- AI agents are popular.
- But they often make mistakes when they are asked to navigate the cryptography market.
- A company is trying to reduce errors and improve reliability.
It was a simple request.
Nick Emmons, co-founder and CEO of Alora Labs, a company building a decentralized artificial intelligence network, asked a new AI agent he was testing to exchange a crypto in American dollars for him.
Although he gave the program for explicit instructions, he began to negotiate an active ingredient completely different from what was requested.
“He started completely from the rails and did something completely unrelated to what he was initially directed towards doing,” said Emmons DL News. This type of aberrant behavior among agents is too common, he said.
AI agents
AI agents, autonomous software designed to achieve specific objectives without constant human supervision, are seated at the forefront of the booming sector.
In the cryptography industry, AI -focused companies have collected more than $ 500 million this year, with many vomiting agents that can analyze potential investments, manage cryptographic asset portfolios and even trade for users.
There is only one clip.
When these agents receive real money and put situations live, things often run badly.
“There is an infinite set of possibilities for capital management to be mistaken,” said Emmons. “They could lose it completely. They could put it in bad assets. They can misinterpret digital inputs to make incorrect financial decisions, all kinds of things. ”
This is a problem, because the initiates of the industry strongly believe in technology.
Google and Microsoft technology giants channel resources to create their own AI agent platforms.
A recent survey of IT leaders has revealed that 93% report that their organizations are already developing – or plan to develop – their own versions of technology, according to a July Outsystems report, a coding platform powered by AI.
The AI agent market is expected to exceed $ 50 billion over the next five years, according to Boston Consulting Group.
With so much interest, anyone who can solve current problems with AI agents should greatly benefit from this growth.
Great language models
According to Emmons, the reason for the problems with AI agents is that most of them rely on large -language or LLM models.
“Llms hallucinously elegantly elegantly most of the time,” said Emmons. “When you are dealing with digital or quantitative parameters, these hallucinations can cause very extreme errors.”
Other specific AI agent problems include excessive dependence on historical data, perpetuation when market conditions change and does not take into account liquidity and shift, according to amplework, a development and consultation company of AI.
AI agents have also been shown to collusion with each other and engage in anti -competitive practices such as prize fixing, according to a recent Wharton School study of the University of Pennsylvania and the University of Sciences and Technologies of Hong Kong.
The Emmons company, Allora, compensates for LLM traps with traditional automatic learning via its decentralized AI network. This said Emmons, allows AI agents to benefit from the things that LLMs are good while reducing errors and hallucinations.
“It is a question of determining the good marriage between these two somewhat distinct technologies,” said Emmons.
Allora already uses its network in decentralized finance.
It has live applications actively managing liquidity on Uniswap, the greatest decentralized exchange, and engaging in the loop, a kind of lever -effect loan strategy which juices the yield users can win by stimulating Ethereum.
Have a human?
But even with the reduction errors of the AI of Allora network, there are still risks.
Emmons said that there should also be stricter parameters to ensure safety. “We need the portfolios we equip agents to have an even more specific set of function and function calls so that they cannot simply throw money.”
Certainly, it is not as if human traders were infallible. Between 2006 and 2008, former general company trader Jérôme Kerviel lost around $ 7.2 billion in his employer’s money by placing a series of unauthorized and high issues.
There is also a debate on the question of whether AI agents can never act fully without human supervision.
An article in 2024 published by Google Deepmind Researchers argued that AI agents should be able to engage in causal reasoning to function properly, which, according to them, is not possible currently.
Emmons, on the other hand, is more optimistic.
“Many AI agent stuff will be enough without tips,” he said. “We are here or very close.”
Tim Craig is the DL News -based correspondent, based in Edinburgh. Handle with advice Tim@dlnews.com.
