Open Weights Models— Good or Bad?

Some models on the market are open-weight models, but are they good enough?

Do we need them? Yes! But are they safe?

Let us understand it in detail.

In open-weight models, the weights of the model are made available by the AI company. This makes it possible to run the models locally on our systems.

Moreover, we can edit the weights.

This is the problem: editing, even if initial weights are frozen! This can be Parameter-Efficient Fine-Tuning, Merging Weights, or Direct Code Manipulation.

As long as people do not edit the weights, it’s good enough.

Even if we can test and reproduce evaluation and testing results with open weights, that is good enough too.

But what about editing?

There are AI EU rules that once a model is tested for all rules and anomalies, it’s ready to be sent to markets. What would AI regulations work for then?

But what if someone, for fun, or with serious intentions, changes weights on a local system?

Say, 32GB, runs 14B to 32B parameter models, and weights are changed by any of the above methods.

What if they add bias to the edited model?

What if they add risks, cyber sensitive things to the system by appropriately modifying the model?

This makes it a new model, as weights are changed now!

But then?

Then it needs new evaluations, new testing on all set standards of models!

How can you let an edited model, especially a model edited for, say, cyber threats, enter our lives?

Yes, open weights are good as long as they are for evaluations and testing.

But if you create a whole set of new Frankenstein with it? Who shall check?

If there is an edited model in each other’s home in the world? Would it be good?

Give open weights to be transparent, but do not allow editing! This is possible! Not quite!

So this means monopoly over AI models?

Well, the safety of systems around the world is a must!

For that, we can’t add bias, discrimination, cyber threats, politically motivated AI, intentional AI, or even harmful and dishonest AI audio, text, or videos.

So, if there is a way to not edit weights in open-weight models, it would be good.

Otherwise, we all will save money, but we all would be at risk.

As anyone would run biased, politically or otherwise biased models! And cause havoc in the world.

We don’t want these people! Tell me! Do we?

What’s the way out then?

How to reduce prize then?

How to make models accessible then?

Even running the edited model needs resources and money!

At the same time, editing weights can be for good too-to fix bias in a system, to fix political alignment in a system, we can just edit the models. This responsibility can be given to the company itself when it’s their showoff time.

So, these are two sides of the same coin! Which one do you choose? Tell me?

Thank you for reading.

Subscribe for updates.