Hacker News new | ask | show | jobs
by medler 17 days ago
Strange way to structure an argument. He seems to be arguing in favor of local models, but most of his examples are reasons local models are bad.
3 comments

How’s that “skipping vendor bloatware to get the printer working” example is bad?

Once upon a time that situation got us the whole Free Software movement going.

That’s the one out of three that’s not bad
Local models and alignment are orthogonal concepts.
How so? If it's not local, it's not yours, a third party owns it.

Why would you expect that third party specifically trained their model to be more aligned to you and your needs than to them and their (business) needs?

>Why would you expect that third party specifically trained their model to be more aligned to you and your needs than to them and their (business) needs?

That is my point. Why do you think Gemma, a local model trained by Google, is aligned to you and not the values of Google.

You can similarly have a model aligned to you which you pay someone to remotely host for you instead of running it locally.

I don't think this is entirely true, insofar as non-local models inherently have a "man cannot serve two masters" problem. The end user and the org running the AI may have different goals. That means even if AI alignment were to be in some sense "solved" for local models, you wouldn't be guaranteed an aligned AI using a non-local model because the "owners" of the non-local model may have deliberately misaligned the AI in their favor and against your interests.
Think of AWS. If you break their ToS and host a pirate website there it will work despite being misaligned from AWS. You are that sole master of the software. AWS is the sole master of the "hardware." Now ultimately hardware has control over the software so technically AWS could just hack into whatever code you are running and do what they want, but in practice they act solely as the master of the hardware and will tell the hardware to "shut down" if they catch you acting out of their alignment.

With local models just because you are providing the compute, that doesn't mean you are the master of the model and it is aligned to your interests. You could just be subsidizing inference for a model aligned to someone else.

Take it up with geohot
If you follow his recent posts he's on a bit of a spiral with how he thinks about AI.