ChatGPT: Its awsome... and it sucks. But still impressive

I first used chatGPT to generate something in a bash script. Worked great, and have been experimenting with a lot since. It's great at anythingwith language, terrible at math, geometry and physics, and good at generating small pieces of code.

[JeffGeerling]

Hey Matthias! I got a Beepberry, a little Pi-powered Blackberry clone board (but without Cellular, and with a B/W display). If you wanted me to ship it to you let me know! I know you worked for RIM and it would be fun for you to see it.

[BruceChastain]

I had a situation with a specif cad software, where I was trying to find a very specific setting, chat GPT very confidently lied to me about what the setting and value was. But the answer was so confident it initially fooled me into believing it. So it's a great at BSing

[SeanHodgins]

Cat certainly is out of the bag.
For the longer websites you were talking about, you can usually get away with just saying "continue" and it will continue writing where it left off, since it has a character limited output.

[mfsolutions]

Thanks for diving in on this Matthias, I know here at my college it is creating quite a stir. I asked it to give me a course outline on engineering problem solving particularly root cause analysis including ishikawa diagrams... it did a great job but had one section on root cause analysis and fishbone diagrams and another section on ishikawa diagrams with a similar discussion of the previous section. Other stuff it generated had similar redundancies... I have invited my students to use it as an organizational writing tool (makes reading/grading essay questions much easier) but I suggest they read/edit the content very carefully.

[DustinCorbin]

GPT Plus now has plugins enabled which has a lot of new functionality, one of which being Wolfram, which makes it way better at math.

[washoecreative]

I am using ChatGPT to learn Python programming. It's like having a personal tutor who never gets tired of your stupid questions. But as you noted, it does often make mistakes. In fact, the very first time I used it I asked it a simple question, it gave me six answers, and only two of the six were correct. Don't use it for your homework.

I have a friend who routinely uses it to produce drafts of legal documents he can then give to a paralegal to clean up. And people who know or suspect they don't write well can use it to quickly produce letters, e-mails, press releases, etc.

[TKC_]

I’ve had it generate technical procedures from instant message logs discussing how to modify equipment. It got like 90% right. I just had to proof it and tweak a little of the technical jargon. I actually did that because my helper said they couldn’t follow the conversation and I was annoyed.

It’s also good at telling you what sections of building codes to read. Those tend to be big dry documents that I know but don’t remember the sections off hand.

[engineer.woodwork]

Schwäbisch is so hard to understand! I was once in Stuttgart in a store to buy flowers, I didn't understand a word of what the owner was telling me. After three attempts, another customer translated what he was trying to tell me. He wanted to give me a discount because I waited for him for 20 minutes after his lunch break. I hoped you would say some words... never heard you talk German or Schwäbisch. And on the subject of AI, I agree that it's exciting and we need to see what it brings in the future. For now, I also use ChatGPT regularly and often have to correct the information.

[tracybowling]

The things you find interesting, I find hard to understand. You're so damn smart that you're fun to watch.

[ajtxander]

I use it as a tool to help write functions for scripts. I use it as a jumping off point, especially if I have "writer's block." It rarely gives me something that works exactly for what I am trying to do, but it is generally a good starting point.

[paulomelo]

Hello Matthias, thanks for the video. Yes Social media has its problems (doom mongers, flat earthers, etc) but without I could not have found you and see these type of videos. These Language models will have their place, but I believe that we must know its problems and understand how to use them.

[altronixvideo]

I've used it to write a few emails and such, it's 90% there, but I found it sometimes uses words or phrases that a human just wouldn't use. Apart from that, as a dyslexic it's opened up my abilities to get written work produced for my various needs.

[lematindesmagiciens]

I have been using Chat-GPT for a month now, and I view it as a very quick and often incompetent assistant. One of its strengths is the ability to cross-reference information. But it can also spew out completely wrong, made up answers. So, I can understand that teachers are getting a little nervous about it. Also, if you further probe Chat-GPT for references on a previous answer, sometimes you will get only broken links as references. In fact, I ended up getting tons of very interesting but unverifiable answers.

[RickRolling-tcvb]

Thanks Matthias, it's a fascinating subject. When I was studying at university and marking papers, I wished for something that could verify that the information cited was actually in the cited paper, because I know a lot of students lied about that and only a subject matter expert would pick it up - and they're time poor and not marking undergrad papers! I also wanted to be able to explore some subjects more fully through their cited works, but that was often laborious and often as duplicitous as the undergrads! I think that is the sort of thing a language model would be better at than humans, if it was trained properly and didn't lie.
The problem academia has to solve now though is a new way of measuring academic progress that is valid and economical, because the old ways only half worked before (see above) and they can't be trusted to work reliably now.
In my own experiments, I asked it to navigate me across Australia by road, listing every town on my route, and it was quite good at that, with my way points, but it did get the mileages wrong. On top of that, it wasn't very helpful at combining fuel prices and mileages for that trip so I could drive the furthest on the cheapest fuel (because it's double price at the popular but remote places), but it had a red hot go and I am encouraged by the possibilities. I find it handy for projects like your bash script too, because I can ask it variations on a bit of a script that I should know but am having a mental blank about today, and I can do that all day without a shred of embarrassment.

[Tugmun]

This is so far over my head I'm lost, but watched every second and enjoyed every most every bit... Thanks for sharing.

[jemand]

I am german and didn't know what you typed but figured out a little bit more when you read it.

[TheMaxwellee]

I asked it to write a heavy metal poem. Pure gold. Cheers for your thoughts and findings.

[brianb]

You are delightful to watch when you are excited by your topic.

[marcuskaufman]

I use ChatGBT to assist in research for technical articles, mainly in the field of electrical engineering. My experience is that the quality of the response is enhanced by well formulated, content-rich questions. The briefer the question, the thinner the response. By way of example, I asked “discuss bonding in aircraft, ” and got a weak “shot in the dark” type response. Then I asked, “discuss surface faying preparation for the purpose of electrical potential bonding between aircraft electrical equipment or components, ” and got a fabulous response rich with detail. So my experience with the tool mirrors sitting down with a subject mater expert. The quality of the conversation depends on input from both parties.

[unremind]

I used the bing version for writing some auto hotkey scripts. Pretty neat how you can describe what need and it figures out the rest.