58 projects · technology and artificial intelligence

Technology and AI: can a model follow the moving ball?

Watch the swaps, choose the cup and see how different video tests ask different questions. Then explore the technology behind devices, networks and public systems.

All projects ↓

Watch and answer

Video puzzle

Under which cup is the blue ball after the swaps?

This clip is a practice puzzle made for “Watch Carefully”; it is not an item from the benchmark below. It does not autoplay.

Choose an answer

Choose a cup after watching the clip.

A test measures the question it asks

The final answer or the full sequence?

In Video-MME-Logical, giving a direct answer and reproducing an exact sequence are separate tasks. Change the view to keep each sample and its scores in context.

Full official set · accuracy2026

Human level95.9%95.9%

GPT-5.422.7%22.7%

Gemini-3.1 Pro28.6%28.6%

On tasks that ask for the final answer: human level 95.9%, GPT-5.4 22.7%, Gemini-3.1 Pro 28.6%.

The two sets ask different questions and use different models or samples; bars in separate views do not form a ranking. The mini set contains 375 samples, with five examples in each task–difficulty cell. Its results are not directly comparable with the full-set scores.

Devices and standards

A code tells you which test a device passed.

An IP rating describes tested protection from dust and water; it does not by itself tell you what repairs a device allows or how it will fare in every situation. Batteries have physical limits that a label cannot remove.

IP68, IP67, IPX4 on your device — what each one means Why Phone Batteries Die in the Cold How many watts is my charger? 5V/2A, 9V/2A and USB-C decoder V30, U3, A2 on your microSD card — what each one means

Networks and public systems

A connection has more than one stop.

One project follows an attacker from a wind farm, through a private network, to another power plant. Another asks who can stop a vulnerable public service. Read each timeline alongside its sources and limits.

How an attacker reached a power plant through a wind farm Who can stop a public system before the attacker does? The operator stays at the keyboard · SilkParasite

Materials and energy

Software runs inside a physical world.

Data centres need buildings and electricity. Fusion needs materials that keep plasma away from reactor walls. See what each project measures, the period its data cover and the assumptions behind its comparisons.

Two years to build an AI data centre, five for a power plant A Tiny Star, Built by Hand Would Tarnița-Lăpuștești make money in Romania? A test on real market data Before the Melt

A measurement is not a verdict

Question, instrument, limit.

A clip checks what you followed in that sequence. A benchmark checks answers on a described set. A standard describes a test. For any wider conclusion, open the project, check its sources and see what falls outside the measurement.

Watch Carefully: Why Multimodal AI Fails Simple Video Reasoning A history of cinema in eleven machines: why the screen was dark half the time How Marian Rejewski broke Enigma: a working machine and his mathematical attack

Explore the complete collection ↓

Technology and artificial intelligence · 58 projects

The complete collection

All projects about technology and AI, plus closely related work on analogue cinema and public digital systems.

