← Journal · · Note
AI Conf 2024: talks worth remembering

Hey everyone! On Thursday and Friday I was at AI Conf by Ontico, a conference about applied ML. It was great, the venue was top-notch, in the center of Moscow, and the talks were also great! Decided to write down a few that stuck with me.
- Beyond the script: continuous testing for adaptive and safe LLM systems https://aiconf.ru/2024/abstracts/12981
Evidently AI post a ton of useful ML and MLOps info on their site, and here they even showed up with a talk. It was interesting to hear how to test LLM applications when people try to pull them out of the corporate context :)
-
Masterclass "Getting high-quality data for training AI models" https://aiconf.ru/2024/abstracts/id5647224 Went to a masterclass on crowdsourcing. Turns out Yandex has a cool tool, Yandex.Zadaniya + Toloka, which lets you automate the data labeling process. You create labeling tasks via API, real people see them and start labeling datasets, and then you can take that and fine-tune your model. We tried it on YoloV, got interesting results. What struck me most was how fast people started completing tasks after publishing them :)
-
Labeling the future: how GPT helps train models https://aiconf.ru/2024/abstracts/id5602978 As a continuation of the masterclass, they showed an interesting way to use GPT for the same labeling operations. Overall the results are good.
-
How "Pisets" went to "Total Dictation", or a few words on robust speech recognition https://aiconf.ru/2024/abstracts/12975 A story about fine-tuning the Whisper model to improve Russian transcription. They trained it on podcasts :) And also open-sourced it https://huggingface.co/bond005/whisper-large-v3-ru-podlodka
-
Evolution of the Transformer: how the most successful architecture in DL is changing https://aiconf.ru/2024/abstracts/id5523026 A big history of the transformer architecture. Interesting to hear about its evolution :)
Looks like you can even check out the presentation here https://mellain.github.io/data/AI_Conf_Transformer_Arcitecture_Full_Slides_2024.pdf
Maybe at some point they'll publish the talks publicly, then you'll be able to watch the videos too :)