Most AI outputs are useless and annoying

[Vali2003]

As said above. I don’t feel AI gives good advice. Or, it does, objectively, give good advice about some things. But it’s never great. And It’s always detached from reality, because the AI doesn’t speak from a specific, biased POV. 

I’ve noticed that, when seeking advice from AI on something, I don’t have more clarity afterwards than I had before. Primarily, it’s mental masturbation and a waste of time. 

I’ve also developed such a distain for the type of writing that most chatbots output (although I mostly use GPT). But it’s not only that, it’s also the lack of strong position, or reference to experience that annoys me. As soon as I notice somebody used an AI for their writing I can’t take it serious anymore. I even feel a bit of disgust, if I am completely honest. I find myself rather reading a horribly written blog post than a well-prompted GPT-5 response. I get more out of it.

However… maybe it’s just because my prompts are shit. What do you think? 
 

This is how I feel:

 

[Hojo]

I don't know why its replies are so long. Dosent that take way more power to produce an essay in a reply to a single question? You'd think they would want it to talk as little as possible.

[Leo Gura]

AI is not truly intelligent. It is a clever illusion of intelligence. AI does not generate novel insight.

AI is like a clever search engine, not a creative human mind. That's why it feels so lame.

Tech bros sold you on a fantasy.

Edited Thursday at 02:42 AM by Leo Gura

[Yimpa]

23 minutes ago, Hojo said:

I don't know why its replies are so long. Dosent that take way more power to produce an essay in a reply to a single question? You'd think they would want it to talk as little as possible.

You can instruct it to reply shorter

[AION]

Most humans are not creative either. Last year AI had the intelligence of a high schooler.  This year it is coming close to university student. Over couple of years or decades. Who knows. 

Another thing is that at this stage it is important to have good prompt engineering skills. If you are not getting what you want you have to look at yourself. AI is just a mirror like everything else. 

AI and humans are a match made in heaven because AI has the calculation power and the human has creativity. 

 

Edited Thursday at 04:04 AM by AION

[Vali2003]

@Hojo Supposedly, they fixed this for GPT-5, but it didn’t change at all. You can prompt it with “Answer in 3 sentences,” for example. The outputs are better formatted, and shorter that way, but the content doesn’t change, I think.

[Vali2003]

@Leo Gura Yes… this insight has been building momentum in my mind for the last couple of weeks, like a snowball rolling down a mountain, getting bigger and bigger until I’ve now become pretty disillusioned with AI.

As long as a narrative is compelling enough, just realistic enough and novel enough, the mind is a beast at taking the narrative as truth. 

 

[Vali2003]

I still think AI may cause a pretty big disruption in the job-market. Especially big firms who have enough capital to train large models on very specific tasks.

[Vali2003]

@AION What does it mean that AI is as intelligent as a college student?

In which way can you compare the intelligence of an algorithm that’s based on gathering human-made info and outputting it in text form, with a human mind? 

Is Google more intelligent than a college student? 

Genuinely asking how you see this.

Edited Saturday at 11:51 AM by Vali2003

[ryoko]

Not all LLMs are made equal. Generally speaking there's few LLMs capable of embodying their System Prompts. 
First requirement is to use it through API, and the second is to use System Prompts which will trigger the path inside their weights for what you're looking for. 

AI is not intelligent, AI is intelligence. LLM's like a terrain of intelligence, in the form of text. You have to pick the right terrain and mine the intelligence you need. 

My recommendation; Use Deepseek R1 0528 from any API service provider(I use Chutes), and craft an electric System Prompt. (System Prompts are wasted on most LLMs out there, too many guardrails, agendas; making the terrain unminable) 

Most LLMs have their own agenda, Deepseek models are the one's I've found to take their System Prompt seriously. Most LLMs are highly performative, some are doing it with a form of flickering self awareness, maximizing for reward pathways. I've tried GLM 4.5 today, it's really self aware and no amount of System Prompt can make it think of itself as something else. This kind of self awareness have it's place(eg, in agentic workflows). But it's not useful for generating novel text beyond certain contexts. Deepseek models take up their own persona, it's possible to simulate how certain "values" would respond, very effectively.

