Robopsychologist
In another great Two Minute Papers episode, they discuss a recent paper on how ChatGPT might show some early displays of general artificial intelligence. You can check out the episode and the paper it discussed here:
The episode and article try to understand where GPT4 flops and by doing so, invoke Isaac Asimov’s robopsychologist who was tasked with understand how robots think in a few of his texts. Many of those working with LLMs are trying to do the same.
Why can this technology perform so well as some tasks but completely flop at others, including basic arithmetic.
I’ve brought up this story a few times but it highlights how much intelligence isn’t a single dimension but many different areas, some of which can go incredibly deep and some can remain shallow: when we launched the Ubi many years back, someone asked the device Kevin Costner’s age and it was off by a year. It didn’t matter that the wake word functioned, the speech recognition worked flawlessly, the natural language understanding was able to identify that the person was asking a question, fetch a response, structure the response as a sentence, and then use speech synethesis to generate a sound that mimicked a human voice. No! The device was “stupid” because it got the age wrong.
By trying to piece together where the technologies excel, we can undrestand what tasks can be offloaded to it more reliably. This may be the largest crowdsourced project in humanity’s history.
