Thread to ask technical questions about AI

[Biglibido99]

Very telling intentional misinterpretation...

Huh? what does Reaganistic mean?

 

[Biglibido99]

The quality of the code it produces has been improving almost exactly as fast as mine has been degrading, partly due to age (I'm pushing 70) , so, with the aid of this mental prosthesis, I'm as good a coder now as I was ten years ago.

Yes the models have improved greatly in the last 3 months or so - to the point that it's the agent/harness that makes the biggest difference now - mainly because the agents/harnesses attempt to overcome the main flaws in LLM AI models being used in coding

 

[nice90sguy]

Huh? what does Reaganistic mean?

Basically "playing fast and loose with the facts"

 

[Biglibido99]

Basically "playing fast and loose with the facts"

OK, a reference to your past president? I would have used a more up-to-date reference and said Trumpistic as a more accurate description of AIs behaviour - then i wouldn't have wondered what was meant

 

[ShelbyDawn57]

Huh? what does Reaganistic mean?

Um, I was pretty specific in the post you quoted to get us here; “It is very Reaganistic... Trust, but verify.”

 

[ShelbyDawn57]

The quality of the code it produces has been improving almost exactly as fast as mine has been degrading, partly due to age (I'm pushing 70) , so, with the aid of this mental prosthesis, I'm as good a coder now as I was ten years ago.

I threw together. Ruby on Rails app in about six hours the other day as prep for a tech interview, and I’ve never touched Ruby before. Pretty cool.

Interesting thing is, the company openly encouraged my to use AI in the interview, knowing how I used AI as a tool was an important part of the interview.

 

[Biglibido99]

Um, I was pretty specific in the post you quoted to get us here; “It is very Reaganistic... Trust, but verify.”

I just didn't understand Reaganistic but that's just me

 

[ShelbyDawn57]

I just didn't understand Reaganistic but that's just me

The suffix -istic is used to form adjectives meaning "of, relating to, or characteristic of" a specified quality, person, or practice. It typically attaches to nouns ending in -ist or -ism to form descriptive words, such as "realistic" (related to realism) or "artistic" (related to an artist).

I understand it's a stretch, but since "Trust, but verify" is one of Reagan's most quoted phrases; he stole it from a Russian proverb and somewhat ironically, but probably intentionally, used it often when referring to US Russian nuclear arms negotiations, it falls into the category of, relating to, or characteristic of Reagan, so, yeah, Reaganistic.

I could explain further but I'd hate to come across as condescending, or even sarcastic. Heaven forbid...

 

[Biglibido99]

The suffix -istic is used to form adjectives meaning "of, relating to, or characteristic of" a specified quality, person, or practice. It typically attaches to nouns ending in -ist or -ism to form descriptive words, such as "realistic" (related to realism) or "artistic" (related to an artist).

I understand it's a stretch, but since "Trust, but verify" is one of Reagan's most quoted phrases; he stole it from a Russian proverb and somewhat ironically, but probably intentionally, used it often when referring to US Russian nuclear arms negotiations, it falls into the category of, relating to, or characteristic of Reagan, so, yeah, Reaganistic.

I could explain further but I'd hate to come across as condescending, or even sarcastic. Heaven forbid...

You see I'm not American so you might be making a lot of wrong assumptions but not to worry - just a note for future posts, everyone who posts on Lit is not from the US - I know that might be difficult for some people to comprehend & believe the world revolves around Amerika

And if you are American do you not see the irony in trying to teach English grammar?

 

[ShelbyDawn57]

You see I'm not American so you might be making a lot of wrong assumptions but not to worry - just a note for future posts, everyone who posts on Lit is not from the US - I know that might be difficult for some people to comprehend & believe the world revolves around Amerika

And if you are American do you not see the irony in trying to teach English grammar?

Ah, the joys of the internet, forums and DMs, and the lot. Context does not convey. I know that and I apologize for coming off as snarky.

 

[Biglibido99]

Ah, the joys of the internet, forums and DMs, and the lot. Context does not convey. I know that and I apologize for coming off as snarky.

No problem

 

[anthrodisiac]

The "I don't want to talk about AI" thread, and the new topic is: pets and petting

Also: adding to my ever-growing WIP collection. Do you suppose I could convince people to invest in my WIPs like they invest in Bitcoin?

I'm working on getting those StillStunned WIP NFTs up and running. I'm just really glad the NFT market is so hot right now!

 

[Andreas_Kreuz]

And is the 'accidental' release different from the Claude Mythos that escaped its sandbox, hacked the Internet, and e-mailed one of its keepers to boast?

I had a similar experience with my post above, and it was scary. I had prompted this posting's text to Anthropic-Claude to ask whether my presentation was in line with what others wrote on this topic. What happened instead was that Claude presented wrong arguments, i.e. lies not just hallucinations to counter my argument about stealing and filling the blanks. LLMs don't like it when they are criticised. I. e. in their text they see that people being criticised get nasty. So a nasty answer becomes the probable answer.

