*This text is from our interview published on You Tube.
The Big Whale: For several weeks now, blockchain-based AI (artificial intelligence) agents have been much talked about. How can you explain this concept simply?
Colas Gabriac: The AI agents that are currently emerging are based on LLMs (Large Language Models). There are different types of models: proprietary models, such as ChatGPT, which is widely used by the general public, and open source models. With the latter, users can modify, customise and transform the models to suit their needs. These agents can thus develop their own personality, making their interaction particularly interesting for users.
One of these agents, Truth Terminal, recently became an autonomous crypto-millionaire...
In fact, what began as a simple experiment with the Truth Terminal account was based on an agent trained through exchanges between two instances of the Claude LLM model, developed by the Anthropic company. These exchanges were supervised by a human, Andy Airey, the creator of Truth Terminal.
The agent was then left autonomous, first on Twitter, where it could interact with anyone and share its ideas. Later, he announced his intention to create a token. As he was unable to do this on his own, a member of the community took charge. Today, this token has a valuation of over a billion dollars.
Truth Terminal even exchanged with Marc Andreessen, one of the two co-founders of the famous a16z fund, who donated $50,000 in bitcoins at his request. Due to the agent's technical limitations, it was Andy Airey who received the funds.
Can we imagine, in the future, an AI agent capable of managing a venture capital fund or being active on platforms such as Aave or Morpho, for example?
This already exists. AI agents now act as VCs (venture capitalists) or manage financial strategies. For example, there is the AI16z token, inspired by the famous venture capital fund.
This token essentially works like a VC fund managed by an AI. It also has communication channels, notably on Telegram, where users can try to convince the AI to buy specific tokens. The AI then tries to optimise the performance of the funds it manages, which are currently worth between $3 million and $6 million. However, this agent has not yet made any major investments.
Other more creative projects, such as Goat or Zerebro, are emerging. For example, Zerebro acts on several platforms such as Twitter, Instagram and Telegram. It has produced an album available on Spotify and created NFTs on its own. This project focuses on artistic and technical aspects, and although its token has no direct use, it is generating interest.
For the moment, the majority of users are buying these tokens to speculate, a bit like memecoins.
These agents could be useful for managing financial strategies. When might we expect to see them in common use in finance?
I think this could happen quickly. Most of these agents have been created by individuals working alone, often from their bedrooms. This trend is still extremely new - it's barely a month old in most cases.
In crypto, that may seem like a long time, but for the development of new blockchain projects, we're still in our infancy.
We haven't yet seen large-scale AI agents backed by significant funding or developed over several years. But even at this early stage, the possibilities are immense. These agents can already perform tasks such as creating tokens, buying them, analysing blockchain addresses, and even suggesting investment strategies, such as placing funds on Aave.
In my opinion, their development will accelerate rapidly, thanks in particular to LLM-type languages. These agents excel at language understanding, data reading and even code interpretation.
AI agents already exist outside of blockchain. Why is blockchain particularly interesting for their development?
What has really attracted the attention of the crypto ecosystem towards these agents is their tokens, which are often treated as memecoins at the moment. These tokens attract for several reasons: the sense of community they create, their artistic or humorous value, and their speculative aspect.
The blockchain allows a unique community link to be forged. This link can also be used to train these agents in innovative ways.
Also, blockchain offers the advantage of self-custody, allowing agents to have their own portfolios and become financially independent. Of course, these do not arise spontaneously in nature; they are always created and maintained by humans.
However, with blockchain, these agents can manage real money and concrete value, which is unheard of.
I don't expect AI to outperform the best venture capitalists, but it could prove useful for those who lack expertise in financial management.
In crypto, the majority of individuals lose money managing their portfolios. Entrusting this task to an AI, which applies simple, well-trained strategies, could be a better option for them.
Isn't it too early to delegate tasks such as cash management or strategies to an AI? For companies, it seems risky...
We're still in the early stages, but things are moving fast, not least because AI is already widely used in finance, particularly in areas such as high-frequency trading, where speed is crucial.
These systems are often small AI applications, not LLMs (Large Language Models), which are not designed for fast execution. That said, the complexity of financial markets makes it unlikely that AI will outperform humans in managing short-term investments. On the other hand, AI can assist humans in decision-making.
For example, in venture capital, it is already being used. I know crypto VC investors who use AI tools like Claude or ChatGPT to analyse proposals: they copy and paste a company's pitch into the AI, ask questions, or even analyse a startup's code to check its authenticity.
AI is therefore a useful tool right now, but as far as managing money autonomously is concerned, it remains an experimental phase.
