Variant Fund: How can the crypto world shape a more tasteful AI model?
Original article by: Alana Levin
Original translation: TechFlow
Over the past two years, there has been an explosion of new AI models that can accomplish many types of tasks—from finding information and answering questions to providing customer support, proofreading documents, generating content, and more.
Many such tasks are objective and have clear optimization functions: finding the correct answer, identifying the most relevant information, detecting any errors or anomalies, etc.
But there are also models whose outputs are extremely subjective, like producing "good" art or developing "interesting" videos. I call these "taste-based models." Taste-based models tend to be harder to optimize because they are a mix of collective and individual decisions; there are no obvious answers or outputs. Therefore, frequent feedback is particularly valuable to help the model stay up to date on cultural preferences.
Nowadays, there are roughly two ways to cultivate a model's "taste":
Based on user-generated content/data (such as Twitter or Reddit feeds), this could theoretically reveal the latest trends in human attention (thus acting as a proxy for taste).
Leverage a community of human “taste makers” to help models proactively train around their preferences.
There are many things that are not ideal about the first approach. Data can be siloed (e.g. Reddit shutting down its API) or biases can be introduced (e.g. only a subset of data is shared). Models can also be overfit to the algorithms of a specific platform, especially if it has a limited selection of data sources. This may not sound important until one starts imagining the massive amounts of new media generated based on Twitter’s popular content. This is not ideal.
The latter approach, a network with human feedback, avoids many of the risks described above. There may still be bias, but only in that it includes the preferences of the community members who opted in to help train the model. Therefore, the key is to ensure that these community members, the people who formulate the "taste", have a vested interest in the model actually developing good taste.
Crypto rails can help facilitate this alignment. Providing ownership in the model/giving participating members a financial stake in the model’s output can incentivize them to actually participate. Cryptocurrencies also make participation more open and easy: anyone, anywhere in the world can contribute, as long as they have an on-chain wallet and a network connection.
One notable example is the Botto project. Botto is an autonomous artist where $BOTTO token holders have the ability to help train the model each week. Training is simple: participants vote up or down on various images, and Botto learns from members’ preferences. At the end of the week, the most popular works are auctioned off, and participants who helped train Botto that week are paid.
Art is just one category of tasteful models. Others might include film, television, other forms of storytelling (novels, short stories), comedy, and advertising/brand campaigns. Even a few years ago, these tasteful models were impossible. The tools were less expressive, slower, and the models could not be reliably expected to produce cohesive or (in the case of video) photorealistic output. Only today are these becoming possible.
Importantly, there is a large (and growing) addressable market for models with taste. Art is a multi-billion dollar market. Content consumed online accounts for trillions of hours of attention each year. If people are already going to spend time and money on these forms of entertainment, it seems reasonable to let them have a share of the production, which would not only create a more engaged user base, but also a more satisfied user base. It would be really cool to imagine a Best Picture Oscar where the main players were the audiences who helped train and develop the storyline, or a whole new award for community-created films.
I think of this as creating a new category of content, rather than replacing existing creation. This is similar to how smartphones and Instagram enabled everyone to be a photographer, and the existence of these new technologies did not eliminate the work of actual photographers, but in fact, may have made the work of photographers more appreciated. The same is true of models with taste: they expand each of the above categories by creating a new form of participation by leveraging new technologies, in this case crypto rails for consumer ownership and economic docking.
We’ve seen thousands of new models emerge over the past few years. There will likely be millions (or more) of new models in the next few years, and at least some of them should strive to engage stakeholders in new ways, from greater openness and accessibility to new ownership structures that experiment with incentives. Tasteful models are one area particularly suited to this kind of innovation, but they are unlikely to be the last.
