upvote
Enjoy it while you can. You can see the AI companies have already started the process of locking us out of this, with Opus 5.5.

God forbid our fans, our fans of all things, might be broken of their dependencies on the internet. The plebians can't be allowed to do that. Every relay must be cloud-connected so it can be monetized and disabled for whatever dumb ass reasons someone might want to disable it.

reply
Just use abliterated/uncensored self-hosted models. This is why we need them and the hardware to run them. Yes, the API models which will further close the walled gardens will evolve faster than the open-weights ones, but I'm hopeful about the community as a hole rejecting this status quo.
reply
Is there hardware you would suggest that you can actually buy today? Pricing is out of control right now, if you can even find stock.
reply
An RTX4090 or a 128gb macbook can generate a pretty reasonable number of tok/sec. Those are not cheap, but within the realm of serious hobbyist. I really hope the AI bubble deflates a bit so consumers can get more powerful hardware at the low thousand $ range.
reply
The other problem is that the manufacturers can also use AI to improve their DRM and do red-teaming to fix "vulnerabilities" before release. Right now they're not because it actually takes some effort but eventually they'll be able to just tell their agents "make sure no unauthorized clients can connect". I expect that several years from now there will be a sizeable market for pre-2027 hackable devices.
reply
Coway sells tons of devices with no smart/connected features. Their best-reviewed ones for home air purification are not connected. Everyone will be better served by supporting such product lines or pressuring companies to create them more than they will be by reverse engineering "connected" features no one ever needed.
reply
> locking us out of this, with Opus 5.5

I hear this a lot but I've never hit this.

For weeks now I've been working on a shared recompilation project of Switch 1 and 2 games, and Opus 5.5 has handled that with no issue.

Over 10 supported Switch 1 games, and 3 supported Switch 2 games, and never once has Opus 5.5 denied anything

reply
> I hear this a lot but I've never hit this.

Opus 5.5 will absolutely refuse to touch anti-cheat systems in video games, even when the end game isn't to cheat but to work on Linux compatibility or to bypass/spoof the checks to enable gameplay.

reply
OpenAI and Anthropic models seem to be the most affected by these barriers, from my own experience.

Mistral, QWEN, and other models (read non-American) tend to have little to no restrictions. There were some tasks where I was actively trying to sabotage my own local network and Mistral had no issue helping. Anthropic and OpenAI refused outright or had to be "convinced" but would often hit further barricades along the way.

reply
Sure but as easy it is to hack these devices as easy it is to harden it.

So this only works on products that was developed before ~now.

The solution is to not buy crap regardless if you can make it work or not.

reply
> So this only works on products that was developed before ~now.

You're under the assumption that the companies producing these give a shit about security. It'll probably increase a bit but I doubt it will be by much. If they have shipped devices that leave the doors wide open for almost any intrusion, they won't invest engineering/AI time and money to make these better than before and certainly won't review the code that the model spit out beyond a quick check if the promised base functionality still works.

reply
I'm not. But no AI will produce anything that is as bad as the state of current products.
reply
> So this only works on products that was developed before ~now.

The Home Assistant hackers are some of the most persistent reverse engineers that I have seen. Almost as much as game crackers and cheat developers.

reply
Apparently there's a website which already addresses your concerns about fans - nothing else - only fans.
reply
at least current models are safe, glm is very capable in terms of RE
reply