ChatGPT and me
Most of the time, it feels like innovation and development are hidden in the background of daily life. Phones get a bit faster, cameras better, more cars have lane assist, a new algorithm makes your feed load 1% faster. ChatGPT is a truly impressive technology and skips a slow plodding development straight to actually useful in regular circumstances.
The first thing I’ll say is people have seen a lot of mistakes generated by GPT 3.5, the free version, and learned to laugh off the possibility that GPT4 could displace many of the aspects of their jobs. I wouldn’t be so sure, and I’d try playing with the best version to see how it can benefit you.
It took me some time to adapt my thinking to prompt when ChatGPT could be useful for me. The best way to think about it is that you have a person working for you full time, and all they do is look up answers to questions you have by scraping the whole internet and comparing and compiling answers. It is also pretty good at math, GPT 3 and 3.5 were bad, but 4 is good. Sometimes it makes mistakes, but you can normally catch those or prevent them by asking GPT to triple check the math. GPT 4 is easily worth the $20 a month.
In my job (synthetic biologist) the way I use I it is for:
1. Get answers to questions I could answer myself with a lot of petty work, which wouldn’t otherwise be worth it.
2. Generate simple code to analyze high-throughput data which I had no real ability to write myself.
For hobbies I mostly use it as a google, and revisit things that I feel like I might be doing in an inefficient way.
An example or two.
For number 1, I had a disagreement with some other scientists in my lab. My evolution experiments were producing weird outcomes. Some context is required to understand. I have a system where I’m mutating a protein which digests a chemical to make nitrogen. The yeast need nitrogen to grow. Theoretically, with mutation and selection the protein should get better and better. In my experiments it didn’t really get better, often it got much worse.
I had my explanations for that, namely that this protein is one of many that is hard to evolve. There was literature data to support that. I had rigorously tested the system and had good controls in my experiments. Everyone else insisted no, there’s something wrong with your system. The yeast must be getting nitrogen from elsewhere.
To me it seemed impossible. As noted in my Fitness post, you can’t cheat a mass balance. I knew the composition of the food I was giving my yeast, but to figure out how much nitrogen was in there I’d have to look at every chemical compound structure, find the set with nitrogen, for those I’d have to figure out how much of the mass was composed of nitrogen, and calculate the proportion and concentration in my final growth media. It’s doable, but would take a solid 2-3 hours of boring, error prone work. Instead, I can just ask ChatGPT to do it for me, voila. My explanation makes more sense since there Is basically no nitrogen in the media other than my chemical. I’ve done this same sort of calculation and ratio conversion for a bunch of other applications.
For number 2, I am a hopeless coder. There is something about it that is anathema to me. One too many layers of detachment from the material world, an abstraction of an abstraction. I draw the line at an abstraction. But this limits me, because coding is useful. I recently generated a DNA library composed of a bunch of known genes in unknown proportions. To sequence this, I used a technology called nanopore, which takes my mixed pool of DNA and sequences millions of pieces of DNA. I only have 5,000 different genes, so I can get an idea of how often each gene appears in the pool.
But what comes out of the sequencer is gigabytes of data which looks like this:
So in order to analyze it I need to take those reads, and find ones which have a segment that match to the list of genes I put into the library and keep a running tally of the relative frequencies with which each gene appears. That’s basically the level of information I plugged into ChatGPT, and because I’m pretty clueless I spent a couple hours troubleshooting my computer setup, largely by copy + pasting errors into ChatGPT. Eventually, voila. What would have taken me several weeks to get done was possible with a few hours. The second time I had such a dataset it was much faster.
For hobbies, I’ve had ChatGPT help guess how long a screen should be burned for screen printing, find a fast way to draw a grid on a picture in photoshop, a better way to automate editing my film pictures, help diagnose some clutch issues on my motorcycle, how to convert a picture to a screen printable one, choose a solvent to free a stuck camera shutter. All things that I could find on youtube or more esoteric 2000’s era chatboards, but god do I hate watching those 5 minute youtube videos with about 10 seconds of information.
It has made me more productive, and faster learning. Right now, I’m using it to build a little evolution simulator which will be helpful for a paper I’m working on. They just added some image generating integrations, and added the ability to speak to it with voice. I could imagine it being used for app mock-ups, story generating, Think of it as a very dedicated employee and give it a go.
