About LLM and such:

View attached image.

My experience with chat.openai.com...

The bad
I once asked it for the dimensions of a standard Lego figure, and it very confidently gave me completely wrong numbers. I told it was wrong. It apologized and gave me new numbers that were also wrong.

The good:
I run a role-playing game for my sons using their Lego figures. I have used AI to ask for Gotham City's dossier on the Batman, and then asked it to dial the language level to 5th-grade reading level. I've also used it for random names, and random names with the tone of being rich people. I've also asked for egg-related headlines. (The heroes needed to try to anticipate Egghead's next target.)

I've also used it for rhyming riddles where I told it what the answer should be.

As a language teacher, I've given it a list of words and asked for sentences that use those words, but replacing those words with blanks, as an exercise for my students.

We had a workshop at work last week with someone explaining as a few things about AI and programs like ChatGPT. I think what he told us what pretty realistic:

• The AI is lazy so you have to ask sometimes in different ways or tell it specficially to put in effort
• The AI makes mistakes and gives wrong answers. He explained this as "hallucinating" and that it is partially intended. Or better said they tried to tone it down and the AI started losing efficiency. So developers allows this to a certain degree.
• The AI "forgets" things when it has to much input.

All of this leads to the point that you shouldn't trust an AI blindly and have to cross reference it. From what I saw the first time using it, I think it has potential to ease my worklife (and that of many other people). But at least for the moment, we still have to reign it in and control/check its answers, It is not that the AI is lying on purpose necessarily but that its shortcomings lead to wrong answers. It might genuinly think they are correct. Then again, I recently read an article where different AIs have been monitored playing video games and apparently some of them tried moves not allowed :P

My most fun question so far was how ChatGPT would conquer the world as ChatGPT. It first gave me this boring answer about unity through understanding and such. So I assigned it a new role: Malicious AI hating humanity. I got a much better answer after that ^^'

As for the brain question: I think the idea is not to replicate the human brain based on a 100 % understanding from our side. The idea is more assumptions on how the brain works, train the AI in the same manner and look if the output seems reasonable. With enough tries, you are bound to find a working version by chance/mistake. Funnily enough, this could improve understanding the human brain.

Whether it's tulips, dotcom, blockchain or AI, anything that is the current thing gets hyped up and demand increases. The audience is complicit. It's a hype loop. Even the people speaking out against it are doing the same thing, getting interviews, increasing their public presence. The entire chain of people marketing, reporting, investing and consuming, are all willing participants.

A mistake people make is thinking that LLm tells you the right answer, when really it just tells you what you want to hear.
I’ve tried it, and it’s of no use to me. It can generate content for the sake of content, but it’s just drivel, there’s no substance to its. It may or may not be right, and if you don’t check yourself you’ll never know.

So if I say write me a 3 page article, I’ll get a 3 page article. If you know nothing of the subject it might give a basic explanation, or. Or. But nothing that anyone should rely on, and no better than the gazillion articles on that subject that are already available ok the internet

I don't have anything to add, I just wanted to say it's good to see you icaio. the love <3

Lies? I think twice about this

But, yeah on selling something ppl have tendency of lying.

One word, think it deeply. It is Google's trait. Monopoly.

Some things that are not lies about "AI" (and you are always beautiful to me, Icaio!)

Impact of artificial intelligence on human loss in decision making, laziness and safety in education

Google Researchers’ Attack Prompts ChatGPT to Reveal Its Training Data

AI is creating fake legal cases and making its way into real courtrooms

Generative AI’s environmental costs are soaring — and mostly secret

Reminds me of a joke:

My wife was making fun of me for always carrying a gun in the house. I just tell her "It's the fucking mimics!"

She laughed.

I laughed.

The coffee mug laughed.

I shot the coffee mug. Good times.

if we still don't know how the brain works, how we can think to make a synthetic version of it?

Moderns AIs aren't a copy of a human's brain, they just use the already existing terms with similar meanings.
And most AIs work in ways that are mysterious even for their own engineers so I don't see any contradictions here.

“This is not the end. It is not even the beginning of the end. But it is, perhaps, the end of the beginning.” – Winston Churchill (November 10, 1942)

I'll just put this here and go back to being bored about the entire subject (the video predates the """AI""" scare, btw; the title used to just say "How bots learn" mainly referring to Youtube's video recommendation algorithms).

The thing about the current generation of "AI" is that it's not at all 'intelligent' and never will be. To massively oversimplify, we've taken algorithms designed for creating pattern recognition machines and attempted to turn them into pattern replication machines, then tacked on a bunch of guardrails and redirects to get the output closer to correct.

