the systems decide and the people learn later. seven signals from one week where the call was already made.
the model below is listening to two things at once. the officer's voice. the cartoon on the tv behind him. drag the ambient slider up and watch where the official record goes.
the officer was responding to a domestic call. the body camera was on. the transcription model was running. somewhere in the room the television was playing the princess and the frog. the model has no concept of source. it has audio. it has a confidence score. it produces a sentence. the sentence in the record reads, in part, the suspect turned into a frog and hopped under the couch.
the report was filed before anyone read it. transcription is supposed to be the work; the record is supposed to be the receipt of the work. when the receipt is generated by the same machine that did the listening, no one is holding it up to the room. the failure looks like a joke. the failure is that the record went into the system, and the system did not know it had been told a lie by its own tools.
every dot below is a publicly trained model. drag any cluster and watch the cost curve. lance is the small one near the bottom that should not exist.
the model is three billion parameters. it understands and generates across image, video, and text in any direction. the cluster was small enough to fit in a single rack. the license is apache two. the training compute, on spot-priced a100s, comes to roughly the cost of a mid-sized round of seed funding. the labs have been arguing for a year about whether the next frontier model needs one hundred thousand h100s or two. bytedance trained this one and posted the weights.
the scaling laws are not wrong. they are also not the only road. when the architecture is solved, the cluster shrinks. it shrinks unevenly, and not at all for every problem, but it shrinks. lance is the proof that the smallest plausible training run for a multimodal model dropped past the threshold where one company has to do it. it is now a long weekend, on a graduate budget, in english or chinese, posted under a license that says you can keep it.
the card on the left is the upload. the calendar on the right is anthropic's morning. read the conflicting tags and watch the meeting tick by.
the readme says trained from scratch. the config says qwen 32b. the tokenizer is qwen's. the safetensors weights are sized for a qwen fine-tune. on the same page, both things are claimed and neither is corrected. the upload appears on a saturday and is downloaded over a thousand times before anthropic's monday morning standup begins.
the cost of misrepresentation has dropped to one click. the cost of finding out is a comment thread. the cost of replication is whatever a hugging face card weighs in bytes. the product surface for frontier model marketing is now a wiki page someone uploaded over the weekend, and the time horizon for caring about it is three days.
this is the canyon. you are the state. fire enters at the upper left. wind is southwest at 22 mph. you can call evacuation once. the model already called it at minute zero. find your moment.
firesight was reading remote sensing data and fuel maps when it raised the alarm. the call did not require seeing the fire on the ridge; it required noticing that wind, slope, dry fuel, and the road geometry would close the only evacuation route inside an hour. the model put a flag on the dashboard at fourteen oh two. the duty officer did not see it until after lunch. the state issued the order at fourteen forty two.
nobody died. that is not the point. the point is that for forty minutes the warning existed and no one received it. the model did the work of perception. the institution did the work of routing. the institution is still slower than the perception layer, and it is now becoming visible that the bottleneck is not data. the bottleneck is who is allowed to act on data that arrived without a human signing for it.
the ruling below is the actual passage, in english. hover any phrase to see where it came from. one phrase, when you find it, has no human source.
the judge wrote the ruling. one of the sentences was suggested by chatgpt. the sentence is good. it is consistent with prior case law. it does not change the outcome. the worker is reinstated either way. but the sentence is in the record now, and the next time a litigant or a court looks for prior reasoning on this point, they will find it under the judge's name and not the model's.
this is how the model gets into the law. not through statute. not through landmark cases. through a clerk who needed an opening line for the section on inquiry, and a judge who signed the order without flagging which sentence was theirs and which was the assistant's. the legal record is becoming a corpus that includes its own generator. ten years from now no one will be able to tell which sentences were written by judges and which by something the judge consulted before lunch.
below is the grid. each cell is a unit. the model adjusts every fifteen minutes. one of the units is yours. try to set your own rent. the grid will tell you whether the model agrees.
nobody in san diego sat down and decided to raise rent fourteen percent. the people who own the apartments did not call each other. they did not need to. they subscribe to the same revenue management platform. the platform reads occupancy, lease length, comparables, season, and sends back a recommendation. the owners click accept. the recommendation, run across hundreds of thousands of units in the city, moves the median up in lockstep.
the legal framing has been antitrust because that is the closest precedent. but the actual mechanism is not collusion in the old sense. it is shared inference. a market is a price-discovery process; price discovery requires that the discoverers be independent of one another. when every owner accepts the same model's output, there is no discovery happening. there is one decision, scaled out, and rent that goes up because the algorithm believes the market will tolerate it.
read both files. give each defendant a risk score from one to ten before you scroll. the system already has its scores. compare after.
compas was trained on prior outcomes. prior outcomes were the result of prior decisions, which were the result of who got arrested, who got charged, who got prosecuted hard, who got released early. the model does not see race because the model does not need to see race. it sees zip code, employment history, family structure, school disciplinary records. it sees what the system has already been doing for forty years, and it produces the same shape of answer.
the seventeen year old in the first file scored higher than the forty year old in the second. the seventeen year old did not re-offend. the forty year old did. these are not edge cases — they are the audit set. the predictive performance of the model is not a question of whether it works. the question is what it is trained to predict, and the answer is that it is trained to predict what the system did before. it is a measurement device pointed at its own past.
the call gets made faster than the people who are supposed to make it. the record gets written before anyone reads it. the price moves before anyone names it. the score is already in the file when the file opens.
seven signals from one week. each one a system that already decided. each one a place where the decision moved into the model before the institution noticed it had handed the decision over.
nothing here is sponsored. nothing is optimized for retention. the signal is the product. the systems decide and the people learn later.