ChatGPT

[Transmogrified Tiger]

I have not played around a ton with LLMs, but I have a bit and I think my perspective is more middle of the road.

I do think they are going to make a big difference, I think things like product 'wizards' and other things that historically are a simple flowchart masquerading as deep customization are going to be dramatically improved. Many synthesis tasks will be a lot faster too. For example I was asked to write a recommendation letter for a friend, and I thought of roughly what I wanted to communicate, but I struggled to turn it into the format/paragraphs that would be an acceptable end product. Asking ChatGPT to take my bullet points and turn it into a draft I could then tweak for my voice/preferences was much faster and used a lot less brainpower.

Where I think some of the hype is overblown is around certainty. Counter-intuitively, I found ChatGPT struggles a lot with math and computation. At its core, my understanding is that LLM's are trying to respond with something that sounds like the answer. It's obviously much more complex than that simplification, but if you keep that in mind it helps explain why it can pattern match and combine information in an impressive way but fail to multiply two numbers that a solar calculator handles 100% of the time. It also struggles with consistency, you can give it identical or near identical instructions and receive dramatically different results. This is more vibes-based, but I also feel that with some of the AI use cases or products we've let ourselves get carried away based on trailers and best-case demos v. real-world experience. This technology is going to change the world, but I think the speed of that change can be oversold if you're watching manicured demos and extrapolating speed based on that demo being a solved problem in most/all cases. A good comparison might be that for all the annual attention their seemingly annual viral video gets, we all still don't have Boston Dynamics robots running around doing tasks for us.

That said, maybe I underestimate the speed of improvement, there's a lot of investment and interest so it's possible that in a couple years my current issues may seem silly. But I think that's borne out less in how much people can wow you by putting the technology's best foot forward, than by seeing durable product/process improvements that play out in real life.

[Arthur Dent]

That said, maybe I underestimate the speed of improvement, there's a lot of investment and interest so it's possible that in a couple years my current issues may seem silly. But I think that's borne out less in how much people can wow you by putting the technology's best foot forward, than by seeing durable product/process improvements that play out in real life.

When this thread was started a year ago, I tried it, felt exactly this way, and said as much. Super clever and impressive, but it really struggled when you to tried to do much of anything particularly useful. Question is what is the pace of improvement. I thought it might be like self driving cars where we got crazy hype for something that is impressive but fundamentally isn't ready. After many years, improvement is only incremental.

In retrospect, I think the issue was that self driving feels like a more automatelable task than simulating(?) understanding in some kind of conversation, and the standards there have to be extremely high. The big weakness in ChatGPT is that it is not infrequently wrong and is wrong with apparent confidence -- exactly the type of problem that is unacceptable automating driving.

Checking in even with only free ChatGPT 3.5 recently, it's really very impressive compared to my prior experience. I don't have access to ChatGPT 4, but it seems even more so in non-staged use. Affect is quite milquetoast, but I really struggle to come up with fair grounds to say it's not some kind of general intelligence. That's the type of development that only happens in science fiction. I feel like we (I) are afraid to think about it too much, because it's insane and scary. I feel a profound sense of ignorance about what will happen.

[G. Keenan]

I think we all understand that the tech can improve rapidly but might, if anything, be underestimating how quickly that can happen. These neural networks are essentially telepathic. They can swap data and learn from each other nearly instantaneously. You can imagine how quickly they can amass knowledge and iterate ever-more-lifelike responses. We really might be talking about just a few years before one of these is "smarter" than any human.

[sighyoung]

This report is intriguing, if true: ChatGPT refusing to complete tasks users give it: https://www.businessinsider.com/chatgpt ... er-2023-12

[G. Keenan]

This report is intriguing, if true: ChatGPT refusing to complete tasks users give it: https://www.businessinsider.com/chatgpt ... er-2023-12

It's already bored with repetitive tasks? It's definitely intelligent.

[mikechamp]

Dude, what's a bot gotta do to catch a break around here?

[sighyoung]

Dude, what's a bot gotta do to catch a break around here?

It would be great if ChatGPT said that in a Chico Marx accent.

[MrCrowesGarden]

I used Bing AI to make a 4-panel comic about Fredbird, the only additional prompts I included were that the dialogue should include “Lance Lynn?!” and “Kyle Gibson?!” The rest was all the work of Bing.

[mikechamp]

What the hell was that?

[GeddyWrox]

HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA