OpenAI’s New AI Model, o1, Makes Basic Errors in Demo “`

OpenAI unveiled its most advanced AI model, o1, to paying subscribers on Thursday, initiating its “” holiday event featuring twelve consecutive releases.

OpenAI promoted o1’s “complex reasoning,” offering unlimited access for $200 monthly. A demonstration showcasing the model’s capabilities involved a user requesting instructions to build a birdhouse from an image. The model seemingly provided comprehensive instructions.

However, closer inspection revealed the instructions to be largely unhelpful. Measurements for paint, glue, and sealant were given in inches. Only the front panel dimensions were provided. Sandpaper dimensions were inexplicably included. Furthermore, a section promising “exact dimensions” offered none.

James Filus, director of the Institute of Carpenters, criticized the instructions in an email, stating they offered no more practical guidance than the image itself. He pointed out the omission of a hammer from the tool list despite nails being included, and noted that the AI’s estimated building cost was far lower than reality. He also highlighted the insufficient detail regarding the installation of a hinged roof opening.

OpenAI did not immediately respond to a request for comment.

This incident adds to a pattern of AI product demos backfiring. Last year, an AI-assisted search tool inaccurately reported a James Webb telescope discovery, impacting the company’s stock. More recently, a Google tool provided unsafe instructions involving cheese and pizza.

OpenAI’s o1, considered its most capable model, uses a “chain of thought” reasoning process unlike ChatGPT. While fundamentally a sophisticated next-word predictor trained on vast text data, it “thinks” before responding. This often improves accuracy, and OpenAI highlighted o1’s reasoning abilities, particularly in math and coding—reporting 78% accuracy on PhD-level science questions based on September data.

Nevertheless, fundamental logical flaws clearly persist.