Hai Duong "Čan" Tran

AI Researcher; ex-@seznam_cz

[Hai Duong "Čan" Tran @PhoBoAI]

@KulaRostislav @z_nohy @ondrejnovotny OCR

[Hai Duong "Čan" Tran @PhoBoAI]

@milanfon @pesvklobouku To, ze to nenechame dojit do stavu, kdy je to opravdu overfitle je uz vec druha. Ale kdyz se to udela nesikovne, tak se nejake pasaze z autorskeho textu proste daji do jiste miry zreplikovat.

[Hai Duong "Čan" Tran @PhoBoAI]

@milanfon @pesvklobouku Asi slovickarim, ale u (L)LMek se optimalizuji parametry modelu, aby se minimalizovala cross entropy loss. Jinymi slovy hledame takove parametry, abychom co nejpravdepodobneji umeli doplnovat predesly text videny v treninku.

[Hai Duong "Čan" Tran @PhoBoAI]

@milanfon @pesvklobouku My to v Seznamu respektovali. AI Act v EU to taky tusim komplikuje, protoze se musi loggovat zdroj dat.

[Hai Duong "Čan" Tran @PhoBoAI]

@milanfon @pesvklobouku Fair use si to jiste miry asi obhajili, ale zpusob jakym ta dila dostali uz koser nebyl. Tusim, ze vetsina tech webu na screenu ma v podminkach/robots.txt zakaz stahovani, nektere i explicitni zakaz pouzivani pro AI. Takovy idnes ma v podminkach zakaz jakehokoli pouziti obsahu.

[Hai Duong "Čan" Tran @PhoBoAI]

@askalphaxiv Or just download the TeX sources if available including all the assets.

[Hai Duong "Čan" Tran @PhoBoAI]

@lukaschadraba @destralux Dokazu si predstavit, ze by to pomahalo prave dodefinovat/docistit vstupni data a pozadavky, ale ne navrhovat a evaluovat ten rozvrh (co je dobry rozvrh?). Co ja vim, tak rozvrharky resi prave ty softove pozadavky, ktere se blbe formalizuji a resi konflikty.

[Hai Duong "Čan" Tran @PhoBoAI]

@lukaschadraba @destralux Vzhledem k cemu by se ten ML mechanismus optimalizoval a jak?

[kalomaze @kalomaze]

the world if tmux scroll was good

[Hai Duong "Čan" Tran @PhoBoAI]

@ununik @bubnikv @durycek @Paralelni_Listy Po vypuknutí války zavřeli všechny pobočky včetně té největší v Petrohradě. Což se týkalo hodně velké části engineeringu. V Česku afair daní veškeré příjmy, takže jde o jednoho z největších daňových poplatníků u nás.

[Gemini Notebook @Gemini_Notebook]

Hola! Vielen Dank fürs Warten. 你期盼的时刻终于到来了.

Audio overviews in 50+ new languages is officially 100% rolled out in beta! We're working to add more languages (patience, young polyglots) but we're excited to make audio overviews more accessible around the world 🌎🌎🌎

https://t.co/GBltX10S8n

[Ameya P. @AmyPrb]

🚨 New paper!

Exciting progress in GRPO variants, smarter training strategies, and curated datasets showing impressive improvements on math reasoning -- Is the hype justified?

Details below👇

[anton @abacaj]

Plz stop making agent demos planning trips or booking flights, how many flights are you even taking??? Make them do repetitive stuff no one wants to do instead

[Hai Duong "Čan" Tran @PhoBoAI]

@RedHat_AI @rodrimora (VllmWorker rank=1 pid=812832) AttributeError: 'Llama4Config' object has no attribute 'vocab_size'

Both vLLM and transformers upgraded to the latest version :(

[wh @nrehiew_]

Cohere's Command A report is an extremely extensive paper on how to train a modern LLM in 2025. But it's a model for very different but specific use cases.

Let's talk about it

[Hai Duong "Čan" Tran @PhoBoAI]

@diegoasua TensorRT-LLM combined with Triton Inference Server ._.

[Mistral AI @MistralAI]

Introducing Mistral Small 3.1.
Multimodal, Apache 2.0, outperforms Gemma 3 and GPT 4o-mini.
https://t.co/BHLAAaKZ9w

[Hai Duong "Čan" Tran @PhoBoAI]

@HKydlicek @distillpub Cool, I was about to yoink the templates from the spaces.

[Klaas @forgebitz]

we are all building the same fcking thing aren't we