Base model
Hi,
Thanks for releasing Phi 4! Are there any plans to release the base model?
Thanks!
they never did that before. by nature, there is no real base model, I think.
I mean, they could release a model that has no RL, but I bet it is always kinda an sft instruct model.
if they could strip away the synthethic data, the base model would not even be interesting.
but even the model without RL, it has some degree of ethical alignment. I like to see it is MIT licensed as before.
Base model release becomes difficult because of safety guidelines. Since it is not post trained, it can say unwanted things. Since MS takes safety as a priority.
No final plans as of now. But we are trying.
Great release! LLM safety seems like field still in a flux. Personally I think tokens don't do direct harm and information is first of all just information, good, bad, right, false. My worries are more about really potent agents, but those are not around.
A question is how much we should entrust people to be able to 'do stuff'. With AI, 'stuff' scales more than other actions, but then again, it's not like the hundreds of millions of people who drive cars all try to run over other people. In the context of LLMs, one can think about the impact on the noosphere as a whole, but then again, in that space, only limited safety makes sense, because humans are by themselves unsafe epistemic agents, which we hope should have freedoms to express themselves. When AI is seen as a form of art, there should be ample room for human expression there as well. As I think I should be able to type bad words into my keyboard, without my keyboard deleting those and telling me I get a fine for thinking bad things. That's AI safety twisted to its negative. Noosphere safety, killing human nature. I don't think we are risk free beings.
That said, certain things can be made more difficult for better safety. But where a reasonable balance is, is not 100% clear.
I feel like adults should be allowed to consume or interact with legal but morally wrong content, just like video game violence.
And you're also free to code, no black hat hacker can do without code. Still we don't ban code.
With regards to noosphere, I think influence should be more elegant than banning tech. With really powerful agents though, I think things can quickly go wrong, not just by wrong action but also by accident. I think the space of thoughts is more resilient than the systemic influence of lots of disruptive actions. Then at least some applications needs to be illegal. Nobody can forbid people to build a robot and create unsafe software that controls it, but it is feasible to prevent commercialization of unsafe robots.
there might also be ways to act on it in distribution, but how I don't really know.
There must also be an assessment of reasonable safety. When they started to do research in physics, they could not anticipate someone could one day build a nuke with it and hence forbid it. or well, perhaps the church did at some point. Maybe part of AI safety is a somewhat religious thing.
But I'm not so afraid of tokens or (legal) pictures, as we as humans manually generate text or images also in culturally embedded and not 100% bias free way, as values are biases and those change over time, can swing back and forth. This is not meant to not advocate for being mindful about the noosphere, but I think there are smarter ways to influence it in a positive way. Humans always have the freedom to do very bad stuff, but they don't end up doing that all day. Why is that? Only because of fines? We can also work on not becoming barbarian cultures, where influence is strong in settings that set basic values in personality formation.
And with advancing AI, I think less biased approaches replace worse ones by merit, not by enforcement. I'm looking optimistic in the future where we have kinda AGI that might go beyond LLMs that outperforms humans in avoiding unethical bias in a contextualized manner. The negative of negatively disruptive actions might be just something new but positive new, systemically.
I would also love a base model. They're very valuable to have