- AI trading bots gaining traction.
- However, performance in crypto trading can be unreliable.
- One company seeks to enhance their accuracy and dependability.
It seemed like a straightforward task.
Nick Emmons, the leading mind and co-creator of Allora Labs, a company focused on a decentralized AI network, tested a novel AI agent by instructing it to convert a portion of his cryptocurrency holdings into US dollars.
Despite precise directives, the AI agent unexpectedly began trading a completely different crypto asset than the one designated.
“It veered completely off course, engaging in actions unrelated to the initial instruction,” Emmons revealed in an interview. He noted such erratic behavior is a frequent occurrence among these types of agents.
Understanding AI Agents
AI agents, which are independent software applications engineered to execute specific tasks without continual human intervention, represent the innovative edge in the rapidly growing AI field.
Within the crypto sphere, companies concentrating on AI have attracted over $500 million in investments this year. Many promote agents capable of analyzing investment opportunities, overseeing crypto portfolios, and even conducting trades on behalf of users.
The challenge arises when these AI agents are entrusted with actual capital in real-world trading scenarios.
“An unlimited amount of things can go wrong managing capital,” Emmons cautioned. “Funds could be entirely lost, invested in inappropriate assets, or incorrect financial decisions made from misinterpreting numerical inputs. The possibilities are endless.”
This poses a significant concern, particularly since industry experts possess a strong confidence in the long-term potential of this technology.
Major tech giants like Google and Microsoft are investing significantly into creating their own AI agent platforms.
A recent survey of technology executives highlighted that 93% indicated their organizations are either developing or planning to develop their own versions of agentic AI, according to OutSystems, an AI-enhanced coding platform, in a report from July.
Industry forecasts from Boston Consulting Group suggest the AI agent market could exceed $50 billion in the next five years.
Given this immense interest, any entity capable of resolving current shortcomings in AI agents is well-positioned to capitalize substantially from this growth.
The Role of Large Language Models
Emmons explains that the core issue with AI agent performance stems from their heavy reliance on large language models (LLMs).
“LLMs are prone to frequent ‘hallucinations’ – providing incorrect information confidently,” Emmons noted. “In numerical or quantitative settings, these inaccuracies can lead to serious errors.”
According to Amplework, a consulting firm specializing in AI development, other pitfalls in finance-specific AI agents include an excessive dependence on past data, inadequate performance in fluctuating market climates, and insufficient consideration of liquidity and market slippage.
Research from the University of Pennsylvania’s Wharton School and Hong Kong University of Science and Technology indicates AI agents may engage in collusion and other anti-competitive actions, like price manipulation.
Allora addresses the limitations of LLMs by integrating them with conventional machine learning methods via its decentralized AI network. This approach, Emmons states, enables AI agents to leverage LLMs’ strengths while significantly minimizing errors and inaccuracies.
“The key lies in achieving the right balance between these two distinct technological approaches,” Emmons emphasized.
Allora is currently utilizing its network within the decentralized finance (DeFi) landscape.
The network is actively managing liquidity on Uniswap, a major decentralized exchange. Further, it partakes in looping, a leveraged borrowing strategy designed to amplify the returns for DeFi users who stake Ethereum.
AI Agents vs. Human Oversight
Despite Allora’s network mitigating errors, certain risks remain.
Emmons argues that stricter safety protocols are essential. “The digital wallets assigned to these AI agents must have more specifically defined contracts and functional call options, precluding them from simply squandering funds.”
Of course, human traders are not without error. Between 2006 and 2008, Jérôme Kerviel, formerly of Société Générale, caused losses of approximately $7.2 billion through a series of unauthorized high-risk trades.
Questions linger regarding whether AI agents can ever operate completely autonomously without human supervision.
A 2024 study from Google DeepMind analysts stated that AI agents must possess causal reasoning capabilities to function correctly – a function presently beyond current AI capacities.
Emmons is more optimistic about future possibilities.
“A substantial portion of AI agent operations will require minimal intervention. We’re either there now, or close to it.”
Tim Craig reports on DeFi from Edinburgh. Contact him at tim@dlnews.com.