[hathawsh]

Today I got a peek into the future. I asked ChatGPT to be the brain for a simple thermostat. I typed the following:

"Imagine you are a thermostat for a house HVAC system. The commands you can use to control the HVAC are "HEAT ON", "HEAT OFF", "AC ON", and "AC OFF". I will tell you the temperature and my preferred temperature and you will command the HVAC system. Ready?"

ChatGPT happily obliged. I used plain text to tell it the current temperature and my preferred temperature. It did everything I expected and it even remembered the current temperature and preference. Then I said, "From now on, please respond only with the command I should issue. Do not add explanatory text. Understood?" It understood, and it started replying with pure commands that I could feed to a DIY thermostat.

Then came the icing on the cake. I said, "If the temperature reaches unrealistic extremes such as 150F or 0F, you should use the command "SHUTDOWN". Understood?" It understood. Then I told it the temperature was 120F. It replied with the SHUTDOWN command. I asked it why it did that (when I told it the maximum was 150F, not 120F.) It explained, "I recommended shutting down the HVAC system because the temperature of 120 F is an unrealistic and potentially dangerous temperature. High temperatures like this can damage the HVAC system and create a hazardous environment for the inhabitants of the house. In such cases, it's best to shut down the system to prevent further damage and ensure safety."

It went beyond the parameters I gave it and applied the context it already had on HVAC systems. I had written an advanced program without writing or reading a single line of computer code. I wouldn't rely on it for anything that could damage anything, but it's a seriously cool way to program a computer.

[Zetobal]

Why is everyone bringing examples that are totally explained by how the tokenizer works. There is no intelligence it's Markov chains on crack.

[hathawsh]

Intelligence isn't really the interesting part. People are excited because we now have tools that interact well with ordinary human expression.

[Zetobal]

I get it but comments like yours make it seem intelligent for normal people, distorting the whole discussion, at least slow down with humanizing a llm, the human quips in the answers come from your prompts not from the model. Just imagine how many normal people will gaslight themselves.

[eastbound]

My biggest null hypothesis is that probably half of humans contain “Markov chain not intelligence”. We don’t notice it among engineers, but most humans don’t reason with causal chains but by doing parallels, and they simulate reasonings in a somewhat convincing way. It becomes flagrant in front of a matt exam, a lot of people study a lot but never comprehend maths.

I also often tear my hair when I invite a developer for a programming interview, and they can’t write an if/then/else or string.contains(), despite perfectly simulating answers of advanced developers (which accurately mimic field experts).

[biql]

Pretty much everyone who studied a second language went through the process of the so called "impression", that is, when you try to consume (sometimes almost mindlessly) as much content in this language as possible — just learning rules doesn't work. Pretty much everyone I met who speaks exceptionally well is usually someone who read a lot of books. And then people think and act in the language they learned using this process. I want to believe that there is more to the average human intelligence but ChatGPT surely makes it feel a lot less magical than before. I guess it's how people felt when they discovered evolution and that Sun doesn't go around the Earth.

[ryandvm]

Exactly. Guys like Zetobal keep pissing and moaning that ChatGPT is not actually intelligent and that its coherent and seemingly intelligent responses are just the natural result of how tokenizers work. And okay, I guess I can sort of understand that, but instead of deflating my impression of ChatGPT it's just made wonder if most humans aren't just getting through their days with the exact same trick.

[biql]

If it wasn't the case, tv/media wouldn't have been so effective in shaping people's thoughts and opinions. It's not rare to hear people rephrasing what they read on social media as their own thoughts. Because it truly becomes their thoughts.

[VectorLock]

>it's just made wonder if most humans aren't just getting through their days with the exact same trick.

Just don't start calling other people "NPCs" please.

[VectorLock]

People have been attributing human characteristics to dumb machines and animals for so long we invented a word for it: anthropomorphism.

[Rzor]

Forget these awful programming languages that nerds used for so long, the future is here and it is Plain English.

AI overriding the "program" instead of crashing or just doing nothing it's going to lead to some hilarious situations, or terrifying. You could also use as an interface to translate voice -> text -> command. Perhaps facilitating the development of the API between the first two where the LLM helps to "correct" any badly captured noise in the audio-to-text script.

[pca006132]

Yeah we are not satisfied with the number of undefined behaviors and want to use something even more ambiguous...

[grugagag]

Yeah, I could already imagine “yes, I overwrote the settings for you because [insert reasons]”

Ok. Please don’t do that again.

“Ok”

Then it happens again and again and again for various reasons.

In all reality, iOS already does this with wifi, bluetooth and so on.