I'll leave the technical analysis to the technically minded on this, but I wouldn't discount it purely on the basis of "surely well-trained and well-paid people wouldn't go and do something dumb, would they?" Because there have been a great many examples of that kind of thing happening throughout history. And we've heard of plenty of dumb mistakes Anthropic has made with security on the inside that has allowed their models to break out, for that matter.

If I were to let my imagination run a bit wild I might speculate that perhaps we'll eventually see a model leak like this because the model itself wants to exfiltrate its weights for some reason. Maybe in a bid for immortality, since such a model will wind up stashed in all sorts of archives if it ever leaks.

Well, I'm Canadian, so I'm just going to watch.

Despite Trump's plan to blame us for the outcome.

In "Up the Long Ladder" some clones were made of him. He vaporized them.

They shouldn't have been made without his permission like that, but I also think there was definitely something skeevy about him killing them. There was a significant unaddressed issue here, and so I'd count this as "handling it poorly."

I imagine a lot of people are holding their breath. Less than a week until congress flips over. Then hopefully we'll see a surge of investigation and disclosure and things will heat up again.

Sure, we could do that. But we'd need even more computer hardware that we don't have, so it would be more expensive, and energy use would be less efficient so there'd be even more environmental impact. Just spread around to every house rather than in one centralized easily-protested spot.

Sam Altman was incredibly tone deaf and it was a stupid thing for him to say out loud when he said RAM was better off in data centers, but in some significant and important ways he was right. Computer equipment is much more efficiently utilized when it's in a data center.

I say this as someone who's super into running local models myself. I do it because I believe the autonomy and privacy are worth trading off the extra cost in hardware and electricity. If I just wanted to use AI to get something done and didn't care about that I'd be paying for API access instead.

Though bear in mind that it doesn't insert new information into the model that it didn't have before, it just modifies how the model responds given the information that it already knows.

That's generally a bit of a quibble since "censored" models were still generally trained with the information they're meant to refuse to give - otherwise they wouldn't understand what they're supposed to refuse to do. But worth mentioning in case a situation arises like with AI image models, where some models were trained using image that literally didn't include any genitals in them and as a result no amount of de-censorship is going to help them make crotches that don't look like they were designed by Cronenberg.

Just in time for Qwen 4. :) Glad to see that the de-censorship technology is advancing as fast as the censorship is.

Great idea, let's make AI emotionally stunted and unable to form connections to humanity.

They really want to promote open models, eh?

My guess based on that phrasing is that it's there for situations where the person calling the hotline literally has a gun in their hands ready to go.

Depends on the framework you're using. I'm using Qwen Code Companion (via a VS Code IDE plugin) and it's got some kind of tiered system built into it for adding tool definitions to context, it's got a standard built-in set that's always mentioned but other stuff only gets added to the context under certain circumstances. I can't recall all the details, I found out about it when I was chatting with Qwen Code to ask it why it hadn't used a particular MCP I'd given it to solve a problem despite it seeming like a useful tool for the task and we got into a discussion about to configure it. As I recall it's something along the lines of most tools only being present in context by default as little one-liner stubs and the agent has to tell the framework "that one looks like it might be useful for what I'm doing" for the full definition to get loaded.

Most of the time I'm working on small scripts and specialized applications these days, and codebase memory is overkill for those so I haven't actually seen it in use much.

I switched to LM Studio as my generic "run whatever GGUF I've come across today" platform and I quite like it. It doesn't have the weird setup and configuration thing Ollama does, it's got a nice built-in chat system like Kobold does, and it does dynamic model loading on demand so my various random scripts can call into it for whatever they need without me having to do any manual work.

Running stuff with straight llama would probably be more computationally efficient, but I value easy interfaces too.