Growing ChatGPT use on the site.

[Iron E]

Originally Posted by s/v Jedi

This is ChatGPT 4o which is the light version, which is good enough for this

See how important it is to expand the context by asking for more detail, options etc.

Not surprised that I have the #1 option aboard

This is where AI excels when compared to "Google"

With Google, would you get the correct answer, an answer optimized for SEO, an answer with affiliate links, or a completely jacked answer from a "Popular" YouTube channel?

[s/v Jedi]

Originally Posted by Iron E

This is where AI excels when compared to "Google"

With Google, would you get the correct answer, an answer optimized for SEO, an answer with affiliate links, or a completely jacked answer from a "Popular" YouTube channel?

it will do whatever you ask it to, but this is still a censored LLM (Large Language Model). I run non censored LLM’s on my Mac but of course they are slow and older tech, but you get anything you ask even if deemed inappropriate, like how to make a car louder, how to make a weapon, why the people in Kosovo are supported in their struggle for independence but not the peoples in Donbass, Catalonia etc. So there’s political censoring as well as moral.

[anotherT34C]

You’re probably unlikely to run a local llama3.1 405B model on your home/boat hardware just yet, but when you do, you’ll have your very own HAL9000.

[s/v Jedi]

Originally Posted by anotherT34C

You’re probably unlikely to run a local llama3.1 405B model on your home/boat hardware just yet, but when you do, you’ll have your very own HAL9000.

Exactly. I can only use the smallest ones, which are already amazing. I recently bought a MacMini but was close to buying a MacStudio instead to run LLM’s aboard.

[Shanachie]

The thing to remember about AI as it is currently designed is that it aggregates conventional wisdom by reading millions of posts, books, articles, etc., and telling you what most people think.


So it's wonderful if you're asking about how to fix a toilet or the proper way to bleed a diesel.


Once you get into areas where moral, ethical and skill judgments are involved, not so much. Most of the time, you get on the one hand this and on the other hand that. Occasionally, you get alarming stuff.


Most important, AI does not think critically. So you're never going to get that surprising insight that you can get on occasion from talking to other people who have actual experience.


Taking it to sailing:


AI can do a wonderful job of telling you how to sail a boat so that you can cross a lake, bay or maybe even an ocean. But if you want to win an important race, you need that penetrating insight that separates the truly great skippers from the mediocre ones.

[anotherT34C]

Now that I think about it for a few seconds, starlink + open ai API + pynema2 and pyserial should get us most of the way to HAL9000.

[MiddayEnglishmn]

Originally Posted by anotherT34C

Now that I think about it for a few seconds, starlink + open ai API + pynema2 and pyserial should get us most of the way to HAL9000.

If you’re using the API theoretically you wouldn’t need the internet. You’d have to make your own database but for a boat that would contain all maritime knowledge.

It would work for a specific function.

[GordMay]

Please pardon my ignorance:
How does one fact check the answers, IA, like ChatGPT provides, particularly when one has little/no prior knowledge, of the subject?
Can it provide links [citations], to [trusted] external sources?
How sensitive is it, to context, and subtle differences in phrasing,of one’s query?
How up to date/current is the information, it sources?

[anotherT34C]

Originally Posted by MiddayEnglishmn

If you’re using the API theoretically you wouldn’t need the internet. You’d have to make your own database but for a boat that would contain all maritime knowledge.

It would work for a specific function.

If you want to use GPT4 to control your boat while having natural conversation with it, it'll need to be able to make API calls. Meta's 70B model can probably be run locally on a really good bit of home hardware, but the power needs might be a bit much for most of our boats.

Gord, you can ask it to provide links to references, and it will oftentimes do this itself. But... it does BS. Best to treat it kind of like wikipedia... use for things that don't matter much, or as a first pass for knowledge gathering. Of course, for my primary use cases; code and shell script writing, IT task automation and management, it's responses will either work or not (usually do work when prompted correctly).

[GordMay]

Computer science prof explains how AI makes mistakes
David Gerhard, the head of computer science, at the University of Manitoba, says artificial intelligence cannot be trusted, to give truthful answers to questions.
➥ https://www.cbc.ca/player/play/video/9.4416670

Thanks, T34C.

[s/v Jedi]

Originally Posted by GordMay

Please pardon my ignorance:
How does one fact check the answers, IA, like ChatGPT provides, particularly when one has little/no prior knowledge, of the subject?
Can it provide links [citations], to [trusted] external sources?
How sensitive is it, to context, and subtle differences in phrasing,of one’s query?
How up to date/current is the information, it sources?

Yes, of course, but you need to ask. Here it retrieves citations and lists them.

The links go to boatingworld anchor test, Fortress reviews, Boating Magazine anchor test and a specific Boating Magazine Fortress test.

[s/v Jedi]

Originally Posted by anotherT34C

If you want to use GPT4 to control your boat while having natural conversation with it, it'll need to be able to make API calls. Meta's 70B model can probably be run locally on a really good bit of home hardware, but the power needs might be a bit much for most of our boats.

Gord, you can ask it to provide links to references, and it will oftentimes do this itself. But... it does BS. Best to treat it kind of like wikipedia... use for things that don't matter much, or as a first pass for knowledge gathering. Of course, for my primary use cases; code and shell script writing, IT task automation and management, it's responses will either work or not (usually do work when prompted correctly).

I looked into this and a MacStudio would deliver a compact, workable platform: https://www.apple.com/mac-studio/specs/

The knowledge of the LLM comes from it’s training. When you feed it BS then that is what comes out. It works the same with humans: our kids are fed BS in school and now all of society is crumbling down as the result.

On Wikipedia: the English version isn’t bad. I would say it is better than the old bookshelf encyclopedias, which were more biased than Wikipedia. With Wikipedia you have many editors, from all backgrounds.

[GordMay]

Originally Posted by s/v Jedi

Yes, of course, but you need to ask. Here it retrieves citations and lists them.

The links go to boatingworld anchor test, Fortress reviews, Boating Magazine anchor test and a specific Boating Magazine Fortress test.

How would I get to the [2] sources cited, without performing my own independent [Google] search? I see no links, nor URLs.

[anotherT34C]

I believe you can click on where it reads 'searched six cites' and it will show you the links. An independent google search is always recommended when it actually matters.

[anotherT34C]

I think if I only wanted the model to interact with the boat's instruments and controls via natural language, then a local model would do. If you want the full "Hal9000 experience" (minus the homicides) I suspect your best bet will be to setup for API calls to the best openai or anthropic model over the internet.