A recently released paper from Apple that has set off a firestorm within the AI community. So it is all of AGI today, and to its future now. The study serves to illustrate how the industry is still figuring out its way through the challenges of large language models (LLMs). These models fail miserably, particularly when it comes to tackling big, hairy, audacious problems. Prominent leaders in the field are already hailing this surprising announcement. Among them are Andrew Rogoyski of the Institute for People-Centred AI at the University of Surrey and Gary Marcus, a well-known critic of the rush to AI.
As mentioned in this recent Apple caucus paper, this is a tragic and disturbing trend in LLM performance. These complex models frequently fail to even solve a problem. This occurs, shockingly enough, even when they have algorithms purpose-built to prevent that from happening. As issues grow in difficulty, models are often still forced to first search in the wrong direction before “learning” how to find the right solutions. While potentially harmless, this behavior is deeply troubling, as it casts doubt on what LLMs, which power heavily marketed tools like ChatGPT, can actually do.
Andrew Rogoyski highlighted the significance of the Apple paper, noting that it flies in the face of many existing assumptions about what LLMs can do. He stated, “These insights challenge prevailing assumptions about LRM capabilities and suggest that current approaches may be encountering fundamental barriers to generalisable reasoning.”
Influential AI researcher and entrepreneur Gary Marcus was sympathetic to this sentiment, calling the conclusions of the Apple paper “pretty devastating.” His caution: if you think LLMs are the simplest possible path to transformative AGI, you’re wrong. “Anybody who thinks LLMs are a direct route to the sort [of] AGI that could fundamentally transform society for the good is kidding themselves,” Marcus asserted.
The Apple white paper also argues that the industry may be approaching a “cul-de-sac” in its approaches to AGI. It suggests a “fundamental scaling limitation in the thinking capabilities of current reasoning models.” This constraint becomes especially evident with wicked issues. As the going gets more demanding, models tend to decrease their reasoning stride. The paper notes, “Upon approaching a critical threshold – which closely corresponds to their accuracy collapse point – models counterintuitively begin to reduce their reasoning effort despite increasing problem difficulty.”
These findings have inspired more serious scrutiny of the race to AGI. They warn that experts are concerned that if the current trajectory continues, it won’t lead to the outcomes they want. The implications for AGI development are pretty darn profound. This will require a major reevaluation and innovation to break through the barriers that exist today.