Category: AI

Challenging the Limits of AI with ARC-AGI-3 Tests | August 17 2026, 04:51

Today I came across ARC-AGI-3 tests for AI models. That is, if you think that AI can do anything for you, consider that you have already earned a large part of the $850,000 prize fund – just submit your beautiful code, and let it be the best. But it doesn’t work that way. Humans handle 100% of tasks, while programs have not yet surpassed the 2.5-3% mark. What are those tasks?

The program is given not a textual task but a mini-game. There are no instructions or descriptions of the rules. At each step, the program receives a “frame” — a JSON object with the current state of the field. It is a grid of up to 64×64 cells, where the cells are encoded with numbers from 0 to 15 (representing different colors or types of blocks). The agent has a standardized interface (actions 1–7), including basic steps and the ability to specify specific coordinates (X, Y). But AI does not know beforehand what these actions do! In one task “Action 1” might shift a block to the right, in another repaint it, in a third turn on gravity. To understand the logic, AI must actively “poke with a stick” at the environment. It performs an experimental action, the environment reacts and returns a new frame. The agent must analyze the changes, update its theory about the physics of this particular world, and make the next deliberate step. The game can end with the status WIN, GAME_OVER, or continue further.

Again – humans score 100%, machines – 2-3%. You can try it at the ARC PRIZE website, there are sample tasks available.

Redefining Recruitment: AI Agents as the Future of Resumes | August 11 2026, 20:19

I have published a new article about how the interaction between companies and candidates for positions within these companies may look. The main thesis is that we’ve sat too long with these PDF resumes. It feels like a greeting from the last century. There is an interesting trend – essentially, interactive resumes in the form of AI agents for candidates and AI agents for companies, communicating with each other 24/7. Agents in the sense that celebrities and athletes have them. I was quite surprised that there is no movement in this direction, even though technologically, it seems, the required level has been reached.

https://hybrismart.com/2026/08/11/the-cv-as-an-ai-artifact/

https://hybrismart.com/2026/08/11/the-cv-as-an-ai-artifact/

Exploring the Quirks of Whisper Large-V3’s Russian Decoding | July 26 2026, 18:49

It’s funny that the Whisper Large-V3 (OpenAI) voice decryption model, when working with Russian, occasionally replaces a random fragment of what it heard with one or several consecutive phrases “The point is that all speeches today speak of the most monstrous disgrace.”

Apparently, it’s also funny that many don’t care about the outcome of the transcription.

At least a thousand examples are found by Yandex and Google

Automating Disk Analysis with AI: Towards Smarter File Management | July 23 2026, 12:24

I contemplate creating a small program that would scan the disk and analyze all folders and files, peering inside the files only if necessary – primarily focusing on their names, and using LLM to tag and briefly describe entire hierarchies, and more importantly, classify, and save the size of files in the hierarchy. It’s assumed that such a tool runs in the evening and the results are discovered in the morning, because naturally all this is slow. But in the morning, I receive a detailed, structured, relatively small comprehensive report on what needs to be backed up, what can probably be deleted (for instance, a clone of some repository from the public domain), sorted by impact (freeing up space, less clutter, etc.).

Of course, throwing tens of thousands of files at an LLM – is an unreasonable waste of tokens. An agency scheme is assumed – the system takes the top level, makes assumptions, checks them, looks at the time files were modified, checks if there have been changes since the creation of the entire hierarchy and asks the LLM very compactly, setting a percentage of confidence in the judgement about the necessity/unnecessity of any given folder. Then the user can manually check something, or ask to take a closer look at something else.

I wonder, would such a thing be useful to anyone else?

The Future of Fiction: Satire and Automation in “The Silver Eggheads” | July 20 2026, 16:48

I came across the book “The Silver Eggheads” (1961) by Fritz Leiber—a satirical novel about the future of literature, featuring “wordmills” and, darn it, so many parallels with what is being done now.

In the novel, wordmills are enormous machines that have completely replaced writers in the production of fiction. They mechanically “grind” words, producing an endless stream of novels—a so-called “wordwooze”, a product perfectly tailored to the tastes of the mass reader. It causes almost narcotic pleasure: the reader swallows the book and immediately forgets its content, remaining hungry for the next one.

Human writers in the novel still exist, but purely decoratively. They don’t write a single line—they only “service” the wordmills (by pressing buttons) and maintain a public image: they wear bohemian outfits, live in attics, stir up scandals, start affairs—all that the public expects from a “real writer.” Writing has turned into show business and cosplay, while creativity itself is left to machines.

Writers, imagining that they truly can write, rise up and destroy the wordmills. And then a catastrophe is revealed: none of them can compose anything readable—the skill has been lost for generations. The publishing industry panics, searching for a solution.

That’s when the titular Silver Eggheads come into play—preserved human brains in egg-shaped metal containers, brains of writers and thinkers of the past, stored for decades in the “Wisdom of Ages.” The publishers hope that these brains will be able to write books again. The Eggheads, however, having become rather snide and peculiar entities over years of pure contemplation, don’t rush to cooperate.

The plot then turns into a whirlpool of intrigues around these thirty brains—they suddenly become the only “productive asset” in the entire publishing industry.

