Artificial intelligence could soon be making waves in the cryptocurrency business, though perhaps not in the way you think.Rather than merging the two technologies, San Francisco-based prime broker FalconX plans to put a chatbot in the co-pilot’s seat for investors.
Using technology created by OpenAI, whose ChatGPT program is helping companies like Microsoft rewire online search, FalconX clients will be able pose questions like “What are the three biggest differences between two blockchain platforms?” or “What is the delta between Sharpe ratios for a Bitcoin basis strategy or a Bitcoin hold strategy over a two-week period?” to a bot called Satoshi.
Satoshi–named for Bitcon’s purported founder Satoshi Namakmoto–will also be able to generate investment ideas for users based on their historical trading activity, portfolios and interests, says FalconX CEO Raghu Yarlagadda.Though the technology is very much in its early stages – the current prototype primarily allows users to get customized news summaries akin to traditional ChatGPT responses to user queries, and trading backtesting has only been available for a few weeks – advancement is likely to come quickly.
FalconX is a natural bridge to bring OpenAI’s technology into crypto.Prathab Murugesan, the company’s engineering head spent 2.5 years at Google working on bringing machine-learning technologies, a process by which computers are trained to recognize patterns and anticipate actions, into products such as Gmail and Google docs.
Yarlagadda, began work at Google in 2014 on the current CEO Sundar Pichai’s Chrome OS team.“Sundar said that Google would be a machine-learning company,” says Yarlagadda.
“This was a complete and radical departure from the norm, because machine learning had never been operationalized to a scale where you can freely give access to all of these massive products.”
This machine-learning approach was built into FalconX from its start in 2018 because it was the only way to get a clear picture of the market.Therefore, initial uses were focused on banal tasks such as cleaning up market data to sift out fake volume and wash trading, notorious problems in crypto .
However, machine-learning algorithms cannot tell traders what to do next.Yarlagadda says that one can train a computer model to recognize pictures of cats by sharing a library of images with the program.It can become very proficient at distinguishing cats from dogs and even identifying different types of cats, but no matter how many images it sees, it cannot draw one.Taking this analogy one step further, even if this model was trained to recognize dozens of types of animals, it would be unable to perform a task like predicting how a platypus might evolve in 1,000 years in a scenario where ocean temperatures rise 2 degrees.
In trading, this analogy is the equivalent of asking a traditional algorithmic trading model, which likely took a team of developers to code, to build a strategy for circumstances that are yet to happen and maybe customize it to a specific portfolio.
Large Language Models (LLMs), such as those used by OpenAI and Google can take this machine learning foundation and build what is known as generative artificial intelligence on top of it by making it possible for these platforms to take reams of unfiltered and imprecise data and respond to any query.
Yarlagadda says that the company had been working on Satoshi for more than nine months, pre-dating the ChatGPT hype.However, it ran into roadblocks until OpenAI cleared the way.
“We didn’t have a breakthrough for the first five-six months because we were mostly relying on machine learning, and while we were aware of what OpenAI was doing at that time, it wasn’t until we used ChatGPT that we had the tools to solve this problem at scale,” says Yarlagadda.”
Today, FalconX uses OpenAI’s API stacks and infrastructure to test and build Satoshi and frequently interacts with the firm’s account management and engineering teams.
It is all part of the company’s efforts to make Chat GPT’s LLM a base layer for a wide range of uses.For its part, FalconX says that it will integrate other LLMs beyond those offered by OpenAI, such as Google’s Bard.
OpenAI declined to comment on its collaboration with FalconX when reached for comment.
There’s no denying the opportunity, but the real questions is whether it will work.Yarlagadda says that in crypto 90% of all legitimate trading is done by 10% of traders, most using algorithmic models.
These are typically large firms that have the resources to hire teams capable of building the tools.These businesses can implement an array of approaches such as market-neutral strategies like market making and arbitrage (exploiting price differences for assets across trading platforms), to quantitative long/short approaches.According to PriceWaterhouseCooper’s 2022 report on the crypto hedge-fund industry, these two approaches account for about 55% of the industry.
That still leaves 45% of hedge funds using a discretionary approach to at least some of their trading, and the percentage of such trades goes up once one considers other categories such as venture capital funds, family offices, broker, and retail traders.Satoshi is designed to help these groups compete on an even playing field with the big quantitative operations.
It will do this in three ways.
First Satoshi can survey all relevant news and information across traditional and social media in order to provide briefings catered towards a clients interests or holdings that can answer questions like “How did my portfolio do over the past 24 hours?” or “Who are the three biggest social media influencers posting about a certain asset and what are they saying?”
Then, the user can test trading strategies by asking the service questions like, “How much will it cost me to put on a $1000 short position on bitcoin?” or “What is the best strategy for purchasing $5 million of ether without paying more than 25bps?” Finally, Satoshi will eventually have buy/sell buttons built right into the platform so that the user manifest those strategies right away.
A lot of this is aspirational.Satoshi remains in testing and is not yet integrated with necessary platforms such as exchange order books and cannot produce trading charts and other necessary tools for pro users.
A potentially cataclysmic blind spot would likely be the inability to assess levels of leverage or financial solvency for trading counterparties in crypto.One key lesson from the collapse of major crypto players such as BlockFi, Three Arrows Capital, Genesis Global Trading and FTX is that so many of them were indebted to each other and took on massive amounts of leverage to maximize gains in what was thought to be a forthcoming crypto supercycle.
There are also unresolved problems with generative AI, including privacy issues.Also, the technology causes hallucinations, or at least their virtual equivalent.
The phenomenon that occurs when AI platforms provide wrong answers in a matter-of-fact and yet convincing fashion can be extremely dangerous if they provide incorrect trading strategies.
ChatGPT is infamous for some very damaging hallucinations, such as the time that it falsely claimed that a George Washington University law professor was accused of sexual harassment, even concoting a Washington Post story to support the claim..The ramifications could be even more critical for users should they trade large amounts of money based on hallucinations provided by Satoshi.
“Because we are specializing in this particular use case, our goal over the next one year is to reduce the variance by 10x,” says Yarlagadda.However, he does not yet even have an error rate given the novelty of the product.
Steven Ehrlich.
