Humans want to predict the future. This is an age-old conundrum. In recent years (since 2018) bona fide machine learning models have begun to best traditional statistical models in forecasting. Transformers for time series have only been in fashion since 2022 when PatchTST showed quality performance in long-term forecasting. Throughout 2023 many teams unsuccessfully attempted to convert LLMs to time series and it is only recently with TimesFM and Chronos that Google and Amazon, respectively, have shown competitive performance in Large Time Series Models (LTM). Competitive performance is not good enough.

Transformers are data hungry models. They require enormous sums of data to perform as well as other model types. This is a double edged sword. Since the models can consume large swaths of data without saturating, they are able to perform better than traditional models at language processing and vision. The issue with time-series and transformers is there is a limited amount of publicly available data to train a large model. However, if a transformer model could be trained and rival the advancement to time series that was done in NLP then the impact would be even greater.


Transformers have had a difficult time transitioning from NLP to time-series due to the limited data available of sufficient quality and quantity. Blockchain provides a promising solution. They are immutable stores of information that are perfect for training a foundation model. It has been shown that quality data with common data augmentation techniques allows for large models to learn effectively. With blockchain data, supplemental open source data and synthetic data, we will be able to have enough information to properly train a foundation LTM.

Market Figures

Time series is a misunderstood term. It is not always clear what it means or where it can be applied.

Definition: a series of values of a quantity obtained at successive times, often with equal intervals between them

What does this mean and where can it be applied? For instance, stock market and financial forecasting is an obvious answer. Food distribution is another, albeit less, obvious answer. Even wearable devices have time series applications (Data in time with a target value). Anything with a quantity that has measurements in time are time series.

With this broad of an application space, it can be difficult to assign a market value for time series and the associated applications. Forecasting is used in every industry from agriculture, healthcare, manufacturing, financial products, ad infinitum. A significant breakthrough in long term forecasting will touch every industry and every person.

Dither the token

Humane (the company) is a reminder that startups should begin niche – even hardware. Solana as a data source made sense. Testing our model in the trenches of crypto also makes sense. Crypto is one of the most pure PvP financial systems in the world. If we are able to prove the effectiveness of our model within this system then we are confident that it will generalize. Allowing users to directly test our hypothesis lends more brains thinking about the output and more data to feed into the input. Although our models are private, we benefit from many individuals using our product.

Dither enables access to our flagship research product: SeerBot. 2,500 DITH gives perpetual access to our Standard model. 50,000 DITH gives perpetual access to our Premium version. Premium will receive the latest and greatest updates, features and products before or even if they are released to Standard. Some features will be premium only.