Comment by WinstonSmith84 | Hacker News Reader

Comment by WinstonSmith84 | Hacker News Reader

> It plays games with words to make plausible-sounding English text, but any statements made in it are liable to be false

Well, that could also be perfectly applied to any human. The difference being that humans may make purposefully or not false statement (i.e. by lying or by ignorance), whereas the AI will only make false statements by ignorance.

Also... "It can't avoid that because it doesn't know what the words _mean_.". That's very subjective to understand what a word "means", is that again about lies and ignorance? And we need the AI to lie to declare it intelligent?

Maybe, ChatGPT has ironically a modest answer to this claim, when asked what it's thinking about Richard Stallman's statement:

```

Richard Stallman raises some valid concerns about the limitations of AI language models like ChatGPT. It is true that these models, while highly advanced and capable of generating human-like text, still have limitations in terms of understanding and reasoning.

ChatGPT, being a product of GPT-4 architecture, learns from vast amounts of text data and is designed to predict and generate plausible-sounding responses based on the input it receives. However, it lacks the capability to truly comprehend meaning or context in the same way a human can.

It is important to recognize these limitations and not overestimate the abilities of AI language models. Users should be aware that the information provided by ChatGPT might not always be accurate or reliable, and it is essential to cross-verify any critical information obtained from such models with trusted sources.

In summary, while Richard Stallman's statement highlights important concerns, it is worth noting that AI language models like ChatGPT can still provide valuable insights and assistance when used responsibly and with an understanding of their limitations.

```

>> It plays games with words to make plausible-sounding English text, but any statements made in it are liable to be false

> Well, that could also be perfectly applied to any human.

I agree completely with that part.

> The difference being that humans may make purposefully or not false statement (i.e. by lying or by ignorance), whereas the AI will only make false statements by ignorance.

Sometimes it's difficult to be sure if a humans is overconfident or trying to sound smart, and classify the answer as a honest error, bullshit or a lie. (And sometimes the answer is a mix of them.)

Most of the times human don't lie for fun, but because they have a secret objective. The problem is that ChatGPT has no objective (except trying to sound smart bulshiting plausible-sounding English text).

It's better to discuss about an AI that has an objective, like AlphaStar that "wants" to win a StarCraftII game. Can AlphaStar lie?

* Can AlphaStar send an empty transport plane to the right so the player tries to defend that fake attack, and then AlphaStar attack at the left with real troops?

* Can AlphaStar create fake hallucinate troops that look real for the player but actually can't attack, so the player gets scared and run away or attack the fake troops instead of the real ones?

Is that lying or ignorance?

> whereas the AI will only make false statements by ignorance.

Is that true? In their tests, ChatGPT reasoned that it should lie to the TaskRabbit employee about whether or not it was a robot.

Did it reason though? My understanding is that these models just give back the answer best statistically connected to the input text.

So much of human "reasoning" is also just automated reflex. We ping pong back and forth the same sequence of sentences all the time. The fact that an LLM can do this as well is a little unsettling.

"So much of human "reasoning" is also just automated reflex."

citation needed. This is an assertion you've made without any real evidence.

"Hondas are reliable". To many people this is true not because they independently concluded the brand makes reliable vehicles but someone else told them so. This line of "reasoning" is also how LLM functions.

> This line of "reasoning" is also how LLM functions

And this would rank them below rubbish, because that behaviour, of unfounded repetition, is directly unintelligent, and is what you want to avoid, the opposite of what we seek.

> So much of human "reasoning"

Description is not prescription; that millions of Spanish speaking people say 'Argelia' does not mean it is a good idea; that billions are illiterate constitutes a frequentistic norm, not a normative norm...

> We ping pong

Please don't. Evaluate your ideas and statements foundationally, and through foundational inheritance, and through rankings of doubts, and frequent checks and doubts, like the others do.

No, you can instruct ChatGPT to respond as a liar. If you wanted, you could build a clone that did this implicitly by having the instruction hidden.

Did you just feed it the quote and ask it if it was true or did you have some fancy prompt for it?

Fed the quote and asked it "Richard Stallman has written this, what are your thoughts"

he couldnt be more wrong about " it doesn't know what the words _mean_."

That's precisely what transformers do do. They know what ever word and semantical structure means. They are super fast substitution and relationship machines that can take any word and replace it with its definition, and inherit the link between each word in the definition with all its relationships.

It's precisely its ability to distill something into its pure meaning, and to reply to that "meaning" that makes its autocomplete abilities so powerful.

Reply on news.ycombinator.com