kjam's blog

In this mega-post, you'll dive into the world of AI evaluations and how they might be used to evaluate privacy.

A few months ago, NVIDIA CEO Jensen Huang announced their version of NanoClaw, which promised to offer better privacy, security and performance compared to the popular open-source tool OpenClaw.

There's been a lot of chatter about guardrails that's more marketing than science, so here's a real take on guardrails so that you can make real decisions about y:our machine learning architecture.

If you've read Practical Data Privacy, you'll know I'm not a huge fan of synthetic data.

I've been exploring privacy and security aspects of AI-assisted coding and also experimenting with those workflows for my own work.

AI usage (both intended and not) is increasing in our products, software and lives.

You've learned in the last few articles about how differential privacy works and some of the common pitfalls of actually using it in deep learning scenarios.

If you're just getting started with local AI and local-first development, one of the initial hurdles will be getting your data local.

So, you wanna do at-home AI?

In the last article in the series on addressing the problems of memorization in deep learning and AI, you learned about differential privacy and how to apply it to deep learning/AI systems.

Differential privacy influenced both privacy attacks and defenses you've investigated in this series on AI/ML memorization.

In the past articles, you've been exploring the field of machine unlearning, investigating if you can surgically remove memorized or learned data from models without retraining them from scratch or from an earlier checkpoint.

Building on our understanding of machine unlearning and its varied definitions, in this article you'll learn common approaches to implementing unlearning.

Machine unlearning sounds pretty cool.

It seems like every week my LinkedIn feed is filled with new just released AI risk taxonomies, threat models or AI governance handbooks.

You've probably at some point heard the term "guardrails" when talking about security or safety in AI systems like LLMs or multi-modal models (i.e. models that include and produce multiple modalities, like speech and image, videos, image and text).

One common way to control memorization in today's deep learning systems is to fix the problem by building software around it.

In this article series, you've been able to investigate memorization in AI/deep learning systems -- often via interesting attack vectors.

I'm excited to open-source a project that I've been using for the past 2 and a half years: a private/personal reader and recommender.

In this article, the last in the problem exploration section of the series, you'll explore adversarial machine learning - or how to trick a deep learning system.

At this point in reading the article series on AI/ML memorization you might be wondering, how did the field get so far without addressing the memorization problem?

You've heard claims that we will "run out of data" to train AI systems.

So far in this series on memorization in deep learning, you've learned how massively repeated text and images incentivize training data memorization, but that's not the only training data that machine learning models memorize.

In this article in the deep learning memorization series, you'll learn how one part of memorization happens -- highly repeated data from the "head" of the long-tailed distribution.

In this article in the series on machine learning memorization, you'll dive deeper into how typical machine learning training and evaluation happens, a crucial step in ensuring the machine learning model actually "learns" something.

Hello!

I recently had the wonderful experience of keynoting PyData Paris, thanks again for the invite!

In this series, you've learned a bit about how data is collected for machine learning, but what happens next?

You probably are already aware that many machine learning datasets come from scraped internet data.

When's the last time that ChatGPT parroted someone else's words to you?

Want to learn more about how, when and why machine learning, particularly deep learning systems memorize data?

Building a newsletter is a fairly common activity these days, with many creators, writers and thinkers making part of their living via subscribers willing to give small amounts of money out per year or month to get exclusive access.

After months of searching, lots of fun (and some less fun) interviews and hours of self-reflection, I am excited to announce I am the new Head of Product at Dropout Labs!

Hello!

When I first was introduced to the idea of adversarial learning for security purposes by Clarence Chio's 2016 DEF CON talk and his related open-source library deep-pwning, I immediately started wondering about applications of the field to both make robust…

I presented a keynote at PyData Warsaw on moving toward interpretable reliable models.

Unless you have been avoiding all news, you have likely heard of the coming changes in European privacy regulations which go into effect in May 2018.

I was invited to give a talk at 404 Dublin, a really cool conference joining community groups w/ tech folks and art installations.

Finding ways to determine fake news from real news is a challenge most Natural Language Processing folks I meet and chat with want to solve.

This past Wednesday, I had the pleasure of giving a keynote at EuroPython 2017.

I gave a long and opinionated tutorial at EuroPython 2017 about how we should do unit testing and validation within a data science scope.

Practical Data Cleaning Resources (O'Reilly Live Online Training) This week I will be giving my first O'Reilly Live Online Training via the Safari platform.

I was kindly asked by the PyData Amsterdam organizers to keynote the conference.

The saddest moment for me at conferences is when I'm in the middle of an interesting conversation with a bright person and I ask her when her talk is and she says, "Who me?"

Over the new year holiday time I had a chance to get away from it all, and snuck up to Finland to sit in a lodge on the Gulf of Finland, sip coffee, take saunas and read.

I'm really excited to announce a new Python video course with O'Reilly on data pipelines.

PyConCZ 2016 was such a fun conference!

I gave a talk about Directed Acyclic Graphs (DAGs) and Dask at PyConCZ 2016.

I participated in a two-day PyDataBerlin Hackathon event in early-October and decided to build a scraper for European Parliament.

I dunno about you, but I often find myself bored with NLP (natural language processing) datasets.