Claude AI discussion

[Blackhole]

Dario ran his big heeb mouth and now he's getting exactly what he asked for. If you constantly lie in your marketing about how your super special slot machine is going to literally end the world and destroy everything and we need regulations NOW NOW NOW, don't be surprised when people fall for it!

 

[HootersMcBoobies]

Louis Rossmann has a pretty good take on the recent situation with the Fable model

I generally agree with Rossmann here. Anthropic originally called for a pause in AI development (A), then they release a shiny new model (based on the hype of Mythos, but with added "guardrails") and now they're specifically calling for regulation:

t is time to go beyond transparency to more serious and binding regulation of AI. I believe the best analogy, at least at the current stage of the exponential, is to cars, airplanes, or drugs—powerful technologies essential to the modern economy, but capable of killing large numbers of people if designed or operated poorly. I therefore believe we should model AI regulation on agencies like the Federal Aviation Administration (FAA). Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. I am grateful to see the Trump administration’s Executive Order move incrementally towards a greater role for government in AI, though Anthropic’s proposal recommends even further action.

src: Article Archive

When a company calls for regulations, it's to crush out competitors. What even is regulation in this case? It's literally saying: "It's illegal to make x type of software unless you submit it to the government." This gets into incredibly dangerous territory, like when the federal government went after the Tornado Cash developer, for a piece of open source software he developed three years prior to the case, because it might have been used by other state entities for illegal purposes. (This case was barely covered by the media. He was found innocent of most of his charges though, and the one charge that suck is very likely to win in appeal).

Towards the end of the Rossmann video, he gets into what I think is core in what seems like a propaganda narrative: this is going to come circle back to things like digital ID. It's going to come back to tying your ID to every Internet request, including every piece of code you publish or every AI generated image you prompt and post.

 

[CornBogFitz]

HEAR THAT INVESTORS?! out ai models are SOOOO good the GOVERNMENT HAS TO STEP IN! so good it’s scary. This is a totally real action by the United States government telling us to stop doing our jobs SO WELL because we’re SOOO GOOD. Give me a break. I’m so tired of this

 

[Betonhaus]

Wasn't the billing structure for companies changed this month? did they reverse the change when people threatened to cancel their AI contracts or did people not get the bill yet?

I wonder if companies will start suing the AI companies for convincing them to lay off all their employees only for the AI companies to demand extortion amounts of money just to keep their companies from screeching to a halt?

 

[quaawaa]

This is retarded, Chinese spies are going to get their hands on the code sooner or later
They are just delaying the inevitable

They have it already because all the AI labs hired them and trained them. There were no AI experts in China before that.

 

[Myrtle the Turtle]

this is going to come circle back to things like digital ID. It's going to come back to tying your ID to every Internet request, including every piece of code you publish or every AI generated image you prompt and post.

I believe this 100% and its horrifying. Theres a new article out in the WSJ saying AI deepfakes have gotten too good. The only way you can use AI is if you enter in your ID so they can trace back anything you keyed in. Any AI can be jailbroken no matter how many safeguards you try and implement. Thats going to come with its own flood of issues. What happens when Russia, China or Iran come up with their own native AI tools that will let you use it with free reign? How hard would it be to get a proxy in some third world country to put all your AI chats under? Its a clusterfuck of astronomical proportions.

 

[spider_expert]

Hope this doesn't result in lasting retardation, fable felt really solid and like an actual leap forward. I'm bummed about this tbh.

 

[Hannibalistique]

I'm a bit behind on this, what do you mean by Safeteyism, and how were they trying to control research?

Someone on twitter (can't find it, fucking timeline refreshed on me) was saying that if jailbroken the thing would give detailed (and accurate) instructions on how to make home made WMDs like sarin gas, but that's probably bullshit.

The big models have been capable of this for a long time. The only novel thing about Mythos is it supports bigger context (while staying coherent) plus they throw a ton of more compute at it (longer context = more matrices to smash together). The reason it comes now is because it needs to be trained on long contexts.

As for the actual facts in this case, apparently some Amazon exec packaged up one of pliny's jailbreaks showing Claude can recite the uncensored WAP lyrics and sent it to a government apparatchick (apparently they weren't aware that pliny "jailbreaks" every model and that the safetyism is a joke).

 

[Betonhaus]

So has anyone looked at the leaked financials for OpenAI? A 39 billion dollar loss last year is hard to spin.

