ChatGPT Opens A Research Lab…For $2!

Here is a crazy idea: Let’s slice up a brain into  small pieces. Okay, let’s rewind a little. First,   let’s not just use ChatGPT, but use it to create  a full research lab. Now, wait a second. That is   of course, impossible. A research lab requires  the work of several people. How would we do that   with just ChatGPT? Well, here is an even crazier  idea: let’s create several copies of ChatGPT,   and ask them to pretend to be the professor,  a PhD student, a software engineer, and more.    Is that possible? Well, have a look at this.   Previously, it worked with video game characters. They made 25 ChatGPT agents,  gave them motivations and memory,   and put them in a simulated town. They wake  up, start their morning routine, some of them   like reading papers, I approve that. And they  held elections too. Wow, what happened there? Well, Tom, one of the characters said: “To  be honest, I don’t like Sam Moore. I think   he’s out of touch with the community and  doesn’t have our best interests at heart. ”  Oh my…I wish the average voter sounded like that  too. So, I show you this because they can have   relationships with each other, and even help each  other. So, in this new work, they made a research   lab to work on tough research questions. And  it did not go as I expected it. Not at all. Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. So it has its own research staff, yes, but that’s  not where we start. Look. We start with a human   putting in an idea. Very important detail. It  starts with you. This idea can be, for instance,   do biases affect the performance of language  models on benchmarks? Now, a PhD student,   which is played by ChatGPT, is asked look up  if the idea has already been solved before,   then a more experienced postdoc researcher,  another ChatGPT is involved with making a plan,   and finally, two more little AIs are  asked to code the whole thing up. Note that there are plenty of moving parts  here to make sure we get a quality work. So,   then, what happened? Well,  absolute magic happened.   The concept wins medals left and right. It  outperforms previous techniques on a variety   of tasks, just make sure to not ask about  anything that is in russian for some reason. Now…about that brain. I think this is absolutely  incredible. It is like taking a huge brain and   instead of using just that, we slice it up  into smaller little brains, and that somehow   does better. This is insanity. What a time  to be alive! Yes, this is Two Minute Papers,   the corner of the internet where we slice  brains up and we are even happy about it.    So how much does the slicing cost? How much  do we need to pay our little AI researchers? And now, hold on to your papers Fellow Scholars,  because this new technique can do all this,   for the grand sum of…2 dollars and 33 cents.   And it is done in 20 minutes. And if you wish   to plug in these fancy new thinking AIs,  they do the literature review a bit better,   I love seeing that nice gradient downwards here.   This takes a bit longer and costs a bit more,   but even if you wish to feel like a king and go  all out, you pay about $13, and in this case,   it takes 1. 5 hours. For proper research work with  an implementation, experiments and more. Crazy. And whenever I don’t have the resources to  do that, I just rent a GPU on Lambda and do   it myself. By the way, the full code and  paper is available for free for everyone,   so this is open science, loving  it. And there is a more detailed   user study in the paper, the  link is in the description. And this, Fellow Scholars, is how I  see the future of AI. And that is,   empowering human thought. This is meant to  help you with time-intensive repetitive tasks,   but ultimately, you are in charge. Now one super important question remains.    Can these AIs really invent  something fundamentally new? Let’s see…well, their ideas tend to be more  novel than ideas of humans, that is good,   more exciting even. Super good. But here come the  bad news for the AIs: their ideas were evaluated   to be less feasible at the same time. I mean, it  is easy to come up with super exciting things,   teleportation, time machines, anything you want.   But ultimately the ideas have to be practical too! So I would give this one to humans. And that is  why the idea has to start with you. For instance,   the Nobel-prize winning AlphaFold was not just  about throwing an AI at protein folding by far.

No-no-no. It required the ingenuity of  tons of brilliant research scientists,   and the interplay of many many  moving parts. So the AI does not   invent fundamentally new things  unless it meets human brilliance. So what do Fellow Scholars think?   Let me know in the comments below.