BitMind

SN34 BitMind sells a counterfeit-spotting desk for images, video, and audio, where strangers compete to build the best detector and other strangers are paid to forge better fakes to test it.

The commodity, explained from zero

A deepfake is a picture, clip, or voice made by an AI model and passed off as real. A bank's counterfeit desk looks at a note and says real or fake; BitMind does the same for media. The product is an API, an address a program sends an image to and gets a verdict back, at api.bitmind.ai with a key, plus an upload page, a browser extension, and phone apps.

It has a price because companies act on media they did not make; the front page puts it as: "We verify the authenticity of every data point before your AI acts on it." Pricing has three tiers: free for 100 requests a month, Pro at 100 dollars a month for 10,000 or more, and custom contracts for "Large organizations with bespoke needs, SLAs, and integrations." The docs name consumers, developers, and enterprises as paths, and finance, social media, and government as sectors. No customer is named.

Why it is on Bittensor at all

The claim is adaptation, not cost. A fixed detector goes stale as generators improve; BitMind says it wants "a dynamic, self-evolving service that instantly adapts to new threats", and the subnet is how: detectors come from an open contest, and a second set of miners is paid to make fresh fakes for them to catch. That is where the comparison breaks: a bank's forgers are on staff and their notes never leave the vault, while here the forgers are strangers paid in the subnet's token, and their fakes feed the weekly test set. Nothing is said about cost against a centralized detector.

How the work gets done

Detector miners submit a classifier, a model that takes an image, video, or audio clip and scores it real or synthetic; they do not run servers. Generator miners make synthetic media on demand and earn a base reward for valid output times a multiplier, capped at two, for how often they fool the detectors. Validators, the judges who decide pay, test detectors weekly on fresh real and generated media and score them on accuracy and on how well calibrated their confidence is, combined into one number, sn34_score. Emission, the subnet's share of new tokens, is split 40 percent image, 40 video, 4 audio, and 16 to generators; within each lane the reigning best model takes 85 percent and the two previous champions 10 and 5, and a challenger must beat the king by a 0.01 margin after a 48-hour defending window.

How you would know it works

The detector at bitmind.ai/detect accepts an image or video upload without a stated sign-in; that is the artifact. The front page claims "95% accuracy on real-world deepfakes" with no date, and its usage counters render as zero to a fetcher.

What is missing

On chain the subnet has no URL, no description, and no Discord; a wallet reader has no front door. The chain contact is intern@bitmind.ai while the docs use hello@bitmind.ai. The docs' API Reference link is broken and the full reference sits behind an account. There is no status page despite a 99.9 percent uptime claim. Registration cost and the deregistration policy are not stated.

Go deeper

• Try the live detector
• README: the adversarial loop
• Incentive: formulas and shares

Metaphor: a counterfeit-spotting desk that pays its own forgers. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.