Recordings

I have given dozens of talks about new papers over the years but only couple of more general recorded talks, here they are:

Machine Learning in Recommender Systems (2018, EN)

Joint talk given with Pavel Kordík and Ivan Povalyev about uses of ML in recommender systems.

2018 was not that far from the Inception moment in 2013 and most models were still pure classical ML
My section is about how to use Deep Learning (yes DL was the hype :D) in Recommenders and starts here: https://youtu.be/_YR3Osnl_Dc?t=1620

So if you want, feel free to take a jump into the past and see:

my work on Online optimization of Hyperparams using Evolutionary Strategy with surrogate modeling by Gaussian Mixtures
- starts at https://youtu.be/_YR3Osnl_Dc?t=2905
- you can read more in my thesis: Optimization of Recommender Systems
Feed-forward Nets
RNNs - GRU, LSTMs
combining embeddings in Sparse Denoising Auto-Encoders
TSNE projections of Inception v3 embeddings to blow the audience mind :D
RL examples like Bandits (HybridLinUCB)
Offline evaluation with the fancy techniques like Doubly Robust Off-Policy value estimation

When I am watching this in 2025 large part of it is actually still state of the art :D

the online optimization algo is still running in production
SASRec is just GRU4Rec with a transformer block
offline evals are still hard to use unless you have part of your traffic running on random sampling

General overview of how to build a recommender.

Does not go deep into any part but covers:

I talk about my views on the “Future of Programming” regarding recent AI advances:

why programmers should strive to understand business more:
- speeding up companies still follows Amdahl’s law = you do not get massive gains if you don’t remove the bottlenecks
- the bottlenecks are typically meetings that delay the execution loop
- => the way to get speed ups from better AI for coding is to empower programmers to do more iterations by themselves
- => programmers have to understand business goals of the project to be able to iterate without getting sign offs from management
- in other words it does not matter whether you implement the feature in 5 minutes or 5 hours if you still have to wait till next day to get the next steps signed off by you product manager
why does it still makes sense to hire Juniors even if we have Cursor:
- because they have AGI + are capable of continuous learning unlike LLMs
and more like how we use LibreChat, n8n, MCP servers etc.

high level overview of the AI Act draft
the actual AI act in effect now has some differences and it’s not clear now (in 2025) how exactly will the implementation look like, so I would recommend newer sources instead if you’re interested