Gaston Paris / Contributor
A study that had been overlooked for several months was recently featured in a magazine. wired They claim to have mathematically proven that large-scale language models “cannot perform computational or agentic tasks beyond a certain level of complexity.” Crucially, its level of complexity is fairly low.
The paper, which has not been peer-reviewed, was written by Vishal Sikka, former chief technology officer of German software giant SAP, and his son Valin Sikka. Sikka-senpai has some knowledge about AI. He studied with Turing Award-winning computer scientist John McCarthy. McCarthy literally founded the entire field of artificial intelligence and actually helped coin the term.
“There’s no way they can be trusted,” Vishal Sikka said. wired.
In response to an interviewer’s question, Sikka also agreed that we should forget about AI agents running nuclear power plants and other tough promises thrown at us by AI boosters.
Ignore the rhetoric that tech CEOs spew on stage and focus on what the researchers working for them are discovering. You’ll see that even the AI industry agrees that this technology has fundamental limitations built into its architecture. For example, in September, OpenAI scientists acknowledged that AI illusions, in which LLMs confidently fabricate facts, are still a widespread problem in increasingly sophisticated systems, and that models will “never” reach 100 percent accuracy.
This appears to be a major blow to the viability of so-called AI agents, models designed to perform tasks autonomously without human intervention, which the industry widely decided last year would be the next big thing. Some companies that hired AI agents to reduce their workforce quickly realized that the agents were not good enough to replace the departing humans. Perhaps it’s because they hallucinate so badly that they can barely complete any given task.
AI leaders argue that stronger guardrails outside of AI models can eliminate hallucinations. They may be prone to hallucinations all the time, but if such blunders are rare enough, companies will eventually begin to trust them to do the work they once left to flesh-and-blood thugs. In the same paper in which they acknowledged that their models will never reach perfect accuracy, OpenAI researchers also rejected the idea that hallucinations are “inevitable” because LLM “can be avoided in cases of uncertainty.” (Nonetheless, you’d be hard-pressed to find a single popular chatbot that actually does that, as it would almost certainly make the chatbot less impressive and difficult to use.)
Sikka is adamant that LLM has hard limits, but agrees with those in the AI industry who say hallucinations can be suppressed.
“While our paper says that pure LLM has this inherent limitation, it is also true that we can build components around LLM that overcome those limitations,” he said. wired.
Learn more about AI: It looks like OnlyFans’ rival has succumbed to AI psychosis. Please try explaining this to your parents.
