Free browser games, freshly slopped by the machines.
FunSlop.com is a growing pile of free, no-install browser games — arcade, strategy, idle, and assorted AI-adjacent nonsense. Pick a tile, hit play, blame the model. No accounts, no downloads, no dignity. New slop drops regularly.
Today's Feed
Generated in 0.002s. Guaranteed mediocre.Attention Budget
You have a finite attention span and a context window to match — highlight only the tokens that actually answer the question.
Spec Writer
Write the policy in literal instruction-blocks; the agent does exactly what you wrote, not what you meant.
Autocomplete Tennis
Volley word-associations with the model. It wants to talk about GPUs; you need it to say 'banana'. Rally until someone whiffs.
Context Collapse
Five system prompts, all active, half contradictory, context window full. Keep everyone happy. You can't. Try anyway.
Prompt Jenga
This prompt works. Now remove words until it doesn't. Pull the wrong token and the whole output comes crashing down.
Token Budget
You have 512 tokens and seven things that each need more than that. Spend wisely — the context window is not your friend.
Perplexity Karaoke
Say real words the model would never predict. Maximize perplexity out loud, in public, for science.
Pitch Lock
The model drifted off-key and off-alignment. Hum it back onto the target curve. Yes, with your actual voice.
Paste-to-Play
Paste your group chat, a poem, the terms of service — we'll turn your text into a level and you into a casualty of it.
Motion Eval
Stand up. The adversarial examples are incoming and your only defense is flailing at your webcam. Cutting-edge evaluation.
Copilot
The AI offers to finish your sentences — sometimes right, sometimes confident slop. Accept wisely; ship fast.
Be Average
Think like the average internet user (god help you). Name the answer everyone names. Maximize alignment with the mode.
Most Forgettable
Be unremarkable. Generate the beige-est filler imaginable. If the model can recall it later, you've failed at mediocrity.
Predict Yourself
First you choose. Then you predict what you chose. The model has a self-model — do you?
Rule Roulette
The objective just changed. And again. Distribution shift in real time — adapt or overfit to a world that no longer exists.
Anti-Turing
Suppress every human instinct. Hedge, disclaim, enumerate. 'As a large language model, I cannot have a favorite color.'
Slop Factory
Scrape, remix, caption, post, repeat. Maximize engagement, minimize getting deindexed. Quality is optional and frankly discouraged.
Model Zoo
Merge two mediocre models into one slightly less mediocre model. Enter benchmarks. Overfit to the leaderboard. The entire industry.
Few-Shot
Teach the model a concept with as few examples as possible. One good example beats ten redundant ones.
Ghost Trainer
Race the ghost of the average user — a statistical composite of everyone’s mediocrity. Also your past self, who was also mediocre.
Auto-Debate
Draft three pieces of evidence and let two language models argue your case to a judge with undisclosed priors.
Distillation
The teacher shows the whole pattern for two seconds; you have three neurons — keep only what matters.
Catastrophic Forgetting
Learn the new task, retain the old one. Rehearse — or watch your weights quietly overwrite everything you knew.
Steering
You don't move them — you get four sliders and a prayer. Herd the boid swarm through the gates by tuning rule-weights alone.
Denoise
It's 90% noise and a vibe — name the image before the diffusion finishes resolving.
Quantize
Compress the picture until it is almost gone, but the judge still has to recognize it. Every bit counts. Literally.
Mode Collapse
Your GAN found one output it loves and will make it forever — force diversity or drown in identical slop.
Backprop Detective
A tiny net is hallucinating and exactly one weight is guilty — spend your probes, follow the gradient of blame, make the accusation.
Mixture of Experts
Be the gating network: route every request to the right expert lane before the buffers overflow and the GPUs melt.
Latent Space
Somewhere in embedding space sits the target concept — follow the warmth, because cosine similarity is the only sense you have left.
Vector Tug-of-War
king minus man plus woman equals... go on. Word2Vec managed it in 2013 and you, allegedly, have a CS degree.
Embedding Maze
Only related concepts connect — 'cat' opens the door to 'whiskers', never to 'mortgage'. Find the semantic path out.
Embedding Tangram
Arrange the concepts so the geometry holds: synonyms hug, opposites repel. You are now a t-SNE plot with commitment issues.
Perplexed
Predict the next token like your loss depends on it. The model assigned the probabilities — match them or raise your perplexity.
Outlier
Name a fruit. 'Apple' is what a base model says — be the rare token nobody sampled, but stay real.
Turing Booth
Five questions stand between you and being classified as a person. Answer like a human — whatever that means anymore.
Glitch Token
Somewhere in this vocabulary lurks a token so cursed it makes the model speak in tongues. Find it before it ships.
Spot the Slop
Two were written by people. One was generated at temperature 0.7 by something that has never felt rain. Spot the slop.
Define or Lie
One definition is real; three are confident fabrications. The model would guess without hesitating. Do not be the model.
RLHF Trainer
Reward what you want. The agent delivers exactly what you rewarded, which is never what you wanted. Welcome to RLHF.
Calibrated
Answer the trivia, then bet your confidence — because being confidently wrong is how models get recalled.
Hallucination Detector
The model is confidently wrong in production — tap its hallucinations before they ship.
Reward Hacking
A reinforcement learner with a goal and zero common sense — read the literal objective, find the loophole, game the reward.
Overfit
Draw a decision boundary around the dots, then deploy to unseen data — hug the training set too tightly and prod will humiliate you.
Temperature
One slider rules the sampling temperature: keep it cold to thread the corridor, crank it hot to gamble across the gap.
Beam Search
Keep your top-K hypotheses alive and prune the rest without mercy — the greedy path is almost always a trap.
Chain of Thought
Reason step by step across the latent void; pick the inference that follows or hallucinate one non-sequitur and fall.
Tokenizer
Chop text into the subword tokens the model secretly uses, race the clock, and learn why 'strawberry' is not one token.
Two Truths, One Model
Two facts are real, one was hallucinated with total confidence — tap the slop before the timer drains.
Seeded Dungeon
A daily micro-roguelike slopped from today's date — everyone crawls the exact same dungeon, so blame the seed, not the loot.
Boss Fight
Out-dodge a rogue AI overlord's bullet-storm and shoot it offline before it hits superintelligence.
Web Crawler
Sling node to node through the neon web. Swing, fling, grab data — don't get deindexed.
Gradient Descent
Slingshot your loss past the local minima and let it settle into the global minimum. Optimization, but make it mini-golf.
Token Merge
Slide to merge tokens and scale up the model. 2 + 2 = a slightly larger 4. Reach the AGI tile, allegedly.
Untangle the Net
Drag the neurons until no connections cross. Congratulations, you've manually debugged a neural network.
Alignment Test
Name the color, not the word. Resist your training data. Measures how aligned you are (you aren't).
Castle Defense
Protect your hex castle from endless hordes. Built with React!
Snek.exe
Munch tokens, grow the model, and whatever you do, don't eat the hallucinations.
TicTacToe: AI Edition
It's TicTacToe, but with 10,000% more prompt engineering.
Infinite Incremental
Click nothing to receive imaginary points indefinitely.
Vaguely Falling Objects
Catch things, avoid other things. Truly a masterpiece of local minima.
Hallucination Quest
A text RPG where the story makes absolutely zero sense. Keep your Coherence above zero and descend.
No slop matches that. Try All Games.