Nobody cares about "AI" (Chatbot: I disagree.)

[Free]

Put the cat on an electric unicycle.

Problem. Solved. Kinda.

 

[Argent Stonecutter]

Yeh, I was thinking of adding a chain drive.

 

[Free]

Uh oh, gonna need another Grok "Hitler" filter installation.

Trump’s new AI chatbot keeps fact-checking him

Truth Search AI feature promises to deliver users ‘direct, reliable answers’

www.the-independent.com

The Truth Search AI feature, which is integrated into the web browser of Mr Trump’s social media platform Truth Social, is designed to deliver “direct, reliable answers”, according to its creators.

The chatbot’s responses typically draw from right-wing and pro-Trump news sources like Fox News and Newsmax, however, it did not support recent statements made by Trump.

On Friday, Mr Trump said that his tariff policy has had a “huge positive impact” on the stock market, but the AI tool responded, saying “the evidence does not support this claim”.

First reported by The Washington Post, Truth Search AI also called the US president’s claim that the 2020 election was stolen “baseless”.

When asked about a recent post by Mr Trump on Truth Social about crime in Washington, D.C., the AI bot responded that it is “not totally out of control”.

 

[Noodles]

As luke warm as I am on AI. I find it incredibly hilarious that these right wing idiots keep maming bots that are supposed to support their stupidity and the bot still points out their bull shit.

 

[Free]

So AI is basically just like us. Only, you know, sloppier.

Researchers Made a Social Media Platform Where Every User Was AI. The Bots Ended Up at War

Time to start over.

gizmodo.com

Social platforms like Facebook and X exacerbate the problem of political and social polarization, but they don’t create it. A recent study conducted by researchers at the University of Amsterdam in the Netherlands put AI chatbots in a simple social media structure to see how they interacted with each other and found that, even without the invisible hand of the algorithm, they tend to organize themselves based on their pre-assigned affiliations and self-sort into echo chambers.

The study, a preprint of which was recently published on arXiv, took 500 AI chatbots powered by OpenAI’s large language model GPT-4o mini, and prescribed to them specific personas. Then, they were unleashed onto a simple social media platform that had no ads and no algorithms offering content discovery or recommended posts served into a user’s feed. Those chatbots were tasked with interacting with each other and the content available on the platform. Over the course of five different experiments, all of which involved the chatbots engaging in 10,000 actions, the bots tended to follow other users who shared their own political beliefs. It also found that users who posted the most partisan content tended to get the most followers and reposts.

 

[Free]

I don't know whether to laugh, cry, or have a sandwich. I think having the sandwich would be best.

Character.AI Gave Up on AGI. Now It’s Selling Stories (archive.today)

Last August, Google swooped in with a roughly $2.7 billion deal to license Character.AI’s technology. As part of the agreement, Character.AI’s two cofounders left for Google’s AI division.

Anand, who previously worked as the VP of business products at Meta, was tasked with picking up the pieces—which he did in part by leaving behind the founding mission of delivering personalized superintelligence to focus on AI entertainment.

“What we gave up was this aspiration that the founders had of building AGI models—we are no longer doing that. That is the hundreds of billions of dollars investment fight, which Big Tech is fighting,” Anand says. “What we got in return was clarity and focus, being able to singularly pursue the AI entertainment vision.”

Instead of reaching for the big shiny ball of "AGI," their purpose now with AI is merely a Hollywood-like template for pseudo-social media fantasies. That is so great.

 

[Noodles]

I have not used character.ai in ages.

They had a Tom Riddle bot, I mocked it for being a Mudblood, until it got mad and started screaming Avada Kedavra.

Then I mocked it more with things like, "Oh, I am still alive, I guess your Moodblood magic doesn't work well enough."

 

[Dakota Tebaldi]

I don't know whether to laugh, cry, or have a sandwich. I think having the sandwich would be best.

 

[Essence Lumin]

The main thing that pisses me off about ai is how much energy and resources it is taking. When people come to their senses and realize it is a bunch of crap we will have much more energy available maybe.

 

[Innula Zenovka]

There's a lot of recipe data on the web, so there's patterns existing that it can pull in remaining consistent with its prompt.

The article you linked seems to be related to Andy Clark's work too. But the way we make predictions is not by parroting language, pre-linguistic animals follow the same course, they build mental models of the world and make predictions based on them. Our underlying thought processes that we narratize in language have to be operating the same way.

LLMs don't make predictions and test them, if they seem to be surprised that's because they're mimicking surprise, and have no mental models.

I realise that, but ChatGPT 5, at least, seems to be capable of some remarkable feats, no matter how it's achieving them. Over the last couple of days it's been a great help to me in debugging long and complex scripts.

The first one was a reasonably simple task. When I tried to run the script it threw a math error and crashed. I knew this was almost certainly because I was trying to divide something by zero -- that is, divide something by a variable that didn't yet have a value assigned to it -- but I didn't really want to look through several hundred lines of LSL trying to find it. So I gave ChatGPT a copy, told it what the problem was and what I thought was causing it, and asked it to try to find it.

