Our Many Hypothesis

Transformer Time-Series Models (TTSM) – Research and Development

Hypothesis

We can build a TTSM using blockchain data because transformer architectures enable the usage of large scale data to create generalized models. This model will generalize to other domains.

Result

Indications are positive. We have the best performing model on the ETT dataset but we are working towards making this model useful.

Extra Tidbits

The main hypothesis of Dither is that transformers models can be used for any type of input data. This has been hinted by the broad application of transformers for other modalities like images, text, video, audio, etc. Many of these modalities have quirks like time-series.

We are in research and development but from what we have seen, it is very possible to forecast using transformer architectures. We have also seen that the current research is limited. This is the least researched domain of any transformer modalities.

The internal TTSM models currently are geared towards large cap tokens but we are working to create a generalized model that will work with any input.

SeerBot – Deprecated

Hypothesis

We can use the transformer architecture to classify meme tokens and their future performance.

Result

Short term results were positive. However, the speed of the market changes deprecated the transformer based classification model. For small caps, it is better to use non-transformer based models for high noise applications.

Extra Tidbits

SeerBot was the prototype that became the Dither Tool suite. It used custom time-series processing, custom embedding layers, classic transformer processing layers and a custom final layer for classification of tokens. It proved to be effective during the initial euphoric memecoin period of 2024 but lost effectiveness over the summer chop. It was not effective enough to be a reliable signal over time despite evidence of efficacy.

The largest issue we found with SeerBot is the limited and shifting signals in the microcap memecoins. Time-series forecasting is not effective in these regimes because of the noise to signal ratio. When PumpFun joined the scene, the ecosystem had changed enough that the original SeerBot model was not effective enough to warrant continued offering. We have since moved to forecasting on larger cap tokens and classifications based on bonding for low cap tokens.

We currently still use the Bot ‘SeerBot’ to serve our premium demos.

SeerLite (unicorns) – Demo

Hypothesis

Adding a human in the loop to SeerBot would improve efficacy.

Result

Our hypothesis is supported. The user effects are as powerful as the machine learning effects of Seerlite.

Extra Tidbits

Seerlite is a free demo. It is an on-demand model of our original SeerBot architecture. SeerBot was an alert system, whereas Seerlite is a query system with a human in the loop. Seerlite is used as an information source, helping verify instead of alert on new tokens.

The query pattern also allows users to re-query the same token. If the first signal is uncertain, querying again over time will allow users to gain more informed signals. Seerlite operates on a shorter time horizon.

Large Caps – Demo

Hypothesis

Large cap tokens are more predictable (easier to forecast) because they have a higher signal to noise ratio.

Result

Although this is a classification model, there is much more signal using on-chain data for large cap tokens compared to micro cap tokens.

Extra Tidbits

We currently have a large cap classification system. It is intended to signal tokens that might turn bearish or bullish over a 4-day window. This system is still noisy. Exit signals appear more effective than entrance signals, which are meant to indicate possible exits or entrances over a 4-day period, not to direct immediate action.

Otto – Demo

Hypothesis

We can use AI models to create a mod system that enhances human mods.

Results

Success! Otto is a prototype mod. With minimum improvements, it would be effective in moderating token telegram channels.

Extra Tidbits

Otto is an automod system that moderates a telegram channel. Otto answers questions about the project, deletes aggressive or overt messages, and interacts in the Dither public telegram channel. Otto is effective but needs further refinement before release as a product.

Internal Quant – Research and Development

Hypothesis

We should have an AI agent that peruses various products and buys/sells.

Results

The models were not good enough. It was around break even. We also had failed transactions which limited the experiment.

Extra Tidbits

During testing in the summer of 2024, our Internal Quant was able to successfully trade a large cap basket of memes while aiming for automated portfolio management without human intervention. However, execution failures and a choppy market led to inconclusive results. Live testing is paused until the release of new models in development.

Bonded Clustering Model – Demo

Hypothesis

We can classify tokens at bonding and use known values to determine the likelihood of rugging.

Results

Positive! Initial trials showed success with an ML clustering model, though the effectiveness waned over time. Testing continues with an expanded dataset. We will soon have a model which is continuously improving. Training is quick enough we may implement nightly updates.

Extra Tidbits

We use a classical ML model (not transformer-based) to cluster tokens that bond on Pump.fun, aiming to detect high-risk tokens. Early trials showed ~62% success, though the effectiveness decreased as tokens adapted. This tool remains in our Premium demos as we continue to refine it.

Sports Betting – Research and Development

Hypothesis

We can use either our new TTSM or our prior method to predict sports better than leading bookies.

Results

Failure: Specialized models are better than generalized models for sports at the moment. We achieved comparable performance but without the expected edge for broader release. We are confident in a new generation of generalized time-series models outperfroming current specialized models. This is not a short term research project.

Extra Tidbits

Dither started with attempts to use AI to predict football outcomes. Early variants of GPT3.5 yielded results, but the project evolved into TTSM exploration. Attempts in July to recreate the model or use TTSM for predictions were inconclusive. Current research in sports betting focuses on improving base models for better viability.

Personalized Quant – Idea

Hypothesis

We can recommend tokens based upon past trading activity.

Extra Tidbits

Using time-series and text embeddings, this idea aims to recommend similar tokens based on prior activity, whether consolidating similar tokens or exploring new trends.

Wallet Text Embeddings – Idea

Hypothesis

We can grade token holders based on wallet text embeddings and prior transactions.

Extra Tidbits

Using text embeddings, it may be possible to cluster wallets based on transaction history and predict the potential behavior of holders.

PhotoAI for Memes – Idea

Hypothesis

We should be able to train a diffusion model on memes and then reproduce custom memes. Similar to how AI photoshoot models.

Extra Tidbits

One failure mode would be if the model is primarily trained on people. It's likely this is not a concern and can be implemented rather quickly.

Meme Image Generator – Idea

Hypothesis

We should be able to use inpainting (AI strategy) to give AI or people the ability to very quickly make memes.

Extra Tidbits

The technology exists for this. The challenge is engineering only. Combine a few OS AI systems and you can build a very powerful meme generator. I believe there's a more streamlined dingboard that could be built for AIs to use directly.

Meme Studio (Auto-token launches) – Idea

Hypothesis

Agents allow for meme studios that are less shady than current organizations which act as meme studios.

Extra Tidbits

AI agents could create meme studios, organizations that launch meme tokens transparently, allowing for neutral community formation and minimizing risks of rug-based CTOs.

Memeability – Demo (Seerlite)

Hypothesis

We can rank the memeability of an image using vision models.

Results

Yes! This is a simple application of AI technologies.

Extra Tidbits

We use a vision model to grade the memeability of tokens, creating a dataset to quantify meme potential.

Image - Price Prediciton – Fail

Hypothesis

We tested whether we could use vision models to correlate memeability to total price.

Results

No! The meme might be necessary but it is not sufficiently high signal to determine the success of a project.

Extra Tidbits

We trained on a vision model on images and peak prices. However, the data was too noisy to deduce a signal on meme quality alone.

Originality – Demo (Seerlite)

Hypothesis

We can determine the originality of a newly launched token using embedding models.

Results

Yes! This is a simple application of AI technologies.

Extra Tidbits

We use text and vision embedding models to assess originality in image, description, and name for new tokens. This could be improved with a better RAG system.