Mariia previously worked as a Data Scientist at Lyft, where she developed ML solutions. After encountering research in AI Alignment, she decided to focus on studying the behavior of large language models. She is currently conducting research as part of a Research Fellowship at LASR Labs, focused on behavioral steering of LLMs and how model values form during reinforcement learning (RL).
When
August 20, 6:00 PM
Format
Offline in Lviv
Registration deadline
August 19
Fee
Donation from 500 UAH to the NGO "Reactive Post"
About:
At the Align, Forget, Repeat Meetup, we’ll explore the latest research in AI Alignment and Machine Unlearning. We’ll discuss why traditional LLM training approaches, including Reinforcement Learning (RL) and Reinforcement Learning from Human Feedback (RLHF), do not necessarily guarantee alignment, what kinds of undesirable behaviours large language models can exhibit, and which methods help make them safer, more reliable, and more controllable.
We’ll also examine whether LLMs can “forget” unwanted information, how the effectiveness of machine unlearning can be evaluated, and the AI Safety challenges researchers are actively working on today.
Who is it for:
- ML/AI Engineers
- Data Scientists
- AI Researchers
- PhD students
- Software Engineers integrating LLMs into products
- AI enthusiasts
To participate:
- Fill in the registration form
- After our confirmation, donate from 500 UAH to the NGO "Reactive Post"
What you'll gain:
- How AI Alignment is studied today and why it is essential for the future of large language models.
- Real-world examples of undesirable model behaviours—from sycophancy to alignment faking—and why they occur.
- The current state of Machine Unlearning, when models need to "forget," and how researchers evaluate the success of this process.
- Opportunities to network with AI Alignment and AI Safety researchers.
- An overview of today's key AI Safety challenges and emerging research directions.
Speakers:
Viktoriia Makovska
Research Engineer, Lapa LLM
PhD Researcher, Ukrainian Catholic University (UCU)
Victoria has over 7 years of experience in software engineering, after which she moved into research on large language models and responsible AI. She currently works on machine unlearning, mechanistic interpretability, and methods for evaluating the effects of “unlearning” — whether a model has truly stopped reproducing unwanted narratives or behavioral patterns, rather than simply losing its general capabilities as a result of the intervention.
Program:
18:40 – 19:10
Maria Koroliuk
How the behaviour of language models emerges, why alignment does not happen automatically, and how researchers work with the internal representations of LLMs to develop safer models.
19:30 – 20:00
Viktoriia Makovska
An introduction to Machine Unlearning, whether knowledge can be removed from a model without compromising its capabilities, and why this has become one of the most important research areas in AI Safety.
FAQ
-
How to participate?
Fill out the form and wait for a confirmation email. Check your spam folder in case it ends up there.
-
Is participation paid?
Yes and no. To join, you need to donate from 500 UAH to the charity organization “Reactive Post.” The donation should be sent after your participation is confirmed by email.
-
When to send the donation?
Once your application has been reviewed and you’ve received confirmation that you’ve been accepted.
-
Can I participate remotely?
No, the event is offline only in Lviv. There will be no recording or live stream.
-
Is there a shelter at the venue?
Yes, the venue has a shelter.
-
When does registration end?
Registration runs to December 15 inclusive.
-
What is the language of the event?
Ukrainian.