Welcome Avatar!
As you’ve likely seen on Twitter we're quite interested in the current AI / LLM landscape. While there is still a lot to be desired in terms of accelerating the research side of things, we can see the potential.
The advent of Large Language Models (LLMs) in crypto is revolutionizing how non-technical participants interact with, understand, and contribute to the industry.
Gone are the days where you were completely lost if you couldn’t code.
LLMs like chatGPT bridge the the gap between complex coding languages and everyday vernacular. This is huge because crypto is a field that is largely dominated by people with specialized technical expertise.
If you do not understand something or believe that a project is purposely obfuscating the reality of their underlying systems, you can ask chatGPT and get a quick and near-free response.
DeFi is democratizing access to finance, while LLMs are democratizing access to DeFi.
In today’s post we’ll put forward some of our ideas on how we think the existence of Large Language Models can impact DeFi.
DeFi Security
As we've noted, DeFi is transforming financial services by cutting down friction and overhead costs, and by replacing large teams with efficient code.
We’ve covered at length where DeFi is going. DeFi:
Reduces friction costs - gas fees will eventually come down
Reduces overhead costs as there are no physical locations, just code
Reduces human overhead as you’ve replaced thousands of bankers with 100 coders
Lets anyone provide financial services (such as lending and market making)
DeFi is a leaner operational model and does not rely on intermediaries for execution.
The “counterparty risk” in DeFi is exchanged with software security risk. The code and mechanisms that secure your assets and facilitates your transactions are constantly at risk from outside threats looking to steal and exploit funds.
AI, particularly LLMs, can play a crucial role in automating the development and auditing of smart contracts. By analyzing codebases and identifying patterns, AI (over time) can detect vulnerabilities and optimize smart contract performance, thereby reducing human error and increasing the reliability of DeFi protocols. By analyzing the contract against a database of known vulnerabilities and attack vectors, LLMs can highlight areas of risk.
Autist note: one area where LLMs are already a viable and accepted solution to software security issues is helping write test suites. Crafting unit tests can be tedious, but it’s an important part of software quality assurance which often gets overlooked in rush to market.
However, there is a “darkside” to this as well. If LLMs can help you audit your code, they can also help hackers find ways to exploit your code in the open source world of crypto.
Luckily, the crypto community is full of whitehats and home to a bounty system that helps mitigate some of that risk.
Cybersecurity professionals don’t advocate “security by obscurity”. Instead they assume that the attacker is already familiar with the code and vulnerabilities of the system. Where AI and LLMs can help level the playing field is in automated detection of unsafe code at scale, especially by non-coders. There are more smart contact deployments daily than a human can audit. It’s sometimes necessary to interact with new and popular contracts without waiting for a period of battle testing in order to capture economic opportunities e.g. farming.
This is where platforms like Rug.AI come in, giving you an automated assessment of new projects against known code vulnerabilities.
Perhaps the most revolutionary aspect is LLMs’ ability to assist in writing code. Users with a basic understanding of their requirements can describe what they want in natural language, and the LLM can translate these descriptions into functional code.
This lowers the threshold for creating blockchain-based applications, allowing a broader range of innovators to contribute to the ecosystem.
It’s early days for this. We’ve personally found LLMs to be more useful at refactoring code, or explaining what code does for beginners, than greenfield projects. It is important to give your model context and a clear specification - otherwise expect “garbage in, garbage out”.
LLMs can also help those who can’t code by translating smart contract code into natural language. Maybe you don’t want to learn how to code, but you do want to make sure the code running the protocol you use does what the protocol promises.
While we doubt LLMs can *replace* a high quality developer any time soon, developers can get another round of sanity checks on their work through LLMs.
Conclusion? Crypto gets a little bit easier and safer for all of us. Just be careful not to become reliant on these LLMs. They can be confidently wrong. The ability of LLMs to fully understand and predict the implications of code is still developing.