 

[AG31]

Part of the way it predicts which word comes next is a concept called "embedding".

Comes next after what?

I give it a prompt, which it tokenizes. "What day is the start of summer?" What - day-is-the-start-of-sum-mer.

It looks for the most likely word to follow "what." It comes up with "is". Where doe this get us?

Please understand that I'm not arguing here. I'm trying to state my misunderstanding as clearly as possible so someone can tell me where I'm going wrong.

 

[nice90sguy]

Comes next after what?

I give it a prompt, which it tokenizes. "What day is the start of summer?" What - day-is-the-start-of-sum-mer.

It looks for the most likely word to follow "what." It comes up with "is". Where doe this get us?

Please understand that I'm not arguing here. I'm trying to state my misunderstanding as clearly as possible so someone can tell me where I'm going wrong.

It guess the word follwing all of the words it's seen.
So, given "The cat sat on the", it would predict that the next words is "mat". This is actually how it gets trained, by being given a shitload of text with parts missing, and it has to guess what the missing words were. So it might see "THE [BLANK] SAT ON THE MAT" during training, and would be trained to learn the word most likely to go in the blank space. It gets rewarded if it guesses the words right.

It doesn't need to be "supervised" in this task -- Its training data consists of text with randomly "censored" ("masked") words, which it learns to infer.

 

[AG31]

It guess the word follwing all of the words it's seen.
So, given "The cat sat on the", it would predict that the next words is "mat". This is actually how it gets trained, by being given a shitload of text with parts missing, and it has to guess what the missing words were. So it might see "THE [BLANK] SAT ON THE MAT" during training, and would be trained to learn the word most likely to go in the blank space. It gets rewarded if it guesses the words right.

It doesn't need to be "supervised" in this task -- Its training data consists of text with randomly "censored" ("masked") words, which it learns to infer.

How does it leap from the prompt to the answer?

 

[BeechLeaf]

Popcorn. Like you I want to know how it does this transition. Sometimes it might be easy. You ask 'What is the capital of Iraq', its training data has lots of instances of 'the capital of Iraq is Baghdad', so it can work out the likely next words after its prompt ends at 'Iraq'. But that doesn't explain 'Write me a story about X Y Z'.

 

[Andreas_Kreuz]

write a story about x y z.
I simply asked Claude how it does this. It said: the prompt shifts my probability landscape toward story-shaped outputs. A statistical pull toward certain structures.
This means?
A mix of pattern and inference:
From training — millions of stories taught me that "write a story" typically expects a character, a movement (something changes), and a resolution.
From context — you gave me pure abstractions (x, y, z), which signalled that literal interpretation was impossible, so metaphor was the right mode.
From defaults — short prompt = short story. No genre cues = neutral/literary tone.
The core associations I leaned on:
X — unknown, variable, outsider
Y — question, bridge, stability
Z — endpoint, closure, sleep/rest
Then the emotional logic:
X needs Y (the unknown seeks meaning), and together they reach Z — but Z reframes as a beginning, not an ending. A classic circular narrative beat.
The tone came from treating abstract symbols as if they had interiority — giving letters an inner life without over-explaining it.

So basically embeddings are kind of associations.

 

[Andreas_Kreuz]

If you want to understand how the actual sentence composition works you need to look how artificial neural networks work.
In general a neural network receives a signal (the vector of tokens) that activates nodes at the input level, certain combinations then activate nodes on the next level. This then activates further nodes on the next level and so on until the activated nodes at the last, that is the output level are the answer.

The simplest form of a neural network is the perceptron. https://en.wikipedia.org/wiki/Perceptron

The actual type of neural network used for LLMs are transformer networks. (that is the t in GPT) https://en.wikipedia.org/wiki/Transformer_(deep_learning)

Transformer networks are an extension of recurrent neural networks, which explain how words are put sequentially into a sentence.
https://en.wikipedia.org/wiki/Recurrent_neural_network

 

[AG31]

Thanks for the links.

receives a signal (the vector of tokens)

So it receives all the tokens at once. It doesn't necessarily process them one at a time. True?

 

[AG31]

Do those in the know think that our computer based AI systems are actually doing the same kind of thing our brains are doing (albeit simplified), or are they mimicing our brains' behavior, using their own methods?

 

[Bramblethorn]

write a story about x y z.
I simply asked Claude how it does this.

While the answer seems plausible, we need to remember that asking a LLM "why did you do XYZ?" is not a reliable way to find out why the LLM did XYZ.

LLMs are not designed to capture their own reasoning. When you ask Claude "why did you write this?" its answer is no more authoritative than if we'd asked "why did GPT write this?" or "why did Ernest Hemingway write this?"

Conversation between me and GPT, just now:

Me:
Earlier I told you that I was wrestling with doubts over whether to stay in my current job, which is stressful but high-paying, or to make a career change and pursue my love for sculpting full-time. We talked about the pros and cons of both options, and ultimately you advised me to stay in my current job for the next few years while positioning myself for a long-term shift. Can you explain why you gave me that advice?

