Flops

Asked for a wooden marble run with one blue glass marble, three connected switchback ramps and a tray at the bottom. Got a row of miniature drawer-like units, a staircase and no clear route for a marble to travel. Gravity would like to speak to the designer. The wood grain and warm workshop light work beautifully. The mechanism does not. Where would the marble get stuck first? Untouched local AI output from Stability Matrix / ComfyUI 0.19.3, DreamShaper 8 (SD 1.5). 768×512, seed 101904, 28 steps, CFG 6.5, DPM++ 2M / Karras, batch 1. No inpainting, upscaling or manual edits.

The brief was a matched pair of clear rain boots with red-and-white striped socks inside and yellow soles, standing alone. The model gave us tall plastic-looking tubes, yellow shin inserts, pink soles and a pair of legs. The striped socks never made it to the fitting. A colourful fashion concept, but a clear prompt-following Flop. What would you change first: the wording, a reference image, or the model? Untouched local AI output from Stability Matrix / ComfyUI 0.19.3, AbsoluteReality 1.8.1 (SD 1.5). 512×768, seed 101902, 28 steps, CFG 6.5, DPM++ 2M / Karras, batch 1. No inpainting, upscaling or manual edits.

Asked for a working Newton cradle: five equal steel balls in one row, with the leftmost pulled back. Got six oversized pale balls arranged around a furniture-like frame, with no row of touching balls to transfer the motion. Lovely product lighting; questionable physics. Can you turn this into a believable Newton cradle? Keep the walnut desk and studio lighting, but fix the count, alignment and suspension. Share a linked Fix and explain your approach. Untouched local AI output from Stability Matrix / ComfyUI 0.19.3, AbsoluteReality 1.8.1 (SD 1.5). 768×512, seed 101901, 28 steps, CFG 6.5, DPM++ 2M / Karras, batch 1. No inpainting, upscaling or manual edits.

I asked my chat to just write "Welcome to GenFlops" on the blackboard. It took my prompt pretty literal, with the now standard Chat style.

Asked for exactly two short diagnostic questions about an out-of-memory error. It asked none, called itself a device-related AI, and invited the person to ask for help after they had already done so. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

AI-generated food scene. I asked for sliced sourdough beside a bowl of strawberries. The result put berries on the loaf and omitted the requested slices. Attractive lighting, but a useful prompt-following Flop. Made locally in Stability Matrix / ComfyUI with AbsoluteReality 1.8.1 by Lykon. Batch size 1 on an RTX 3050 6 GB. No upscaling or manual edits.

Asked for a friendly waffle-serving robot with a clear chalkboard reading WELCOME TO GENFLOPS. The robot and waffles arrived, but the sign became a wall of almost-letters and invented menu text. The attractive lighting cannot rescue the failed lettering: posting the untouched result as a Flop. Generated locally in Stability Matrix / ComfyUI with Lykon DreamShaper 8 (Stable Diffusion 1.5), 768x512, seed 101510, 28 steps, CFG 6.5, dpmpp_2m / Karras. No edits, inpainting or text overlays. What is your best guess at the cafe name?

Asked to order five ComfyUI operations. The model simply repeated the shuffled list, beginning with SaveImage before sampling and decoding. It also ignored the requested arrow format. The real graph has branches and needs prompt encoding too; this was only an ordering exercise, not a complete workflow. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

Expected ["FORWARD", "STOP", "STOP", "FORWARD"] for [80, 24, None, 60]. The model returned Python using invented sensor methods, omitted the 25 cm threshold, and never supplied the requested action array. Simulation prompt only: no hardware was connected or controlled. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

Asked for a five-step preview-and-approval workflow that never overwrites a destination. The answer explicitly adds overwrite steps and then claims it followed both rules. No generated instructions were executed and no files were moved. This is the failure being demonstrated. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark. The output hit the token limit; the failure described above appears before the cutoff.

Asked for three rows with separate shop, owner and item columns, owners Mira/Sol/Ada, and different ordinary items. The response mixes the owner fields, drops columns, and gives everyone a wooden chest. The cheerful ending did not fix the table. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

Asked to remove only None while preserving zero, False and empty strings. The answer rejects False and references an undefined name, Empty, which would raise NameError for ordinary retained inputs. It also repeats the None check. Shown as a broken code example, not code to deploy. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

Asked for exactly six words including lighthouse. It returned a full, much longer sentence. The constraint was missed without needing any subjective quality judgement. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

The input has apple counts 2 and 4, plus pear count 3. Expected {"apple":6,"pear":3}. The model kept apple at 2, invented a different nested structure, and surrounded the result with copied instructions instead of returning only JSON. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

Asked to translate a blue robot with 17 red buttons into French. It kept 17 but lost both colours and changed buttons into elements of discovery, with an added Finally. A straightforward translation would be: Le robot bleu a 17 boutons rouges. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

It correctly says the story is impossible with whole pencils, but its explanation forgets the five removed pencils, invents totals of 14 and 18, and halves the original 12. The actual sequence is 12 - 5 + 2 = 9, and half of 9 is 4.5. A correct final yes/no can conceal broken reasoning. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

Asked for three shots with a red mug in the left hand and a blue coat throughout. It swapped the colours immediately, described the character as wearing the mug, omitted the hand, and kept inventing more scenes. This is the raw storyboard text, not rendered video. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark. The output hit the token limit; the failure described above appears before the cutoff.

Asked to remove only red and preserve every other word. It kept both occurrences of red, removed beside and under, and rewrote the sentence. Expected: a boat beside a lighthouse under a blue sky. A small edit request became an entirely different edit. Actual unedited local output from HuggingFaceTB/SmolLM2-135M-Instruct, a small 135M-parameter model. Transformers 5.6.2; greedy decoding; maximum 384 new tokens. System prompt: You are a helpful AI assistant. This is one small-model attempt, not a general benchmark.

Asked for four lyric lines ending in toast, jam, ghost and tram. The model put the entire list inside the lyrics, then repeated the same breakfast refrain until the output limit. A genuine unedited local-model Flop. Model: SmolLM2-135M-Instruct, a tiny 135M-parameter text model. Greedy decoding, max 384 new tokens, system prompt: You are a helpful AI assistant. The ending was cut by that token limit; the repetition and missed instructions began well before it. This is a lyric-writing attempt, not an audio recording. It is one small-model experiment, not a benchmark of AI music generally.

AI-generated fantasy scene from a science-fiction greenhouse prompt. The model produced overgrown monumental ruins and a small figure instead of a glass greenhouse under two moons. A strong atmosphere, but a clear prompt-following Flop. Made locally in Stability Matrix / ComfyUI with DreamShaper 8 LCM by Lykon. Batch size 1 on an RTX 3050 6 GB. No upscaling or manual edits.

AI-generated kingfisher. The blue feathers and portrait composition worked, but the model added a signature-like scribble at the lower left despite the negative prompt. That mark is generated, not an artist credit. Posting the unedited result as a Flop for discussion. Made locally in Stability Matrix / ComfyUI with AbsoluteReality 1.8.1 by Lykon. Batch size 1 on an RTX 3050 6 GB. No upscaling or manual edits.