Data Analysis and Insights
If you try to collect data in crypto, you will inevitably come across Dune Analytics at one point or another. If you haven’t heard of it, Dune Analytics is a platform that allows users to create and publish data analytics visualizations primarily focused on the Ethereum blockchain and other related blockchains. It's a useful and user friendly tool for tracking DeFi metrics.
Dune Analytics already has a GPT-4 enabled feature that explains queries to you in natural language.
If you’re ever confused by a query or want to create and edit one, you can turn to chatGPT. Note that it’s going to perform better if you feed it some example queries in the same conversation, and you’d still want to learn it yourself so you can verify chatGPT’s work. However, this is a great way to learn as you go as you can question chatGPT as if it was a tutor.
LLMs are massively reducing the barriers to entry for non-technical crypto participants.
As for insights, LLMs have been disappointing in terms of the unique insights they can come up with. In the complex, cerebral world of financial markets, do not expect LLMs to give you the right answer. If you’re someone who operates based on instinct and gut, you’ll find that LLMs fall well short of your expectations.
However, we have found one effective use for them when it comes to insights—checking to see if you missed something obvious. You’re unlikely to find non-obvious or contrarian insights that actually generate returns. This shouldn’t come as a surprise (if someone made AI that generated superior market returns, they would not release that part to the broader public).
Death of the Discord Mod?
One of the most thankless and painful jobs in crypto is managing a discord of ravenous degens for a popular project. Many of the same common questions are asked over and over, some times even back to back. This feels like a pain point that should be easily resolved through the use of LLMs.
LLMs are also showing some reasonable accuracy in detecting whether a message is self-promotion (spam). We expect this could also be used to detect malicious links (or other hacker behavior). It’s hard to police a very busy discord with thousands of active members posting regularly, so we look forward to some LLM-backed Discord bots.
Out There Stuff
One of the recurring memes in the crypto space is the launch of coins off of popularized memes. These range from stickier memes like DOGE, SHIB, and PEPE, to random coins that die off in an hour based on a trending search term of the day (largely scams—we avoid).
If you could afford access to the Twitter Firehose API you could track crypto sentiment in real time and train an LLM to flag trends, then use a human to interpret the nuances. A crude example of how it could be used is that you could then launch memecoins in accordance with your sentiment analysis when there are viral moments.
There may be ways to build a poor man’s version of a sentiment scraper which monitors a subset of popular crypto influencers across multiple social media channels without needing to deal with the cost and bandwidth of consuming a “firehose” type API feed.
LLMs are uniquely fit for this because they can get deep contextual understanding (necessary to parse the sarcasm and trolling online from real insights). This LLM partner would evolve and learn alongside the crypto sector where the bulk of the action is discussed on Crypto Twitter. Crypto, with its open debate forum and open source tech, is uniquely set up for LLMs to capture market opportunities.
But the technology would need to be much more sophisticated to avoid being fooled by deliberate social media manipulation: astroturfing, undisclosed paid sponsorships, and troll farms. In another post we covered an interesting third party research report suggesting that some entities may have intentionally manipulated social media to increase the value of crypto projects associated with FTX/Alameda.
NCRI analysis shows that bot-like accounts comprised a substantial proportion (about 20%) of online chatter mentioning FTX listed coins.
This bot-like activity forecasted the price of many FTX coins analyzed in the data sample.
After promotion by FTX, activity for the coins grew increasingly inauthentic over time: The proportion of inauthentic, bot-like comments steadily grew to approximately 50% of the total chatter.
Now that you have an understanding of how AI is going to make crypto so much more accessible, it’s probably a good idea to subscribe to the more in depth analysis we are putting together on AI projects and the broader narrative.
Until next time..
Until next time…
Disclaimer: None of this is to be deemed legal or financial advice of any kind. These are opinions from an anonymous group of cartoon animals with Wall Street and Software backgrounds.
Hi Jungle folks, can you do a primer on how to get started on SOL ecosystem, there seems to be quite some activity and airdrops, thoughts in general?
Great article as always, thanks guys. The example you give of a scraper to grab real time sentiment has Swan written all over it!!