Search for a project or topic

58 projects

↑ Back to the experience

Energy and infrastructure
Eight electric cars over 35,594 km: every charging stop of the EUROCHARGE 2026 tour, on one road
Eight electric cars, four countries, eight timed days. The data of the EUROCHARGE 2026 tour puts every stop on one road: how long it took, what it cost, how much battery was left on arrival. One car spent 3 h 15 min at stations over the tour, another 10 h 3 min.
Systems and tools
How a train steers without a steering wheel
A train wheel is shaped like a cone, and the shape does the steering. Push a wheelset sideways, watch the wave about ten metres long it draws, then test the idea with two cups.
Devices and standards
Phones and teenagers' mood: what the public data show
One survey analysed 372 ways: the link between screens and US high-schoolers' sadness is small on average, but it differs by screen, sex and hours a day. You choose what counts as “screens” and as “mood” and watch the result move. Plus the experiments, school phone bans and the numbers for Romania.
Models and limits
Pax Silica: the US declaration on the artificial-intelligence supply chain, who signed it and what has been done
The full text of the declaration, its 25 signatories on a map through time, the eight layers of the technology chain and what has been announced or launched in each, with every claim tied to its source.
Systems and tools
One sound, five accepted forms
In one example from the public Monsoon collection for Hindi (hi-IN), the same transcript has a 33.3% or 0% error rate depending on which forms are accepted as the reference.
Networks and data
One sale, two keys
How a Price Paid Data transaction connects to a UPRN and an INSPIRE ID, what HM Land Registry's tables mean, and where their promise stops.
Systems and tools
Lift 1,200 tons so a ship can pass
Open the bridge on steam: hold the valve, lift two arms of over 1,200 tons so the ship can pass, then lower them — one minute from 1894, with the machine's figures behind it.
Devices and standards
How many years of your life will your phone take?
Enter your age and daily phone hours: the page shows what your number means across a year and by age 80 — whole days, unread books, laps of the Earth on foot.
Models and limits
A history of cinema in eleven machines: why the screen was dark half the time
Until digital projection, a two-blade shutter kept the screen black for half of every film. Stampfer's disc, Muybridge's horse, the 1895 Lumière programme, the Maltese cross, the hand crank, sound on film, Technicolor and widescreen, each as a working model.
Energy and infrastructure
When Humans Vanish
Tomorrow morning nobody wakes up: eight billion people vanish and the planet keeps everything we built. A scrolling journey from the first hours to a billion years — the power grid, the flooded subway, the animals, the clearing sky, the forest and the last message, Voyager — in Romanian and English.
Models and limits
How Marian Rejewski broke Enigma: a working machine and his mathematical attack
An exact Enigma that decrypts a real German message from 7 July 1941, then Rejewski's attack run on your device: the doubled key, the loops the cables cannot change and the 105,456-row catalogue, built in seconds. Plus the codebreakers' passage through Romania in 1939.
Systems and tools
The space station flies over your house. Here is when to look up
You can see it with the naked eye, like a very bright star. Pick your town for the next passes, to the minute.
Systems and tools
1 November 1928, 5pm — the first broadcast
150 watts, and ten years to 274,314 registered owners.
Systems and tools
585, 750, 925 — what does the stamp on your jewelry mean?
Type 585, 750, 925, 916, 999 or any other mark and instantly see the possible metal, purity, and karat equivalent.
Networks and data
Tyre DOT Code — Read the Manufacturing Date and Age
DOT 2424 is not just “from 2024”: it is a tyre made in a single week, 16–22 June 2024. Type your code for the interval and the age.
Models and limits
The operator stays at the keyboard · SilkParasite
What professional AI-assisted cyberespionage looks like: seven RAT families, on-demand modules and commands carried through Google Drive.
Devices and standards
IP68, IP67, IPX4 on your device — what each one means
Type the IP code on your phone, watch, or speaker — IP68, IP67, IPX4 — and get which test each position passed and what the rating does not guarantee.
Systems and tools
V30, U3, A2 on your microSD card — what each one means
Type what your card says — SDXC, U3, V30, A2 — and get what each mark guarantees and which ones say the same thing.
Devices and standards