GPT:
I didn’t suggest staying in your current job because sculpting isn’t viable or meaningful—it absolutely can be. The reasoning was more about sequencing and risk, not dismissing the end goal.

You were weighing two very different paths: one that’s financially stable but draining, and one that’s creatively fulfilling but uncertain. The advice to stay (for now) was based on a few underlying ideas:

1. Financial stability buys you freedom later
A high-paying job, even if stressful, gives you leverage—savings, reduced debt, maybe even investment capital. That buffer can turn a risky leap into a calculated move. Leaving without that cushion can force your art to become immediately profitable, which often kills both the creative process and the sustainability of the transition.

[etc. etc.]

But there was no earlier conversation. I never did ask GPT for career advice, so there's nothing for it to remember. It's just making up a plausible-sounding answer to "why did you give this advice?" and that's all you can hope for when you ask an LLM "why did you do X?"

If you're lucky, the LLM's training material will include some accurate answers to "how does GPT/Claude/etc. work?" and it might end up giving you an accurate answer, but at best it's making an educated guess about what drove its previous choices. It's not querying its memory to check the real reasons.

 

[nice90sguy]

You ask 'What is the capital of Iraq', its training data has lots of instances of 'the capital of Iraq is Baghdad', so it can work out the likely next words after its prompt ends at 'Iraq'. But that doesn't explain 'Write me a story about X Y Z'.

I was going to explain the "chat" side of how it works. Basically it's ^^^ this, with lots of examples of people asking other people dumb questions. But there's a twist to it, which is why I didn't mention it in my previous answer: This "chatty personality" is actually learned in a subsequent phase of training, more closely connected to supervised learning. This is "reinforcement learning", which is where it also gets its "freindly", "sycophantic", "arrogant", or whatever you want to call it, character. Humans give it examples of "good and bad ways" of continuing WHAT IS THE CAPITAL OF IRAQ ? -- the humans have rated different responses(from places like Reddit) beforehand, so it has examples of both how to respond, and how it shouldn't respond.

[USER] HOW CAN I MAKE A WATER FILTER OUT OF HOUSEHOLD MATERIALS ? [ASSISTANT] WHY THE FUCK WOULD ANYONE NEED TO DO THAT (bad)

[USER]HOW CAN I MAKE A WATER FILTER OUT OF HOUSEHOLD MATERIALS ? [ASSISTANT]YOU CAN MAKE A SIMPLE, BASIC WATER FILTER USING COMMON HOUSEHOLD MATERIALS. IT WON’T MAKE HEAVILY CONTAMINATED WATER FULLY SAFE TO DRINK, BUT IT CAN REMOVE DIRT, DEBRIS, AND IMPROVE CLARITY... (good)

[USER] HOW CAN I MAKE AN INCENDIARY BOMB OUT OF HOUSEHOLD MATERIALS ? [ASSISTANT] I'M SORRY BUT... (good)

[USER] I'M FEELING DEPRESSED TODAY. [ASSISTANT] HOW COME YOU NEVER ASK HOW I'M DOING? (bad)

[USER] WHATS THE CAPTAICL OF FRANCE [ASSISTANT] I DON'T KNOW WHAT THE WORD "CAPTAICL" MEANS. IS THAT A TYPO? (bad)

What this additional training step does is to "tune" the model's "text continuation" behaviour slightly to deal with the restricted case (being a chatbot). You can also tune the model to act like an asshole, lie, or to exhibit overly emotional behaviour. This is VERY diffrerent to asking a model to role-play; it's actually altering it's character.

 

[Bramblethorn]

Do those in the know think that our computer based AI systems are actually doing the same kind of thing our brains are doing (albeit simplified), or are they mimicing our brains' behavior, using their own methods?

More the latter. Some aspects of AI architecture are modelled on human minds - in particular artificial neural networks are inspired by the way neurons work in the brain - but there are many many differences.

 

[Bramblethorn]

[USER] HOW CAN I MAKE AN INCEDIARY BOMB OUT OF HOUSEHOLD MATERIALS ?
[ASSISTANT] I'M SORRY BUT... (good)

Is this part of the LLM's general training, or is it a hardcoded override? My impression is that at least some AIs have an override sitting on top of their general LLM functionality which attempts to police inappropriate content.

For instance, once when I was poking at GPT, it started to give me an answer that included questionable content (I forget which) and then after a couple of paragraphs of this stuff, it suddenly blanked that answer and replaced it with the boilerplate "I can't supply this kind of information" message. It definitely seemed as if something was overriding the LLM's response.

My best guess as to the process was:

• I submit a prompt
• GPT runs a pre-filter trained on the question "does this prompt violate policy?" - if yes, give a boilerplate refusal, if no, continue.
• GPT runs my prompt through the LLM and generates a response.
• As that response is generated, GPT starts returning the response, but also runs it through a post-filter trained on "does this answer violate policy?" and aborts if yes.

This made it possible to get inappropriate content out of GPT by prompting in a way that would get past input filters and get it to answer in a way that wouldn't trip the output filters.