I recently ran across an article on the Seeking Alpha investing site with the provocative title “ AI: Fakes, False Promises And Frauds “, published by LRT Capital Management. Obviously, they think the new generative AI is being oversold. They cite a number of examples where demos of artificial general intelligence were apparently staged or faked.  I followed up on a few of these examples, and it does seem like this article is accurate. I will quote some excerpts here to give the flavor of their remarks.

In 2023, Google found itself facing significant pressure to develop an impressive innovation in the AI race. In response, they released Google Gemini, their answer to OpenAI’s ChatGPT. The unveiling of Gemini in December 2023 was met with a video showcasing its capabilities, particularly impressive in its ability to handle interactions across multiple modalities. This included listening to people talk, responding to queries, and analyzing and describing images, demonstrating what is known as multimodal AI. This breakthrough was widely celebrated. However, it has since been revealed that the video was, in fact, staged and that it does not represent the real capabilities of Google’s Gemini.

… OpenAI, the company behind the groundbreaking ChatGPT, has a history marked by dubious demos and overhyped promises. Its latest release, Chat GPT-4-o, boasted claims that it could score in the 90th percentile on the Unified Bar Exam. However, when researchers delved into this assertion, they discovered that ChatGPT did not perform as well as advertised.^[10] In fact, OpenAI had manipulated the study, and when the results were independently replicated, ChatGPT scored on the 15th percentile of the Unified Bar Exam.

… Amazon has also joined the fray. Some of you might recall Amazon Go, its AI-powered shopping initiative that promised to let you grab items from a store and simply walk out, with cameras, machine learning algorithms, and AI capable of detecting what items you placed in your bag and then charging your Amazon account. Unfortunately, we recently learned that Amazon Go was also a fraud. The so-called AI turned out to be nothing more than thousands of workers in India working remotely, observing what users were doing because the computer AI models were failing.

… Facebook introduced an assistant, M, which was touted as AI-powered. It was later discovered that 70% of the requests were actually fulfilled by remote human workers. The cost of maintaining this program was so high that the company had to discontinue its assistant.

… If the question asked doesn’t conform to a previously known example ChatGPT will still produce and confidently explain its answer – even a wrong one.

For instance, the answer to “how many rocks should I eat” was:

…Proponents of AI and large language models contend that while some of these demos may be fake, the overall quality of AI systems is continually improving. Unfortunately, I must share some disheartening news: the performance of large language models seems to be reaching a plateau. This is in stark contrast to the significant advancements made by OpenAI’s ChatGPT, between its second iteration (GPT-2), and the newer GPT-3 – that was a meaningful improvement. Today, larger, more complex, and more expensive models are being developed, yet the improvements they offer are minimal. Moreover, we are facing a significant challenge: the amount of data available for training these models is diminishing. The most advanced models are already being trained on all available internet data, necessitating an insatiable demand for even more data. There has been a proposal to generate synthetic data with AI models and use this data for training more robust models indefinitely. However, a recent study in Nature has revealed that such models trained on synthetic data often produce inaccurate and nonsensical responses, a phenomenon known as “Model Collapse.”

OK, enough of that. These authors have an interesting point of view, and the truth probably lies somewhere between their extreme skepticism and the breathless hype we have been hearing for the last two years. I would guess that the most practical near-term uses of AI may involve some more specific, behind the scenes data-mining for a business application, rather than exactly imitating the way a human would think.