Meanwhile, the main trouble becomes clear: the Eggheads do not want to write. Two hundred years of pure thought have turned them into haughty contemplators—to them, human novels are trivial and petty. The first attempts to make them write yield either mockeries or texts so strange that they are unsellable. The characters eventually realize that the Eggheads need not orders but motivation: experiences, sensations, participation in life.

Funny book, I should read it sometime.

Innovative DIY Program for Live Transcription and Screen Capture Analysis | June 18 2026, 04:47

I made a really cool thing for myself. I launch a program, it turns on the microphone and listens. I switch to, say, a browser, comment on what I see on the screen, periodically pressing a hotkey to take a screenshot. Meanwhile, my program makes a time-stamped transcript of my comments, saves the screenshots with time stamps, then it recognizes the screenshots, extracting therefrom the spellings of various words, brands, identifiers, people’s names, so as to then transform the transcript of my speech into correct text. And all this – local models, running on my laptop, which means, absolutely free.

After I finish talking to the computer, I start processing the transcript, which takes the raw transcript and text-recognized screenshots as input and outputs a processed transcript, which now looks presentable (Gemini API is used here). One could even go a step further and automatically cut out fragments from the screenshots that were discussed, and insert them in the text exactly where they were mentioned.

Or here’s another thing I can do: just turn on a video on the speakers and the program immediately makes such a transcript for me. Google on YouTube the video “Angular HttpClient Under The Hood. Design Patterns & Source Code Overview” starting at 3:51 – I just put it on autopilot for a couple of minutes, then stopped my script.

Transforming Image Proportions with Generative AI: Smart Redesign Solutions | June 16 2026, 10:08

I published an article about how to transform images with changing proportions. Using generative AI, of course, because transforming a square into a rectangle can either result in data loss, their extrapolation, or by stretching and compressing the image itself. Here, I describe a method where smart extrapolation is performed. When processing hundreds and thousands of images, this approach is not without errors, but their number is relatively small, and it turns out to be much more advantageous to focus on manually correcting the erroneous ones than to do all the work manually right away.

This is specifically necessary during a redesign, when it turns out that the new design slightly mismatches the old one in size, for instance with banners, and the number of these banners is measured in hundreds and thousands.

Automating Banner Crop/Resize Across Breakpoints with Generative AI

AI Revolutionizing Decision-Making in Sports and Business | June 14 2026, 02:06

Today, I pondered how AI is changing age-old, even centuries-old concepts about how people should make decisions in various situations, especially in sports and probably in business. It’s far more interesting than just automation. It’s more about fixing bugs in how people have long considered something to be correct and true.

For example, in the game of “Go,” it was believed for decades that invading the corner (3-3 point) was crude and premature. AI then proved otherwise: early capture of the corner is efficient, and chasing after “beautiful” shapes loses to pragmatic control over the center. Or consider the famous 37th move by AlphaGo in the match against Lee Sedol, which was very strange: people did not play that move because they thought it was “playing into empty space.” It was first taken for an AI mistake, but then recognized as brilliant (there are plenty of analyses on YT). In esports, OpenAI Five demonstrated that aggressive early buyback of fallen heroes in “Dota,” which people considered a waste of gold, works.

Pure mathematics almost erased the mid-range shot from the NBA: it has an accuracy of about 40-42% and yields ~0.8 points per attempt, while a three-point shot with even 35% accuracy brings 1.05 points per attempt, and clubs have restructured for pure profit. Well, this is not AI, but mathematics and statistics. The under-basket shot (lay-up/dunk) turned out to be statistically the most effective.

In soccer, there’s the xG – expected goals metric; AI debunked shots from 35 meters and from outside the penalty area as ineffective (chance of scoring ~5% and 20% respectively) and ultimately teams patiently bring the ball into the penalty area, where the xG of the shot increases to 15-40%. It turns out, DeepMind had a project with Liverpool, a system advising coaches on corners – TacticAI. Expert assessors in 90% of cases preferred TacticAI’s recommendations over the tactical setups used in practice.

So, interestingly, if this continues, will a team or athlete using more powerful AI have an advantage due to more successful methods than a team that does not have such knowledge? Will AI game methods be so complex that they can’t be “stolen” to another team through outside observation – just like in the case with Go?

Helicopter Installs Anti-Drone System on Moscow Residential Building | June 06 2026, 16:26

I read the news that a “Pantsir” anti-drone system was installed by helicopter onto the roof of a residential high-rise (“House in Sokolniki”) in Moscow. Yes, it’s a full-fledged Pantsir, specifically its anti-drone modification (SMD-E), but I couldn’t resist making this AI photo.

[SKIP]

The Mystery of Tal’s Havana Incident: Chess, Reality, and AI | June 02 2026, 00:50

In one of the chess communities, they posted this photo.

It seems to follow a theme. The Olympics, Havana, Tal really did get hit on the head with a bottle in one of the bars – and he was out of action for several days – but then he returned to the board.

But there are seven obvious differences from reality. The most interesting detail – too many pens and fingers on the right hand. But the most interesting – AI would never correctly portray it, no matter how hard it tries.