There are some things that do worry me and it’s not the hype around taking jobs away (really those in danger are ones that are always in danger – Buzzfeed writers notwithstanding). Some of it comes from me testing out the platform, some from reading others, and some just inherent in all AI.
Let’s dig in.
No Such Thing as a Free Lunch | Open AI launched in November 2022 and starting in May 2023, use of the API will cost you $20 a month. Competitors are following suit with Microsoft charging an enhanced Teams for additional license fees. Hopefully, we don’t speed run Cory Doctrow’s principle – “Here is how platforms die: First, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die.”
Confidently Average and Sometimes Even Wrong | The model can lie to you with as much confidence as it can tell the truth. Note that ChatGPT stopped indexing data in 2021, so recent events aren’t included.
In my first trial, I asked about x-ray binaries (I have an inhouse MIT Physics PhD from to verify it). ChaptGPT was incorrect with basic description. I asked for sources, and it gave me the same one three times which only generally referenced with work. It provided repetitive text when asked follow-up questions. It seemed to correlate differing but opposing viewpoints into the same response. It just didn’t seem to understand deeper context. Two weeks later I read a Medium article where a physicist had many of the same issues I had.
I later asked it to write an article on cyber enterprise risk and to point out future improvement. It did give me a good summary at a high level. It’s future improvements I’m pretty sure came from a MITRE webpage on the ATT&CK framework. It made up references, page numbers, and quotes to papers that don’t exist (Psst – teachers, this is a way to get around AI generated papers by forcing students to quote passages on specific pages). The summary I’d say was at a high school level – nothing too deep, generic, but not wrong.
My overall worry here is that it was confidently wrong. Laymen wouldn’t be able to point out what was wrong – ChatGPT created a false feeling of expertise. In people terms, this is the Dunning-Kruger effect which occurs when a person’s lack of knowledge and skills in a certain area cause them to overestimate their own competence.
New/Innovation | ChatGPT can generate working code, IF there is a large index of fully functional code for it to start from. I put it to the test and asked it to write a basic cloud formation template for deploying a Kubernetes cluster. It started fine, then started loop of repeating script, and just ended. As a senior coder, it could cause more work especially in edge cases and places where code hasn’t necessarily been written before.
The mission of OpenAI is “ensuring that artificial general intelligence benefits all of humanity.” I want that to be true.
Stay tuned for Part 4.
