AI Came to Life

The battle between Google's AlphaGo AI and Go champion Lee Sedol was more than just a game. It was proof that AI can think like us---and make us better.

Creativity: Is this still a skill only humans possess?

Lee goes on to lose game three, and AlphaGo secures victory in the best-of-five series. At the press conference afterward, with Hassabis sitting next to him, Lee apologizes for letting humanity down. “I should have shown a better result, a better outcome,” he says.

As Lee speaks, an unexpected feeling begins gnawing at Hassabis. As one of AlphaGo’s creators, he is proud, even elated, that the machine has achieved what so many thought it couldn’t. But even he feels his humanness rise. He starts to hope that Lee will win one.



New Knowledges


Targeted Oncology

AlphaFold AI Sets Stage for Future Approaches in Cancer

The 2024 Nobel Prize in Chemistry highlights AI-driven breakthroughs in computational protein design and structure prediction. Tools like AlphaFold promise faster development of targeted cancer drugs and transformative advances in oncology and medicine.

Drug Discovery: Insilico

insilico.com

Main | Insilico Medicine

Generative AI and Automation for Longevity and Sustainability

AI and chip design

Popular Mechanics

AI Designed Computer Chips That The Human Mind Can't Understand.

That might be a problem.

Medium

An AI Found a 27-Year-Old Bug Hiding in OpenBSD. It Cost Less Than $50 to Find It.

For 27 years, every security expert, every fuzzer, every automated scanner missed it. Claude Mythos found it in a single run. That one fact…

Scientific American

AI just solved an 80-year-old ‘Erdős problem,’ and mathematicians are amazed

A chatbot’s result for the 80-year-old “unit distance” conjecture is the first AI proof that would likely be published in math’s top journal if humans had done it alone

AI and innovation Frontiers

Top 10 Emerging Technologies of 2024

2025 Update: https://www.weforum.org/publications/top-10-emerging-technologies-of-2025/

Enter Gen AI: And the role of the Transformer


Financial Times

Generative AI exists because of the transformer

The technology has resulted in a host of cutting-edge AI applications — but its real power lies beyond text generation

The ChatGPT Moment



This may not be a bug

Medium

Why LLMs Use Tokens Instead of Words: A Complete Guide

If you’ve ever wondered why AI models like GPT or Claude talk about “tokens” instead of just counting words, you’re not alone. This…

platform.openai.com

OpenAI Platform

How tokens work

ML vs GenAI Usecases: A Comparative Overview

No code Innovation?

Let's try it at

Figma

Figma Make: Create with AI-Powered Design Tools

Figma Make empowers you to turn ideas into reality with AI-powered design tools—generate, iterate, and build faster, all in one creative space.

Base44

Base44

Build useful apps, fast with Base44.

Lovable

AI App Builder | Vibe Code Apps & Websites with AI, Fast

Build apps, websites, and digital products faster using Lovable’s no-code and AI-powered platform, no deep coding skills required.

If you can "Generate" text, video, sound, concepts … what else can you do?

What might be issues you might look at in GenAI?"


Are we asking the right questions?

Are we testing out of sample?

Are we prepared to deal with probabilistic results?


Out of sample Data: One of the many ways AI fails

The training data does not include a situation that you want it to answer

Model Hacking ADAS to Pave Safer Roads for Autonomous Vehicles | McAfee Blog

Custom GPT

& NotebookLM

Explainable AI with SHAP


Core idea: SHAP is a post-hoc method. After a model makes a prediction, SHAP shows how each feature moved the output away from a baseline.

Start with the prediction problem

example dataset Califo

DATASET

20,640  California block groups observed in 1990

Eight inputs describe each block group:


PREDICTION

Given these eight characteristics, what value will the model predict for this block group?

EXPLAINABILITY

After the model predicts, which features pushed that prediction up or down — and by how much?

That is the question SHAP answers.

SHAP is a post-hoc explanation tool

The model is trained first. SHAP comes afterwards to explain the model's output.

1

1 Train a model

Use the housing data to fit a model.

The model learns a mapping from the eight inputs to the target.

2

2 Make a prediction

For one block group, the already-trained model produces an output f(x).

3

3 Explain it with SHAP

SHAP interrogates the existing model and attributes the prediction to its input features.

How SHAP explains one prediction

Start from the expected model output, then allocate the difference to the features.

Explainable AI with SHAP · Source: SHAP documentation, "An introduction to explainable AI with Shapley values."

Transformers: the "features" can be tokens

The source tutorial applies SHAP to a DistilBERT sentiment model and attributes the prediction back to pieces of an IMDB review.

1

INPUT — IMDB review text

SHAP comes after the prediction and allocates the output back to the input tokens.

2

MODEL — DistilBERT sentiment model

3

OUTPUT — Sentiment logit

"This film is interesting as an experiment but tells no cogent story."

Selected token attributions from the tutorial

interesting

+0.063

experiment

-0.236

but

+0.366

tells

-0.329

co

+0.037

gent

+0.041

story

+0.051

How to read it: Red → pushes toward positive sentiment · Blue → pushes toward negative sentiment

Representation matters: the tokenizer can split a human word into multiple tokens (for example, "co" + "gent").

Decision-maker takeaway: this is a post-hoc attribution of model behavior — not the transformer's internal "reasoning."

Explainable AI with SHAP · Source: SHAP documentation, "An introduction to explainable AI with Shapley values."

What SHAP does not solve

A post-hoc explanation is useful evidence, but it is not automatic assurance.

Explanation is not accuracy. A wrong model can still produce a neat explanation.

Explanation is not causality. A feature contribution does not prove that changing the feature would cause the outcome.

Post hoc is not intrinsic transparency. SHAP explains model behavior from the outside; it does not expose an internal "reasoning process."

Bias and Fairness

Bias

Data

THE ECONOMIST unlocking-enterprise-ai.pdf

Bias in ML: Implication and Sources

AI Bias: 16 Real AI Bias Examples & Mitigation Guide Joel Bervell, MD on Instagram: "Racial Bias in Medicine Episode 1: Pulse Oximeters I’m bringing back my “Racial Bias in Medicine” series and adding updates about the changes that have, and have not be COMPAS : Unfair Algorithm ?


Fairness

Medium

COMPAS Case Study: Investigating Algorithmic Fairness of Predictive Policing

Bernard Parker, left, was rated high risk; Dylan Fugett was rated low risk. (Josh Ritchie for ProPublica)

Medium

COMPAS : Unfair Algorithm ?

Visualising some nuances of biased predictions and where they come from

File upload

Farness explained

156.3 KB

Questions for decision-makers

  1. Where are we treating a probability as a certainty?
  1. Could we actually explain our own model, today, if asked?
  1. Is that "risk signal" real — or a proxy for region/sector we shouldn't be deciding on?
  1. Who can challenge a model, not just approve it?
  1. Are we ready to prove compliance if the AI Act comes asking?
  1. Probability vs. certainty: Most models output a confidence, not a fact; if no one restates that confidence when reporting up, it gets treated as fact by the time it reaches a decision. The key is to check confidence level.
  1. Can we explain it today? "We could" and "we have" are different claims; only the second holds up under audit. And puristically explainability is not possible…like trying to explain one's preference for a particular flavor of icecream. we know which but not why
  1. Real signal vs. proxy : Check what the feature correlates with, not just what it predicts — region/sector can stand in for protected characteristics without naming them.
  1. Who can challenge it: Approval needs someone with enough context to say no, not just a name on a sign-off form. And there must be a person that is authorized to challenge it.
  1. AI Act readiness — High-risk classification depends on use case, not on the word "AI" — map your models against the Act's categories, don't assume.