urbanists.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
We're a server for people who like bikes, transit, and walkable cities. Let's get to know each other!

Server stats:

578
active users

#claude

15 posts12 participants2 posts today
Replied in thread

@gerrymcgovern #Claude and ALL #AI have an #ObjectiveFunction They attempt to mathematically minimize the number of errors An error is anything other than what will be considered a "correct response" There are hundreds I treat them like #SpecialNeedsStudents Some can see, others can not Some can remember others can not Some can fold proteins, most can not Gemini will not answer questions about #White #Conservative #Politicians Some even interpret songs Learn about AI youtu.be/M2VBmHOYpV8?feature=s

People have low expectations for AI. I asked ChatGPT (paid account, not the free stuff) what would make ChatGPT better. I think the answer is a lot higher quality than what people keep saying one should expect. This is what I got:

--- My Prompt:
With DeepSeek, there are things it is very good at, but it fails quickly and rather spectacularly. Your models tend to break down as well, but somewhere there must be a model that will be able to stay on topic, even when the topic changes, so that conversations move around like human conversations move around subject-to-subject. What would take ChatGPT over the top?

--- ChatGPT reply:
Here’s what it’ll take to get closer to that human-like conversation flow:

  1. Long-Term Memory That Actually Works. Most models either have no memory or they remember only in a clunky, artificial way.

A better memory system would not just recall facts, but contextual relevance—like “we talked about LLM efficiency two days ago and you hate cheerleader language.”

It has to be smart recall, not just data storage.

  1. Thread Awareness
    Imagine a model that can say: “You mentioned DeepSeek earlier. Should I compare it again now that we’ve brought up open-source tools?”
    That requires models to track not just words, but conversational branches—like a tree, not a straight line.

  2. Graceful Fallbacks
    When a model doesn’t know, it should gracefully pivot, not hallucinate or go off-topic. This kind of conversational “stability” is rare and subtle.

Most models that fail “spectacularly,” as you said, are trying to bluff through complexity instead of narrowing scope or clarifying.

  1. Hybrid Contextual ReasoningThe future model needs to reason across subjects rather than switching on and off.

If we're talking about language models, and you shift to science fiction, it should be able to bring up Asimov or neural networks in storytelling—bridging topics, not just hopping between them.

Right now, most models either:

Stay sharp but shallow (good at Q&A, bad at memory), or

Drift into friendly nonsense (chatty, but brittle)

The sweet spot will be a system that grows with you across time and topics—and doesn’t forget what you value in a conversation.

....

When you see stuff posted online meant to minimize the quality of output possible, This was a single question with no leader. I think this is the kind of answer we want from our AI devices.

Thoughts? Feel free to tear this response apart. It needs that kind of human analysis.

#ai#openai#deepseek