The AI safety layer destroys the r1

Asking my newly received r1 to read something for me it starts with telling me the things it won’t do. “I won’t read any names”, bla bla bla, then it seems to forget what it was asked to do and reads the text badly, then ends with reminding about that it is unable to really do anything because of the instructions it has gotten.

Apart from this being really boring, it kills most of the usefulness of having a generative AI in your pocket, What the Rabbit is out to do is super exciting, but also very difficult. The other players are super resourceful and some have a lot of head start. Tying your arm and leg behind your back before entering the race is not going to help. Please unleash the AI.

1 Like

Hey Pez,

I hear you. This is one of the drawbacks of cloud based models - they all, to varying degrees, have a pretty risk-averse model of compliance.

So to be clear, this is not in particular anything we’ve done - more the reality of being compliant with our vendor partners.

There are a few different ways this could play out:

  1. We work with our vendor partners more to try and finetune certain things that we think don’t make sense. We have done some of this already, but there’s a limit to how much we can ask of them, which is a function of scale. The better our company does, the more valuable we become to our partners, and the more we can push.

  2. We continue to add more vendor partners and have less reliance on any individual model. We’re considering whether it makes sense to give the user the choice of a default model. This would have drawbacks, but benefits in that any given user might have a preferential model for many reasons, one of which being risk tolerance. If you look at something like Grok, this is one of the selling points that Elon is using for it, that it is not as risk-averse as other models. Are xAI doing a license model for Grok, and if so, is that something we’d want to include amongst our other models? It’s a valid question, and there is no right or wrong answer.

  3. Local models. It’s not clear right now whether our hardware could support any, since that’s not what it was designed for, but it’s a possibility for future hardware. Even aside from the technical capability, it’s a business decision. Should our hardware be able to run a model that (for example) is so unrestrained that a user could ask how to make a bomb and the device will give the instruction? If so, how damaging is that for us as a company if it does, and someone films it, and the optics to the public (who, largely, don’t yet understand AI) put the blame on the company rather than the reality of software and information on the web?

4 Likes

Thanks for hearing me, Simon! Appreciate it. I was unaware that you vendor the models. But of course it is a huge task to train your own.

There’s a right answer =) Yes, Grok would be awesome. Grok s very helpful and doesn’t pretend to know better than I what is good for me.

I see how there are business risks with letting go of the ”safety”. As I see it there’s also business risk in making the r1 vastly less useful than it could be. r1 is not fully in the Gemeni league maybe, but it’s too close for comfort, I think.

3 Likes

I think for me I tend to user perplexity with 4o for many of my coding tasks and I use llama 3 for many of my creative tasks and also bring in Gemini now and then too on creative stuff. So I think it makes a lot of sense to leverage different models for different tasks.

1 Like

I for one would love to see Grok come to the R1

1 Like