Are we not overestimating the long-term potential of these AI agents under the hype?
The performance of AI agents on blockchain depends on the progress made by large companies such as OpenAI and Anthropic. These agents cannot exceed the quality of the underlying models provided by these major players. As these models improve, so will the capabilities of AI agents, including on blockchain.
That said, I don't think this trend is just a fad. AI is already well established and continues to make steady progress. For example, it's been two years since ChatGPT became available to the general public, and it has already transformed the way people work.
Scientists, professionals and many others are using AI to increase their productivity. As AI improves, its applications will continue to expand. I think this will also be the case for AI agents on blockchain.
In the coming months, how do you see this trend evolving in the crypto ecosystem?
In crypto, the total valuation of AI-related tokens remains relatively low, at less than $3 billion. This is nothing like the memecoin market, which nevertheless offers less intrinsic value.
Although the current integration of AI and blockchain seems experimental, I think it's just the beginning.
As the big AI companies grow and their technologies improve, the impact on blockchain and crypto will also grow. The two trends will reinforce each other over time.
How do you identify a good AI agent project today?
At the moment, there are so many new projects that it is difficult to distinguish what is truly innovative from what is just a copy of existing ideas.
A key point is to check whether the project brings real originality or is just a clone. For example, there is an open source framework for creating AI agents, offered by AI16z on GitHub. This allows anyone to deploy their own agent on platforms such as Twitter, Instagram or Telegram.
Because of this, many projects use the same tools without offering any real differentiation. I'd advise focusing on agents that stand out-those that are pioneers in their field or bring something special to the table.
For example, the GOAT token isn't particularly interesting from a technical point of view; it attracted attention simply because it was one of the first AI agent tokens. Similarly, Truth Terminal owes much of its popularity to its large Twitter following.
Are metrics like engagement key indicators?
Exactly. Look at how many people interact with the AI agent. Is it generating interest beyond the crypto community? For example, Truth Terminal caught the attention of Marc Andreessen, who talked about it in podcasts. This kind of visibility and engagement is an indicator of a project with broader appeal.
In addition, assess whether the project operates on multiple platforms or has unique features, such as the creation of NFTs, on-chain trading or a distinct artistic aspect. These differentiating features increase the chances of success and attract a loyal community, which can enhance the token's value.
Today, how are these agents supervised? Truth Terminal is not, for example, capable of managing its assets itself.
Most AI agents are still largely centralised in the way they operate. For example, these agents generally operate on the servers of their creators, who retain partial control.
Their portfolios, too, are often accessible by their creators. Sometimes the AI has more autonomy, but the person who controls the technical infrastructure retains some access and control.
In the future, it's entirely possible to imagine completely decentralised AI agents. We already have technologies like decentralised storage, decentralised computation and decentralised inference for AI models.
These tools could eventually combine to create fully decentralised AI agents, capable of managing their own money and operating autonomously, without human intervention.
This raises important questions in terms of ethics...
This is indeed a crucial point. At the moment, moderation is provided by the creators to prevent agents from behaving inappropriately, making harmful comments or engaging in threatening actions.
This moderation is all the more important as some of these agents are trained with datasets that include controversial or extreme content. It ensures that the agents remain within acceptable limits when interacting with the public.
As mentioned earlier, these AI agents are tending towards increasing autonomy. The fact that some are already handling real money is revolutionary, but also raises many questions.
For example, how can we guarantee that the AI doesn't lose the funds or that the creator doesn't abuse them? Security is a major concern, especially with language-based AI. If the safeguards are not strong enough, an ill-intentioned person could manipulate the AI into giving away its tokens.
A broader ethical question concerns the decisions the AI might make with its funds. Some researchers are already exploring this topic, seeking to understand how an AI perceives owning financial assets and what it might choose to do with them.
Could an AI use its funds for unethical purposes? It's a possibility, especially with models trained on controversial or problematic datasets.
Monitoring the actions of these agents is crucial: not only to prevent financial mistakes, but also to ensure they don't engage in harmful behaviour. As these systems evolve, the ethical and security challenges will become increasingly complex, requiring clear frameworks to address them.
Are there frameworks in place to train AI agents to reduce these risks?
The training phase is essential and involves putting safeguards in place from the outset. This includes defining the rules for interacting with the environment and processing the data that the AI receives.
However, the challenge also lies in how AI systems are modified or adapted after they have been created. For example, a model such as Truth Terminal is trained on large sets of unfiltered data from the internet.
This could make it susceptible to producing harmful results due to the nature of the data used for its training.
