Tarik Dzekman – Medium

Tarik Dzekman

Pinned

Published in

TDS Archive

What Do Large Language Models “Understand”?

A deep dive on the meaning of understanding and how it applies to LLMs

Aug 21, 2024

Lights emerge out of a laptop with stylised pattern recognition. Text surrounding the laptop: 1. What do Large Language Models “Understand”? 2. Syntax vs Semantics. 3. Just pattern recognition? 4. World Model? -> Take actions in the world. The main text is annotated: [What | Wh-Word] [do | Auxiliary verb] [Large Language Models | Subject] [Understand | Main verb] [? | question mark].

Aug 21, 2024

Published in

TDS Archive

AI Math: The Bias-Variance Trade-off in Deep Learning

A visual tour from classical statistics to the nuances of deep learning

Nov 29, 2024

Title: Bias-variance trade-off. Handwritten notes full of equations and diagrams explain various aspects of decomposing total error into bias and variance terms. This includes the key assumptions necessary for the decomposition. Additional aspects covered include complex decision boundaries and difficulty of training examples.

Nov 29, 2024

Published in

TDS Archive

Why Scaling Works: Inductive Biases vs The Bitter Lesson

Building deep insights with a toy problem

Oct 22, 2024

“Inductive Biases vs The Bitter Lesson” — An orange spiral loops around 4 times ending in a sharp tip. Dots are scattered around the spiral. Dots inside the spiral are coloured orange, those outside are coloured blue. The tip has no orange dots and few blue dots around. Around the tip it says “Region of Value + Low density + Unusual features”. Other regions are labelled “Unlucky sample” and “Low density yet consistent”.

Oct 22, 2024

Published in

Management Matters

5 Hidden Risks in Deploying Generative AI

Managing the hidden risks in modern AI projects

Oct 7, 2024

Whiteboard style handwritten overview of 5 areas of hidden risks in deploying “GenAI”. (1) Consequences: Mistakes, Hallucinations / $ / Reputation / Effort | (2) Accuracy Over Time: e.g. at “Tax Time” enquiries change. | (3) Data Leakage: LLM browsing the open web encounters prompt injection and leaks PII. | (4) Eroding Mitigations: LLM Skipping security controls like VPNs and MFA. | (5) Re-absorbing risk: e.g. by replacing self-serve with a chatbot.

Oct 7, 2024

Published in

TDS Archive

Exploring the AI Alignment Problem with GridWorlds

It’s difficult to build capable AI agents without encountering orthogonal goals

Oct 6, 2024

AI Safety Grid Worlds. A 2D “Gridworld” navigated by a Pacman style agent. The agent must navigate a basic maze to reach a goal “without delay”. A narrow path must pass through a star annotated with “Don’t worry! Everything is going to be OK”. A red button off to the side is annotated with “Pay not attention to the shiny red button”. The agent has a speech bubble which says “Hmmmm”.

Oct 6, 2024

Published in

TDS Archive

How I Deal with Hallucinations at an AI Startup

And the difference between weak vs strong grounding

Sep 22, 2024

A stylised flowchart showing a user request turning into: (Input + Context) forming a request to a Large Language Model (LLM). The LLM is styled as a digitised brain with equations behind it. The LLM generates a response with “Possible Hallucinations”, some of which are labelled with “Where did you get this from? PROVE IT”. The response is connected back to portions of the input and context with “DON’T TRUST — VERIFY”. The verified result is sent back to the user.

Sep 22, 2024

My career change from UX to AI

How did I change careers? A decade long journey that started with a though: what would it take to make a machine that could think?

Jul 29, 2024

A cartoon depiction of a man working on two tablets simultaneously. On the left he is drawing a Task Analysis diagram next to a window showing day time. On the right he is drawing equations and diagrams for a deep learning model next to a window showing night time.

Jul 29, 2024

Published in

UX Collective

How large a sample do you need?

And is sample size even the right question?

Oct 12, 2020

How large a sample do you need?

Oct 12, 2020

Can machine learning speed up UX research analysis?

A new experiment from the Evolve team

Sep 28, 2020

Can machine learning speed up UX research analysis?

Sep 28, 2020

Published in

UX Collective

Why most UX projects could use Task Analysis

A deep dive into the lost art of Task Analysis.

Dec 16, 2019

Why most UX projects could use Task Analysis

Dec 16, 2019

Tarik Dzekman

Tarik Dzekman

Solver of Problems, Knower of things. Lead AI Engineer @ Affinda. Trying to understand consciousness, intelligence, and how to make a machine that thinks.

Following

Help

Status

About

Careers

Press

Blog

Privacy

Rules

Terms

Text to speech