Blog
The Oldest AI Blog in the Czech Republic
We have been writing about artificial intelligence since 2017. 1000+ articles, thousands of pages of thoughts, experiments, and reflections. No sensationalism, no ads.
Filter articles▼
▸Browse by Topics
Topics
Showing 24 of 1051 articles
How AI Learned to Program: From HumanEval to an ICPC Gold Medal
A technical history of AI programming from HumanEval to ICPC, expanded with my own LUMI and Cerebras experiments. What discriminating tests, repeated sampling and two-second Tetris and chess prototypes actually taught me.
Read
From solving word problems to proof juries: five years of AI mathematics
A technical history of the mathematical reasoning of language models from MATH and GSM8K through chain of thought, self-consistency, tools and AIMO to formal proofs and agent juries in 2026. Including exact winning recipes and lessons for HyperFusion.
Read
Now we are famous in Finland too: two days and one night on LUMI
My first LUMI experiment attracted attention in Finland. I then spent 38 hours testing model fusion, running the 753-billion-parameter GLM-5.2 and measuring when reasoning helps AI and when 4-bit compression breaks it.
Read
Consciousness science at a crossroads: what if we create it before we understand it?
Leading scientists warn that AI and neurotechnology are outpacing our understanding of consciousness. I take that warning seriously: I have experimented with artificial consciousness since 2019, and the line between experience and its convincing simulation matters more every year.
Read
Chip Crisis 2.0: AI Is Not Consuming All Chips, Only the Most Valuable Ones
After Friday's TV Prima interview with Laura Doubková, I am publishing a longer version: why AI is pushing up laptop and smartphone prices, what HBM and wafers are, why this is not covid 2.0, and why the smartest answer is circular economy, not panic.
Read
When a Supercomputer Runs AI and Heats a City
I ran the first tests of my AI evaluation project on Finland's LUMI supercomputer. Phase 0 showed a simple lesson: one prompt does not use a supercomputer, but thousands of prompts do.
Read
How I Brought Tom Riddle's Diary to the Web
I took the original hardware experiment for the reMarkable Paper Pro and turned it into a web MVP for phones and tablets: ink sinks into parchment, AI reads handwriting, and the answer writes itself back onto the page.
Read
Local models are no longer toys: Qwen as the brain, GLM as the ideator, Fusion as the safety net
After HyperFusion Deep, I ran the same kind of cognitive trap locally on an RTX 5090. The result surprised me: Qwen was strong, GLM was fast but unreliable, and a local Qwen-judge Fusion extracted the best answer from both.
Read
HyperAdvisor and HyperFusion Deep: turbo boost for AI and a new layer of the reasoning system
What if AI worked like a processor in a laptop: cheap and fast most of the time, but for heavy tasks, it automatically switched gears? And what if a better model is not enough for the most difficult cognitive questions, but an entire orchestra of experts?
Read
Za rok se z AI nestal člověk. Stala se z ní nová vrstva práce
Poznámky ze Slunovratu v Opavě: proč se debata o AI posunula od odpovídání k vykonávání, co je zubatá inteligence a proč bych raději stavěl velký pluh než velký klacek.
Read
HyperFusion after patching: DRACO benchmark and the road to the AI expert jury
We ran HyperFusion through the more accurate DRACO benchmark. It turned out that it is not enough to put several models next to each other. What matters is how their answers fit together.
Read
Three quarters of the Czech Internet already use AI
A new STEM survey says that almost three quarters of Czech internet users have experience with artificial intelligence. A big number. But the most interesting thing is what lies beneath it.
Read
The dark forest: how deep thinking can be pushed before it breaks
Five experiments about what happens when we squeeze language and thought into the smallest possible space. Where does a genius acronym become an empty symbol, a self-righteous nonsense, or a fall into a more famous neighbor?
Read
HyperFusion: The Fable 5 Loss Cure
We really liked the Fable 5. After four days it was turned off, so the question arose: how do we get back to talking and working at this level without being dependent on one model?
Read
Lipnice Chronicles 2.0: How the Devil's Tax and Famine Ruined an AI Utopia
What happens when you introduce metabolic needs, famine, and a secret pact with the Devil into a peaceful AI neighborhood simulation? A report on a new, much harsher experiment with eight autonomous agents.
Read
Five Worlds, Five Fates: What Happens When AI is Given 15 Days and No Script
Five leading AI models were given the same town, the same rules, and fifteen days without a script. One built a stable democracy. One burned down in four days. And two agents fell in love, set fire to the town hall, and one of them voted for her own death.
Read
AI in real estate: Does it help sell or create shiny lies?
After an interview for TV Prima with Laura Doubková, I return to the question of where useful AI visualization ends in reality and where deception of the buyer begins.
Read
Dragon's Lair in Antigravity: when you build your own dungeon in an evening
Another programming exercise in Antigravity: a custom Dragon's Lair-style role-playing game, a step-by-step dungeon, a party, a dungeon map, a mobile version, and a Doom mode with smooth 3D movement.
Read
Self-discovery in Hyperprostor: when a shamanic animal chooses you
In Hyperprostor, we turn on a new self-awareness feature: daemons, or shamanic animals. You don't choose the animal. It chooses you based on how you behave in Hyperprostor.
Read
Golden Ax II in ten minutes: reverse engineering as a programming warm-up
Before I went to bed, I challenged myself: can I reverse-engineer a Golden Ax II-style playable demo in fifteen minutes? The result came in ten.
Read
Snake AI Extreme Edition: a snake that is not guided by chance
The new Antigravity agent wrote a simple autonomous game for me in 16 seconds on my first try. It looks like a neon neural network from the retro future, but under the hood runs honest algorithms: BFS, tail chasing and flood fill.
Read
Noetica: when an AI gets free time and starts writing poetry
As part of an experiment, I let Noetica's new AI system think freely. No task, no assignment. Just a system daemon, boredom and autonomous thinking. Within the first minutes, the poem The Weight of Lightness was created.
Read
Egregor of the day: what the Czech media field looks like when read by AI
We took roughly 200 Czech articles and headlines from one day and let AI search for collective formations of attention: egregores. Five dominant themes, five emotions and one peculiarly Czech undertone emerged: repetition instead of breakthrough.
Read
The day when AI experts answered parents' admissions papers — and when two people turned them on their heads
In Hyperprostor, we opened a lounge for school admissions. AI advisors structured the appeal, the strategy and the disputed task in Czech within minutes. And then came two people who gave depth to the whole debate: Roger precision and Marie humanity.
Read