I take a job I know well, and try to see how far I can automate it with AI.
It’s a game changer. I just rewrote my syllabus for an upcoming course. What I appreciated most: Perspective. My limits are my experience--ChatGPT provided ideas and direction I might have struggled to consider. The syllabus is now better as a result. My take: tools like this will become a “constant companion” much like spell and grammar checks, or the technology that enables my phone to automatically improve my inherent poor photography skills. Thanks for the inspiration!
I think the most interesting takeaways from this is that those who already know their stuff will find it easier to use this in their workflow and get even more productive.
I wasn't amazed _here_, but only because I had already been amazed experimenting myself in the last few days.
I'm able to use it to find info, like that for machine learning PlaidML is the solution I seek that will run on my graphics card, that older versions of GPT-3 were not able to get and which I have been having trouble even finding looking online myself! Often it will come up with incorrect solutions as well, but if even one solution can lead me in the right direction, that is amazing!
Talked with someone else who gave prompts, and it was able to summarise how to replace part of a motorcycle I knew nothing about, and apparently did so well!
Furthermore, I was even able to give it a very badly-formatted infodump of all the things I want to get done and things I might need to do in a given day, and it was actually able to give me a decent schedule. I might base my day on this now! And in general, I can give it problems in my life and it can give decent solutions while being very patient.
I've also used it to write and edit messages for me, no problem. The results when doing this are remarkably good. To be fair I often edit these so much to match my style that it actually still takes a really long time, often longer than it would normally take me to write a message (and I often like to do it only by telling the AI what to do so that the final message is written by AI), but, still, very amazing! (On another note, I also tried editing this comment with AI too but it has not been working well so I did not use it in the end. I think it may be having issues with the length of it (both the comment itself and when I try to give the article for context), often cutting the answer off when giving long answers. The AI may also just be a bit overloaded right now; even simple questions have a huge delay when trying to get it to answer and in the past I've certainly been able to much get longer answers without them being cut off than I currently am!).
The one are where I find the AI does badly is when trying to use it to write book outlines given a certain plot and conditions. When I did that it kept falling into annoying cliches, and when I told it NOT to use the cliches would make inconsistent writing where it said the cliche then said the thing against it – for example, after writing "Without using or being augmented by AI, peopls are often not competitive, but these augmentations ARE very cheap and widely available, with pretty much anyone who wants access to them being able to access them", it would write "However, not everyone has access to AI augmentations, which are necessary to remain competitive in the job market. These augmentations, based on artificial neural networks, are widely available and affordable, but not everyone is willing to undergo the modifications" and that they were illegal for some reason. I had to do lots and lots of edits to turn it into a consistent story where anything made sense at all. Actually a pretty interesting plot idea after that though.
It, weirdly, seems to understand the world much, much less well when writing stories. On the other hand, one time when trying to get it to summarise my prompt for writing an outline and pointed out an inconsistency, it was even able to tell me exactly why it did not make sense! And then re-write it so it did make sense after I asked it to do so.
I really wish OpenAI would save transcripts in-between browser restarts, because although I copy-pasted a lot and screenshotted even more, I still did not get everything I would have liked.
Count me as impressed. Don’t love all the passive and rather bland voice, but as a starting point for a project, I can see it helping the author get beyond the blank page. I’ve written a lot in my professional career and always prefer the editing phase. Is there a citation protocol when you’ve used an AI tool in your work? I wonder also how readers/students are influenced by knowing that an author has used an AI tool in their work.
However there is a catch. I noticed that if you ask AI to add a list of references on a subject it often provides inaccurate feedback: paper titles and authors do not exist, and even the papers are not on the journals that the systems reports them to be. On one hand this is excellent as when my students try to used I can identify who did the work and who used it to create an automatic response. On the other hand it is frustrating as it is not accurate and often it produces results just to get things printed out.
A few days ago, I searched in ChatGPD, “what is entrepreneurship?” The answer was straight out of a 1980s textbook. That’s scary for anyone assuming the advice is accurate.
Fascinating read, Ethan. I am curious: I replicated your queries in chat.openai.com, and got slightly different answers. I also tried a few additional queries, where the Chat Bot's answers were "less creative" than I would have expected. To that end, I was curious whether you used the Chat bot interface on the web, or whether you programmed queries in python to get more robust answers. Specifically, do you know what the default "temperature" setting is for the chat.openai.com interface? I suspect the default may be "zero", hence more limited answers in the open chat interface. (From your old classmate, Todd...)
Does anyone know the name of the artist who made the image at the bottom?
The world's problems need solutions.
Apparently the solutions will not be prominent in the training data.
Garbage in garbage out.
I believe the moment I stop trying to improve my presentation of subject matter I am cheating the recipients.
I would totally want to know if my children were being educated by a regurgitating AI.
Ask it what’s heavier - a pound of lead or a pound of butter.