duzzi said..boardsurfr said..Here's an interesting little ChatGPT experiment: I asked ChatGPT a question that I had previously asked several dozen scientists interviewing for a position in my lab. All of them had at least a B.S., some had masters or doctorates, and all had at least some basic knowledge of chemistry and statistics. The question:
"I have a small molecule with 2 functional groups at the end that can be modified, for example by attaching biotin. The modification reaction (for one group) is 60% efficient, so that 60% of the molecules will have a biotin at this group. What is the expected distribution of products on the molecule with two such groups, with respect to being labeled with biotin?"Of the dozens of people I interviewed, not a single person came up with the correct answer (can you?). Only a few came reasonably close. A few did not understand the question at all, not even after trying to explain it in much more detail.
Here is ChatGPT's response:
This is the correct response. ChatGPT did not only understand the question, but it answered it
better than dozens of college-educated job seekers I posted this question to (and those were usually the top 5 or 10% of applicants, since we typically got >50 applications per job opening).
Even more interesting is that this was the second try at this response. The first response was only partially correct: the bot did not realize that group 2. consisted of 2 distinct populations, labeled at one end or the other, so it have 24% instead of 48% there, and the numbers at the bottom were also mixed up). I did point out the inconsistencies, but got the "Something went wrong" error instead of an answer. I got the response shown above after re-loading the page, and re-generating the response.
Most of the people I ended up hiring were quite intelligent and well educated, and some were outstanding. But not a single one of them got as close as ChatGPT in the first answer, and
none of them managed to get the right answer after I pointed out what was wrong with there initial answer. To me, that's more impressive than ChatGPT passing medical exams, or scoring in the 90th percentile in the (law) board exam. Actually, I find it somewhat scary.
That has been the problem, and my experience, with all the bots I used or worked on (I worked a couple of years with a team developing an AI for music composition). That is: assuming you can formulate a precise question, the only way to know if they give the correct answer is to know the correct answer already.
Sure, it is impressive that you can have a linguistic back and forth with a program, but if you did not know the correct ratios of molecules you would have just to go home with the incorrect ones.
The reality is that unsupervised machine learning on anything that is not just sheer computational math is decades away. And even with computational math of any sort it is questionable unless the objective of the computation is VERY well defined (e.g chess, or go). Right now this wave of chat bots are fun, and can be used to maybe help in a number of applications, but fundamentally there is just a lot of hype from the usual scums in Silicon Valley trying to make a buck.
PS Your example is too trivial when the question is posed unambiguously. It is equivalent to ask: "I have a coin that has 60% chance of showing a head when tossed. What is the probability of seeing 0, 1 or 2 heads for two tosses?" That is question that many people without a little background with elementary probability theory might have difficulties to answer correctly. (Compound probability of independent events is not a simple concept) But it is an absolute triviality for anybody who did or any computational engine of sort like ChatGPT ... unless of course it does not understand the question, and gives the wrong answer, like it did at first.