How did Romania learn to fly? Vuia, Vlaicu and Coandă
The Montesson hop, the Cotroceni flights and the propeller-less Salon machine. Three lives, four years – Romania on the map of flight.
Energy and infrastructure
Why Phone Batteries Die in the Cold
Phone batteries do not lose electrical charge in the freeze: the electrolyte thickens, internal resistance surges tenfold, and the processor cuts off power to prevent critical voltage collapse.
Networks and data
Open knowledge Romania
Find data, archives and tools for your next question. See who publishes them, what you can reuse and what needs a closer look.
Models and limits
PISA 2025: every country's report card
PISA 2025, across eight themes — performance, socio-economic equity, gender, AI and digital, engagement, climate, family, environment — for 91 countries and economies. Romania has one of the OECD's largest socio-economic gaps (122 points), despite strong family support and above-average AI-literacy teaching. Pick any country and see its own dashboard.
Models and limits
Watch Carefully: Why Multimodal AI Fails Simple Video Reasoning
Humans: 95.9%. GPT-5.4: 22.7%. Gemini-3.1 Pro: 28.6%. Ten simple video puzzles that cutting-edge multimodal AI still cannot solve.
Networks and data
The Invisible Architecture of Romania's Sovereign Cloud
Inside Romania's 136 State Databases, the Sovereign Cloud, SRI UM 0296, and the €1M Citizen Shield.
Networks and data
Fully online. At the counter.
The Online filter returned 613 records. In 502, the Available channels field named a counter and contained no online link.
Systems and tools
Before the Melt
Extending a desktop PC's lifespan prevents 265 kg CO₂e; smelting in a furnace recovers just 25 kg CO₂e.
Models and limits
The Radio Accent of Satellites
5.17 million transmissions prove every satellite has a unique hardware signature. Researchers identified all 66 Iridium spacecraft across a month and rejected a ground SDR spoofer forging valid protocol packets.
Energy and infrastructure
Fifty Beats per Second
The power socket dictates time for millions of microwave and radio clocks. In 2018, a 113 GWh shortfall stole 346 seconds (5m 46s) from 400 million people without causing a single blackout.
Models and limits
Who decides it’s an emergency? | AI, LLMs and Romania’s 112 system
Romania is buying a 112 platform with LLM components. Follow a fictional call and move the boundary between transcription, assistance, classification, priority and dispatch.
Models and limits
Before the Write · Pre-Positioning on Siemens S7 Industrial Controllers
Five U.S. federal agencies warn that threat actors are actively pre-positioning against Siemens S7 industrial PLCs using AI-generated tooling disguised as legitimate monitoring software, preparing for future disruptive write operations rather than mere scanning.
Systems and tools
The Tested Sky · Drone Defense Architecture and the Cochstedt Airport Proving Ground
Germany has inaugurated the Technology Centre for Drone Security at Cochstedt airport: a permanent live-airfield proving ground connecting police, scientists, and industry to test counter-UAS stacks in realistic civil aviation conditions before operational deployment.
Systems and tools
The Message Inside the Message
Even if an intelligence service legally obtains the keys and decrypts every Signal message, users can hide a second conversation inside the cryptographic transcript. An interactive simulation of anamorphic encryption.
Networks and data
The Shape of Traffic. What encrypted packets reveal when the destination is hidden
The Tor network hides destinations and contents, yet encrypted packet streams retain a temporal profile. A peer-reviewed USENIX Security ’26 study shows how target watchlist maintenance becomes radically cheaper via few-shot learning.
Networks and data
The picture is the transmitter
A screen can look still while tiny pixel changes make its video cable radiate data. A safe visual simulation of the TrojPix mechanism.
Systems and tools
Same Missile, Three Names, Depending on Who Launches It
The American system, unchanged since 1963, that gives every missile a three-letter code: where it launches from, what it hits, what kind of weapon it is. 315 active codes, an interactive decoder, and a new letter added in March 2025.
Models and limits
Could AI build a better version of itself?
What recursive self-improvement actually means, what Google DeepMind and OpenAI/Anthropic have shown publicly, and why public evidence does not establish whether a September 2026 rumor about DeepMind is true.
Systems and tools
Decode your coin cell: what CR2032, CR2025, CR2016 mean
Enter CR2032, LR44 or any IEC code for size, chemistry and voltage, with a to-scale drawing and a replacement verdict.