It found the error in a minute or two (it would have taken me far longer) but it also pointed out that there was a nearby, but unrelated, logic problem that would certainly have caused issues when the script ran. It was simple enough -- I'd copy-pasted something from elsewhere in the script and forgotten to change one of the variables it referenced -- but I was impressed that ChatGPT found it without being asked.

The second one was more complex. I have a long and complicated script that's always required laborious tweaking and debugging every time I've changed anything to do with the access control checks. I needed to make some more changes for an update and decided to ask ChatGPT if it could find what was causing the script to stall because items weren't being properly cleared from a list of tasks once the task was completed.

It found two clear bugs, either or both of which could cause that problem under particular circumstances (it didn't identify the circumstances, but once I knew what the problematic bits of code were I could tell what had been happening), and suggested various neat solutions. This fixed a problem in minutes that I've been poking at for hours over the last couple of years.

However it's achieving this, I find it a remarkably useful tool.

 

[Argent Stonecutter]

Pattern matching with neural nets has a long history, and LSL is similar to many langues it can match against. Again, this is using it as a sophisticated search engine. The problem is when you attribute agency to it or treat its supposedly subjective responses as anything but parody. And the way its pattern matching is applied it will throw together completely erroneous and unrelated fragments to produce a credible but wrong result if there isn't a good match already in its training data. And its apologies are *always* that kind of output.

 

[Erich Templar]

Argent, although I largely agree with what you are saying, you are a bit like a broken drum at this point, and it's a little wearying to go through a thread like this knowing that every time I come across one of your posts it will just be repeating what you have said in every other post, just phrasing it slightly differently.

A bit like AI, actually...

 

[Innula Zenovka]

Pattern matching with neural nets has a long history, and LSL is similar to many langues it can match against. Again, this is using it as a sophisticated search engine. The problem is when you attribute agency to it or treat its supposedly subjective responses as anything but parody. And the way its pattern matching is applied it will throw together completely erroneous and unrelated fragments to produce a credible but wrong result if there isn't a good match already in its training data. And its apologies are *always* that kind of output.

I think we're at cross purposes. I'm not attributing agency to ChatGPT. I'm saying I'm finding it an increasingly useful tool for debugging scripts, regardless of how it does it. Similarly, I've found it's been very helpful when I've asked it to help me brainstorm the overall design of complex inter-region communications systems for a network of HUDs and objects.

I'm aware of its limitations, and I think I find it useful because I'm sufficiently experienced with LSL to know when ChatGPT's advice is useful and when it's off-beam. So with the brainstorming I'm able to assess its recommendations, rejecting those that are nonsensical or off-beam, accepting some as good ideas I'd not thought of, and accepting others as good ideas but not the best way of doing whatever it is.

I don't attribute agency to the in-world LSL Compiler either, but I find it a useful tool, too, for spotting errors and typos and trying to show me where in the script to look for them.

 

[Erich Templar]

...I'm sufficiently experienced with LSL to know when ChatGPT's advice is useful and when it's off-beam...

This is eventually (soon?) going to become a major problem. AI works well enough when the person using it is knowledgeable enough to be able to spot when it is going off the rails. As people begin to rely on it more and more, that kind of knowledge is going to fade away, and we'll be left with people using AI to create ... sub-optimal solutions.

 

[Erich Templar]

Relevant: Why LLMs Can't Really Build Software - Zed Blog

 

[Dakota Tebaldi]

Mozilla is currently busy forcing AI features into its browser. With the most recent update, a lot of Linux users began reporting that their computers were starting to sound like jet engines whenever Firefox was running, with all the fans running at full speed. Turns out Firefox was spiking their CPU usage up to 130% running a LLM to do tasks like search browsing history or automatically name tab groups.

Mozilla repeatedly stresses that the LLM is a "local model" that "runs privately" on users' computers and doesn't share any data with Mozilla, which is certainly good, but it doesn't seem to have occurred to them that LLMs are very resource-expensive and most people's computers are not really good to run local models on.

 

[Noodles]

ChatGPT, please give me code for a bare bones browser with tabs that is not based on Google's spyware code.

 

[Dakota Tebaldi]

LibreWolf is Firefox with all of the bad stuff stripped out.

By default, for privacy reasons there's also several options turned off in it that are usually on by default in most browsers. So you might have to turn some of those features on, if it turns out some websites you like to use are broken when you try to visit them. But turning them on is easy and at least you know that only the things you absolutely need are turned on.

 

[Free]

More serious legal malpractice on the AI front.

Judge criticises lawyers acting for boy accused of murder for filing misleading AI-created documents

A judge has criticised lawyers acting for a boy accused of murder for filing misleading information with the courts after failing to check documents created using artificial intelligence.

“It is not acceptable for AI to be used unless the product of that use is independently and thoroughly verified,” Justice James Elliott told the supreme court in Melbourne.

The documents related to a 16-year-old boy, who was on Thursday found not guilty by way of mental impairment over the murder a 41-year-old woman in Abbotsford in April 2023.

Defence apologised to the judge for the error and re-filed documents with the court.

But Elliott said “the misleading information caused by AI did not end there” and the revised documents referred to made-up laws.

“Revised submissions were not reviewed by either side … and referred to legislation that did not exist,” the judge said.

Is anybody lawyering correctly anymore?

 

[Argent Stonecutter]

Vibe Representation.