Gonzo ML paper reviews
Two practitioners break down fresh machine learning and AI papers — no hype, just what the paper actually does.
What it is
gonzo-ML ("gonzo reviews of ML papers") is a Telegram channel about machine learning and AI research, run by Grigory Sapunov (per the channel description, formerly head of development at Yandex.News and now CTO of Intento) and Alexey Tikhonov (formerly an analyst at Yandex). The core format is a paper review: title, authors, links to the preprint, code and model, then a Russian-language breakdown of what was actually done and why it's interesting. Short finds, links and the odd meme fill the gaps.
Why we picked it
Everyone writes about AI; few people open the papers. This channel starts with them. Recent reviews cover a study on distilling into byte-level models that challenges standard tokenisation and suggests new scaling laws; WeatherNext 3, a global weather model trained on raw observations ("something useful for the national economy, for a change"); and the Atria Dawn report on how agents help researchers improve the agents themselves.
It isn't dry, either. The authors share things they find beautiful, like Epoch AI's piece on the plunging price of "thought" from language models, recruit help for CayleyPy, a project using LLMs to push mathematics forward, and run a string of memes about a digitised fruit-fly brain writing code. The result is a lively feed for anyone who wants to understand AI research beyond the headlines.
Where to start
- Any post with a "ЧТО сделали" ("What they did") block — that's the classic review format.
- The WeatherNext 3 review, if applied problems interest you more than architectures.
- The byte-model distillation post, for anyone following LLMs and tokenisation.
- The short link posts between reviews — a good way to find more to read.
- Short on time? The first paragraph of each review usually carries the gist.
Who it's for
ML engineers, researchers, students and developers who read Russian and want to follow AI through the papers. Some background in machine learning helps; this is an AI research channel for people happy to read about architectures and experiments.
Quick answers
What language is it in? Reviews are in Russian; paper titles and author lists stay in English, as in the original.
How often does it post? Going by the public preview, almost daily, with reviews mixed in among shorter notes.
Do I need a Telegram account? No. The channel's web preview is readable in a browser without signing up.