Devices and standards
How many watts is my charger? 5V/2A, 9V/2A and USB-C decoder
Paste the OUTPUT values — slashes, spaces or ⎓ — for every profile in watts, the maximum stamped and the naive sum struck through.
Energy and infrastructure
How an attacker reached a power plant through a wind farm
Reconstructing a Polish energy-sector attack from a wind farm through a private APN to a separate CHP plant and Siemens PLCs. With evidence, uncertainty and questions for Romania.
Networks and data
The Code That Expects Dirt
A QR code keeps three anchors for orientation and adds recovery data for errors. Turn it, stain it, and see what remains readable.
Systems and tools
Encrypted for whom?
A call can be encrypted and still be interceptable when a media server accepts the wrong identity. The WebRTC evidence, without alarmism.
Networks and data
Can it still be repaired?
A repair passes through five gates. Public documents show where the path stays open and where local uncertainty begins.
Networks and data
One control panel. Two businesses.
Follow one shared panel across two operations, then watch one template change reach 15+ tenants.
Systems and tools
The Myth That Built the Camera
Every machine closes one gap and opens the next.
Networks and data
Who can stop a public system before the attacker does?
Rules cover fragments. The service as a whole still lacks an owner who follows its risk from end to end.
Systems and tools
Pocket Beacon · How Phones are Hunted in Warzone Operations
Two invisible paths through which an ordinary phone can reveal military personnel movements.
Models and limits
He Worked for the FSB. Then an AI May Have Read 13,000 Emails
The indictment says Denis Obrezko worked for the FSB, then became deputy director at Yutek-NN. His phone contained a file with LLM summaries of more than 13,000 stolen parliamentary emails.
Models and limits
“God Helped Us, and AI Will Do the Same”
A public dossier based on 57 interviews with 27 former Boko Haram members about the institutional use of frontier AI in combat.
Energy and infrastructure
A Tiny Star, Built by Hand
An interactive guide to fusion energy with a kids mode, field atlas, plant cutaway, plasma-control game, reactor tradeoff game, fuels, materials and caveats.
Devices and standards
Windows Remembered the Machine
A device identifier links an intrusion to a chain of logs and accounts.
Models and limits
Who Thought First?
When AI enters before the verdict, the page shows which traces must be preserved so the source, transformations and human accountability can be reconstructed.
Networks and data
Romania, Turla and Russia’s cyber ecosystem
Romania confirms Turla targeted the country. An interactive dossier on GRU recruitment, credentials stolen through Lumma and sabotage attributed to the FSB.
Systems and tools
One Opened Email, a Whole Server Compromised
The UNK_MassTraction campaign uses Roundcube vulnerabilities to turn the victim's browser into a pivot toward the server, from XSS to memory-only implants.
Systems and tools
One Target, Two Flags
How China- and India-nexus cyberespionage actors reached the same Pakistani police servers and converged on Balochistan Police.
Models and limits
Two years to build an AI data centre, five for a power plant
From the first bulldozer to the first operating building, against the grid connection request: two clocks, measured.
Networks and data
Everything looked legitimate
When an authentic thing is used in a malicious relationship, trust can move from one victim to the next target.
Devices and standards
The machine said no
The chip could not hold the world, so someone invented a trick. We still play that trick.
Models and limits
Would Tarnița-Lăpuștești make money in Romania? A test on real market data
A transparent model on OPCOM day-ahead prices tests how much energy arbitrage Tarnița-Lăpuștești could have captured and what share of annual cost the modelled revenues leave uncovered.

About the collection

This collection brings together relevant projects from the site’s library. Each entry has a title, a description and a link to its available edition. The complete projects carry their sources and evidence limits.

Reviewed 2026-10-15 · Collection as text

Sources

Video-MME-Logical · evaluare și setul mini de 375 de exemple
Primary results source; it separates direct answers from exact state sequences and cautions against overreading cells with five samples.
Kwan et al. · Video-MME-Logical (arXiv:2606.27828)
Description of the original benchmark and temporal-logical tasks.
Privește cu atenție · videoclipul exercițiului
The clip on this page is a local practice puzzle, not a benchmark test item.