Are we close to reaching superhuman artificial intelligence? Investors are losing patience

A new and contentious public debate has begun: is ChatGPT an ‘artificial general intelligence’ (aka ‘AGI’)?

Opening for the proposition is Elon Musk, who has argued in a lawsuit against OpenAI that ChatGPT is at least as intelligent as a human being, placing it in the category of AGI. In opposition is OpenAI and Microsoft, which stands to lose its investment in OpenAI if the proposition wins, as OpenAI’s agreement with Microsoft grants them a share of OpenAI’s revenue only until they build AGI, at which point profits are to be distributed to the public good.

The public, along with the American courts, is now invited to weigh in on the questions: Does the fact that ChatGPT can pass multiple tests created for humans at or above the average level qualify it as having human level intelligence? Can AI be considered ‘human-like’ in its intelligence when it still has errors in accuracy and reliability, given that humans also don’t have perfect accuracy or recall?

Regardless of what side the reader comes down on in the Musk-Microsoft debate, it’s clear that AI models don’t yet understand human concepts and human intent. Technologically, this is not surprising. These models are constructed as predictors trained on datasets, and regardless of how they compare to humans on tests, do not generate outputs, whether they be text, images, or actions, out of conceptual understanding.

This debate follows closely on the heels of Google’s Gemini disaster, in which Gemini generated images of German SS soldiers as multi-racial, thus visually erasing the horrors of genocide from the history of the Holocaust. This was a consequence of Google including in system prompts that humans depicted should be multi-racial, in an attempt at ‘aligning’ their AIs. Gemini followed its instructions rather than understanding and implementing their intent.

Behind the doors of R&D offices at major enterprises, a different narrative is unfolding than the one Musk has brought to the media. C-suites earmarked large investments in GenAI projects with high hopes, building on the promises of ChatGPT’s hype. Exciting prototypes and proofs of concepts were developed. Yet, one by one, enterprises are finding it challenging to eke out high enough levels of reliability and accuracy from these systems to deploy them autonomously into real-world use.

Whispers of a possible coming ‘AI winter’, a period of decreased interest and investment in artificial intelligence, already haunt the halls of venture capital. Intelligence as inaccurate and unreliable as the average human being is not sufficient for building robust software products.

While the question of the definition of ‘AGI’ is fascinating, the more urgent question we have facing us is actually whether AI is fit for use as is. The attention of the industry, and the public, should remain stalwartly on improving the reliability and safety of these systems. AI’s value, after all, is in what it can contribute of value to human lives.

Rebecca Gorman is founder and CEO of Aligned AI