Jev Cards: Can You Beat a Vision Model at a Picture Quiz? (Show HN)
What Is Jev Cards?
Jev Cards is a browser-based picture quiz where a human player competes against a vision model named Jev 1.14. The game presents 490 picture cards covering themes such as fruit, flags, maps, chess, and illusions. Jev sees the same picture as the human and returns an answer in roughly a quarter of a second. The site's tagline asks, "Can you beat Jev?" and at the time of the snapshot, it reported that Jev had won 2,463 of 3,103 games. The page also advertises "Live tonight: people against a model" and a "Sprint cards Versus seconds" mode.
How Jev 1.14 Answers Picture Cards
Jev 1.14 is a decision model rather than a text-generating model. It reads the picture and the question and returns a probability for every option—or a box, for "tap the place" cards—in one call. It never writes text. This design means the model does not produce explanations or chain-of-thought; it directly outputs a probability distribution over the possible answers. The reported response time is about 0.25 seconds per card.
Card Themes and Game Modes
The game includes 490 picture cards across multiple themes. The listed themes are fruit, flags, maps, chess, and illusions. Commenters also encountered music theory questions. The "Sprint cards Versus seconds" mode measures how quickly a player can answer a set of cards. The core challenge is "You vs Jev 1.14," where the human and the model see the same card and race to answer.
Player Feedback on Scoring
Players reported that the scoring system can produce counterintuitive results. One commenter, geooff_, wrote:
"I 'Lost' to Jev while getting 10/10 questions and it got 8/10. Interesting scoring?"
This suggests that the win condition is not purely based on accuracy. Another commenter, jrflo, described a speed-accuracy tradeoff:
"Very cool demo! Although I feel that incorrect answers should be penalized more. For instance- I got 100% in about 45 seconds on the sprint. Then, I clicked through as fast as I can (only limited by finding the right targets) and got a score of 42s, technically better than my first score but I only got 20%."
In this case, a faster time with much lower accuracy was scored as better than a slower time with perfect accuracy. This indicates the scoring formula weights speed heavily, or at least does not sufficiently penalize incorrect answers.
Player Feedback on Accuracy and Speed
Several commenters observed that Jev is faster but not always more accurate. edaemon wrote:
"Interesting. Jev is certainly faster but it was wrong quite a bit more than I was."
inerte similarly noted:
"I got a lot of answers wrong (half of them maybe?), and the ones I got it right Jev was still faster."
casey2 reported a specific game where the human answered 5 questions correctly while Jev got 0 in a row, yet the app declared Jev the winner:
"I thought I did terrible 5, but jev got 0 in a row. then the slopcoded app says 'Jev has won 2,613 of 3,300 games, this one included.'"
These comments highlight that Jev's speed advantage can decide the outcome even when its accuracy is lower than the human's. The reported tally in that comment (2,613 of 3,300) differs from the snapshot's earlier figure (2,463 of 3,103), suggesting the game's win counter updates over time.
UI and Content Feedback
Players also gave feedback on the interface and card content. nextaccountic asked for a way to exclude music theory questions:
"Could you add an option to remove musical theory questions? I mean, I know Jev can ace those as well, but I can only click at random, that's unfun."
arjie suggested a layout change:
"Would prefer if the audio button was at the top right of the page etc."
semiquaver reported a bug where a card showed only one option, and the app marked the player wrong even when they believed they had chosen the correct answer:
"The first one I opened only gave me a single option to press and when I pressed it the app told me I was wrong and that Jev had pressed the correct button (which was invisible to me I guess) I saw several other questions where I pressed the right answer and the app scored me as wrong. Broken slop."
coolThingsFirst noted a surprising fact in the card content:
"Wait a minute Orca is a Dolphin. Damn."
What Jev Cards Demonstrates About Vision Models
The Jev Cards experiment shows that a vision decision model can answer picture quiz questions in about 0.25 seconds and win most games against humans. The value of such a model is highest when tasks require very fast answers and do not demand perfect correctness. casey2 summarized this tradeoff:
"if you need very quick answers obviously humans can't compete. Actually on further thought after getting 15, if you've seen the exact problem before you can answer quickly, so the usefulness is diminished to problems that don't need to be 100% correct, need to be solved very quickly (even routing might not need to be this quick), problem form often changes, but is still general knowledge, and very low marginal benifet for 99% correctness over 50%."
This suggests that speed-focused vision models are best suited for scenarios where the problem format varies, general knowledge is sufficient, and the marginal benefit of near-perfect accuracy over moderate accuracy is low. Examples might include quick routing decisions or interactive games, though the commenter cautions that even routing may not need such extreme speed.
Summary of the Show HN Discussion
Jev Cards is a live experiment in human-versus-model picture quizzing. Jev 1.14 is a decision model that returns probabilities in one call and never writes text. The site reported Jev had won 2,463 of 3,103 games at the time of the snapshot. Player comments focused on scoring that favors speed over accuracy, occasional incorrect scoring and card bugs, and requests for content filtering and UI adjustments. The discussion also raised broader questions about when fast, moderately accurate vision models are more useful than slower, highly accurate ones.