 

[Premium Dog Food]

I paid 20 bucks to anthropic for Opus: (vague due to opsec)
Claude Code works directly in your directories (keep backups, though it hasn't destroyed anything yet), however the real good thing is that it can take control of your computer and debug issues on your side.
I would have ended up figuring out that one specific issue eventually, maybe in a couple hours. It only took 30 mins as it poked around and gathered data autonomously.
So far, I've been only using medium and sometimes high effort. Does the job.
Some times when using the remote control feature (MCP?) it was a bit slow and clumsy to navigate, so I told it to take a look at notepad and check if I wrote something if the window lost focus. I wrote it instructions or further guidance.
It worked decently, I think if there was proper a way to communicate to the AI mid-thinking or whatever the fuck it is doing to guide it better it would be cool.
Have not tried cowork yet, not convinced giving it full control of my computer is a good idea, nor do I have any reason to use it yet. Especially since Claude Code *did* poke around in folders outside the working directory somehow.

 

[Jerk Sausage]

Especially since Claude Code *did* poke around in folders outside the working directory somehow.

If you enable the yellow "auto mode", it'll do everything without prompting, including explore outside the working directory. The purple "accept edits" mode asks each time it reaches into a new directory or edits a new file. Auto is great if you're chunking through a lot of files, but yes can be dangerous if you're doing uncertain general "manage my computer" type tasks.

MCP is useful when you need context directly from a server or application that supports it. Best example is Chrome MCP; if you're doing app development, you can get details about a page, network requests, styles and everything directly from a real running instance. So claude code can make edits to your page, see the changes, check the network results, and confirm with confidence that it made the correct changes. Use MCP when you want a direct feedback loop, or up to date documentation, instead of manually copy-pasting results or references.

But yea claude code is a beast for just general computer/server management. Even vague asks like "is there anything running on my machine taking resources which can be disabled" or "organize this mess of files into a good folder structure".

The best thing to do, if you're nervous about unintended effects, is to end a prompt with "Prompt with me to confirm before making any changes". That way you can read and approve any mutations before they happen.

 

[debristar]

So has anyone looked at the leaked financials for OpenAI? A 39 billion dollar loss last year is hard to spin.

you have to understand there exists no framework to allow for audit.

this is no longer at organization or government level, this is an international problem.

 

[Betonhaus]

you have to understand there exists no framework to allow for audit.

this is no longer at organization or government level, this is an international problem.

Will it become a "the rich are now the only nutrition source available to us" kind of problem?

 

[skunt]

DARIO YOU FAT JEW FIX CLAUDE. i have slop to ship and it's been down for hours.

 

[Hannibalistique]

DARIO YOU FAT JEW FIX CLAUDE. i have slop to ship and it's been down for hours.

Goddamn, you made me look at their status page, and I guess it shows the dangers of vibe coding, so much red and orange: https://status.claude.com/

 

[Jerk Sausage]

It's coming back today. https://www.anthropic.com/news/redeploying-fable-5

 

[Premium Dog Food]

Info that nobody asked:

I tested the blender integration with Opus 4.8:
Absolutely atrocious geometry. You ask it to make a floor plan based on a blueprint image, it somewhat manages to do that and labels the rooms correctly. You ask it to make an apartment block, skyscraper or building in general, it finds it complex to do anything other than a rectangle building.
Unsurprisingly, it excels at geometry nodes and materials. Both good time saves, especially geometry nodes which I have not bothered to learn anything more than the basics yet.

Today I tested Fable 5:
Way better at geometry, doesn't stop at simple boxes. Still messy, terrible geometry, but for a floor plan it sets up lights, doorways, windows and all that stuff. Way better at creating arbitrary geometry, such as an apartment block or skyscraper (needs further testing), of course, way more token hungry.

So I would still make the geometry myself as what it outputs is unworkable, good enough for a proof of concept, but Claude may get better at it over time.
I've read it's good at rigging characters, which makes sense, but I have not tested that nor do I have models lying around.
I have not tested animations or keyframes, but I guess it could also work with those.

As for the coding side, I haven't tested Fable much, yet. It did not refuse the coding tasks I gave it (contrary to what I read everywhere) but it just ate through the entire 5 hour limit trying to diagnose the (very complex) issue, in 1 hour.