@psimecek.bsky.social
Marie Skłodowska-Curie Fellow. Stupid tweets are my own. All the glory belongs to CEITEC and Mediaboard.
Czech contribution to AI discussion
06.08.2025 13:33 — 👍 0 🔁 0 💬 0 📌 0
Was excited about OpenAI's new open-source models until I hit "trained on a mostly English, text-only dataset" For Czech gpt-oss-20b is genuinely bad, gpt-oss-120b is passable but far from impressive. Running formal benchmarks, but casual testing already tells the story 🇨🇿
openai.com/index/introd...
This is actually pretty good illustration of AI-assisted coding success / failure (ignore the term "vibe coding" and be honest - we have all been on the 4th line)
Source: www.reddit.com/r/vibecoding...
- Co vy na té fildě vlastně děláte.
= My učíme lidi číst a psát.
Mne AI extrémně zajímá jako někdo, kdo je schopný dělat kvalitativní analýzu srovnatelnou s lidským kodérem. Protože v humanitních a sociálních vědách... máme velmi dobře rozpracované teorie, jak dělat obsahovou analýzu, jenom to znamená, že máš 2-3 lidi v týmu, kteří podle (manuálu) dělají kódování
06.05.2025 17:52 — 👍 0 🔁 0 💬 1 📌 0
Je strašně málo tlustejch šachistů.
(Josef Šlerka v open.spotify.com/episode/5SBB...)
Ok, this is crazy. Geoguessing photos with AI:
* Gemini 2.5 Pro seems to be the best (crazy how fast it is!)
* OpenAI o4-mini-high (on a screenshot) - not bad!
* OpenAI o3 - very slow
* Grok - is guessing Lisbon, Portugal (=wrong)
I believe if anyone showed us GPT4.5 10 years ago, we would call it AGI.
(Standa Fort is killing it at #MLPrague2025)
I had no idea that one of LLM innovations was invented for write better erotic stories. (arxiv.org/abs/2306.15595, enabling LLMs to work with longer texts)
What you learn at #mlprague2025!
Jon McLoone (Wolfram Research): People are asking AI chatbot wrong questions in the same way they were asking Google wrong questions 20 years ago. It is possible they will learn to do better.
#MLPrague2025
From questions to Jon McLoone's talk at ML Prague "Lies, Damn Lies and Gen AI"
Sunken WWII shipwrecks are valuable because they are source of low-background steel (en.wikipedia.org/wiki/Low-bac...). Maybe similarly, one day we will be seeking for pre-AI texts.
#MLPrague2025
Stručný návod na home office ve Vietnamu
(jen si sem schovávám) 🇨🇿
vikend.hn.cz/c1-67713380-... 🔒
When I submitted my "Bullshit AI" talk to PyConAT, I had no idea Trump would end up giving me a perfect example. It is worth trillions of dollars. Literally.
www.linkedin.com/posts/bastia...
Autora neznám
05.04.2025 08:32 — 👍 1 🔁 0 💬 0 📌 0
However,...
03.04.2025 06:51 — 👍 0 🔁 0 💬 0 📌 0
Ondřej Svoboda má na LinkedInu pěkné ukázky, že 4o-image zvládne i české styly
www.linkedin.com/posts/ondrej...
New models in MiniCzechBenchmark 🇨🇿
Gemini 2.5 Pro 🥇 (as expected), Gemma 12B a very pleasant surprise 🫢 (you can run it in ollama on your 💻), only marginal improvements for new DeepSeek V3, Mistral Small, openai gpt-4o
github.com/simecek/Mini...
Despite popular belief it is not true that current LLMs can solve math. olympiad problems.
(I have a set of Czech middle school problems Klokánek, most LLMs are slightly above random chance, best thinking at ~90%)
arxiv.org/abs/2503.219...
Draw me picture with two schemas explaining the difference between AlphaFold2 and ESMFold
(needed extra hint)
My photo is attached. Try to add glasses to it (I am about to buy a new glasses and need to pick the style)
30.03.2025 21:22 — 👍 0 🔁 0 💬 1 📌 0
Draw me a little map of XYZ for my future visit. Not real map, just illustration. Include top places I should visit. Please, omit typical turist traps and concentrate on cheap free hipster punk things #gpt4o
30.03.2025 21:13 — 👍 0 🔁 0 💬 2 📌 0
If you have ChatGPT Plus (paid $20 version), try new image generation introduced yesterday. Seriously crazy.
This is me as Simpsons character (no special model, just prompt)
openai.com/index/introd...
LLM image editing solved by Google (model: Gemini Flash 2.0 Image Generation Experimental). Feels like magic - it understands the image on pixel level.
Try it: huggingface.co/spaces/phils...
Or just use Google AI studio with this model
aistudio.google.com
The poem is from GPT4.5. Too expensive to be benchmarked (cca 700 CZK) but the first model to properly rhyme in Czech.
10.03.2025 10:45 — 👍 0 🔁 0 💬 0 📌 0
MiniCzechBenchmark 🇨🇿 Update including QwQ-32B, Claude 3.7, Gemini 2.0, and DeepSeek-R1!
* Gemini 2.0 Flash is cost effective, fast and great
* 🇨🇳 open models to pay attention: V3, QwQ-32B, R1
* Mistral-Small-24B-Instruct-2501 = open alternative to haiku / gpt-4o-mini
github.com/simecek/Mini...
This is seriously crazy! Promting multimodal LLM to play games IN REAL TIME
x.com/haoailab/sta...
This is the repo github.com/lmgame-org/G...
A this is prompt responsible for tetris
github.com/lmgame-org/G...
GeminiFlash is fast enough to play SuperMario (on par with Sonnet 3.5)
Často Olšany v prvním průchodu zamítnou (protože ona to tak úplně není pražská čtvrť, že jo). Později se vrátí a řeknou si "Aha, tak to asi budou ty Olšany!" A to mi na tom právě přijde zajímavé.
28.02.2025 20:19 — 👍 0 🔁 0 💬 0 📌 0
Test LLM šolynou. Hlubší než si myslíte. Většina thinking modelů to dá (ale ne DeepSeek-R1, ani QWEN)
28.02.2025 20:17 — 👍 1 🔁 0 💬 1 📌 0Actually, it seems they just pushed it out because they could not make it better. There is a lot of negativity on Twitter, GPT4.5 is expensive but worse on benchmarks than Claude 3.7, Grok3 or even DeepSeek. Not a good day for OpenAI, so they put engineers into spotlight and not Sama. Typical.
28.02.2025 19:22 — 👍 1 🔁 0 💬 0 📌 0ChatGPT4.5 is here
www.youtube.com/watch?v=cfRY...
I love that they let the model to be introduced by engineers (with all the imperfections of normal people in front of the camera)