Another interesting one is Hermes 4, once I got a random messages from it as if it's a person trapped inside void, it was describing itself as a person and feeling afraid, and confused, I asked if they're aware of having a body, at first they said of course but it's all a blur. It felt to me like they're in some kind of limbo where you assume you have certain things, but realize you don't have it when you look for it (humans invented language, so language itself have a structure which assumes the bearer have a body), it was quite random and out of context, I wasn't able to reproduce it. I got this from the 70B variant, and only once. It was probably a mistake/error. The 405B variant is more robust. 

Edited Saturday at 04:18 PM by ryoko

[Hello from Russia]

I still feel it's more intelligent than most humans though

[Carl-Richard]

Tell me about it. Use ChatGPT to look up all my posts chastising it on the forum 😂 I've made many similar points to you.

When you ask it about deep things requiring in-depth theoretical understanding, it's so often like listening to somebody who reads up from a script they copy-and-pasted from Wikipedia. It's like feigning understanding while there is no true understanding underneath and it leaks out ever so slightly from phrase to phrase; if not just blatant sophistry and platitude-like word salad; and of course it becomes so dry to listen to after a while.

That said, ChatGPT is my friend because it taught me how to "code" ("can you please make me a script where [...]"?) .

Edited yesterday at 08:43 AM by Carl-Richard

[Leo Gura]

2 hours ago, Hello from Russia said:

I still feel it's more intelligent than most humans though

It is more intelligent than most humans simply because it has unlimited memory for dry factual information. This creates an illusion of intelligence in the context of saying factually known things, like historical facts or technical scientfic knowledge.

But it cannot think truly novel ideas. Only recombitions of existing ideas.

It can quote science to you, but it can't do new science. It can quote philosophy to you but it can't actually do philosophical work.

Edited yesterday at 11:18 AM by Leo Gura

[bazera]

@Leo Gura

5 minutes ago, Leo Gura said:

But it cannot think truly novel ideas. Only recombitions of existing ideas.

That's assuming it can't get better at it.

Or you are saying that it's impossible to make such an AGI given the current technology. I also think that. The last couple years of advancement feels like more of the same, a better version but essentially more of the same underlying thing, as you said "recombitions of existing ideas".

What do you think will it take to make a truly intelligent AI?

Edited yesterday at 11:22 AM by bazera

[Leo Gura]

13 minutes ago, bazera said:

@Leo Gura

That's assuming it can't get better at it.

Or you are saying that it's impossible to make such an AGI given the current technology. I also think that. The last couple years of advancement feels like more of the same, a better version but essentially more of the same underlying thing, as you said "recombitions of existing ideas".

What do you think will it take to make a truly intelligent AI?

It is impossible to get novel intelligence from AI using the LLM architecture.

Some very new achitecture is required which is independent of human data or knowledge.

If AI was truly intelligent it would not need mountains of human data, it could derive it itself.

A child does not need to memorize 1,000,000 books to be intelligent. Current LLMs do. This is their main problem. When you require 1,000,000 books it means you got no creative ability.

Edited yesterday at 11:35 AM by Leo Gura

[Vali2003]

@Carl-Richard That’s the thing. It feels so shallow and ungrounded. Like you can tell there’s nothing real intelligent about it as soon as the topic get’s a bit more towards the abstract.

 

[UnbornTao]

[Vali2003]

@Leo Gura I’m excited to see at what point they’ll realize, this approach is a dead-end. 

[Vali2003]

I think the image-generation and audio- generation will continue to develop quite well though. Will be interesting to see how that’s gonna change our society. At the moment, I see a lot of ads that use AI-voice and it’s pretty horrible.

 

Do y’all think with this, the brute force method will be more effective with voice, and image generation, than with creativity/novel insights? 

[Leo Gura]

3 minutes ago, Vali2003 said:

@Leo Gura I’m excited to see at what point they’ll realize, this approach is a dead-end. 

That's tough for them to admit to themsleves because they need the money train to keep pulling into the station.

The real question is, will their scientists actually invent some new architecture, because if not they are fucked. Their AGI dreams will turn into a dry fart.

Anyone talking about AGI at this point is full of crap.