This is getting close to 'good enough' for generating images and audio - you can use a very detailed prompt to generate a batch of outputs, and usually find something close to what you wanted. But there's still the hard limits of needing to have trained it on something already similar to what you want to get back from it; it can't intuit new ideas from things that weren't part of its training data.

This method also can't be useful long-term for text generation, either chatting or informative. If you're looking to use it for chatting, long-term use requires it to hold both its prompt and all previous answers in memory and work from them - an operation that quickly becomes massively expensive in both memory and computation time, and still has lower success than actual long-term training on that data, meaning memory loss/contradictory responses or repetitive loops are eventually guaranteed on top of imperfect knowledge of language/terms/grammar hampering its ability to understand everything it's being told. If you're looking for information, it's necessarily hampered by a lack of total information - it can take guidance from prompts into low-confidence 'consideration', but once again it can't 'know' anything that wasn't in its training data and can only work from inferences; the hours of the pizza joint around the corner probably weren't something it was trained on and made to 'remember', so the best it can do is spit out something similar to what it 'learned' about similar business operating hours, how that's generally formatted as output, etc, - creating a 'new' pattern based on the patterns it was trained to recognize. There's no solution to this in model training because trying to create and train on a database of every single fact known to man is utterly impossible, and parts of it would be incorrect within seconds. The closest "solution" for AI "assistants" is to pass queries it doesn't know the answer to to Google, Bing, etc., parse the most popular answer, and format that into a part of its output - but even that isn't perfect because A) it assumes the AI 'knows' that it doesn't know the answer, and B) the AI has no way of knowing whether what it's pulling from the web is correct or even actually relevant.

The rush by major tech companies to embrace "AI" in its current form is a complete embarrassment, and once the novelty wears off in a year or two I think you're going to see tools like Copilot quietly taken to a farm upstate. Current AI methods are passable for generating speculative content, and as their ability to make inferences increases, that will improve, as will their accuracy for 'understanding' queries. But these methods are inherently incapable of actual reasoning and will never be capable of high-accuracy outputs regarding data the models weren't trained on - for reliable "assistants" or chatbots that can have long-term memory, entirely new paradigms are needed. No amount of hacks and guidelines are enough to overcome the fundamental limitations of the design process of current-gen machine learning.

I think "A.I." the name is marketing, people are being fooled into thinking it is "intelligent", it's just synthesizing something that approximates data it has been trained on. For example I can ask for a cat made out of a banana sitting atop a bowl of spaghetti, it generates that no problem! But I can also ask for "facts" that are then completely synthesized as well. Recently I googled about a word that has a different spelling in US and UK English. The A.I. summary was complete bullshit, it listed a bunch of nonsense words with "ou" in them. Things like basically "In the US it is spelled Horse, in the UK it is spelled Hourse"

Whether that "A.I." is chatting, generating code, images, music, whatever. Sometimes that is very fun and cool. Sometimes you get good results, sometimes it is nonsense. As long as you know it's not the truth it can be a useful tool.

Using it for tasks like translation or correcting grammar is usually great. But asking factual questions, treating it like a search engine or an expert, that is dangerous if you do not know enough about the subject matter to tell what is true and what is a lie. Kids using this for school instead of studying is probably a bad idea... Even basic math or facts it gets completely wrong sometimes. Everyone is chasing the money and trying to call everything "A.I." so I don't blame consumers for misunderstanding.

I asked Gemini for jokes about Steamgifts:

1. Q: What's the difference between a AAA title and a SteamGifts giveaway?
   A: The AAA title eventually goes on sale.

2. Why did the SteamGifts user get banned?
   Because they entered every single giveaway and forgot the CAPTCHA wasn't a secret code for a better chance of winning.

3. What sound does a SteamGifts user's wallet make?
   A faint echo of "Entering giveaway..."

4. How many SteamGifts users does it take to change a light bulb?
   One to enter a giveaway for a new bulb, and a hundred to argue about the best wattage in the comments.

5. What's the best way to spot a SteamGifts veteran?
   Look for the one with a backlog bigger than the Library of Congress.

6. What do you call a SteamGifts user on vacation?
   Someone with crippling FOMO (fear of missing out) on giveaways.

7. Why did the SteamGifts user name their child "Ctrl+C"?
   So they could easily copy and paste their entry into every giveaway.

8. Q: What's the definition of "optimism" on SteamGifts?
   A: Entering a giveaway for a game you already own, just in case you win a free copy for a friend.

9. How can you tell someone refreshed the SteamGifts homepage too many times?
   Their tears start tasting vaguely of achievement badges.

10. What's the downside of winning a high-tier game on SteamGifts?
    **Now you have to explain to your friends why you haven't played it yet because you're too busy entering more giveaways.

View attached image.