As always in such cases this is 100% bull**. Either something is not working out for them and they have to delay in which case they could've just said so, or th

As always in such cases this is 100% bull**. Either something is not working out for them and they have to delay in which case they could've just said so, or this is some sort of pretense to show how "responsibility minded" they are. The reality is that bad actors have the resources to train their own stable diffusion on a dataset of whatever they want to deep fake and such delays do not slow them down one bit. What it does slow down is normal people using those models. From the smallest thing like mobilenetv3 through whisper, stable diffusion, CodeGen, and bloom those are huge productivity equalisers between the huge corpos and the little guy. Also the same thing can be said about frameworks like huggingface's. Just recently I was looking for a way to classify image type (photo or not photo[clip art, cartoon, drawing]) in an android app. Of course first hits on Google stear towards Microsoft Azure's paid API service. I was unhappy with having to use an over-the-Internet-API (with potentially sensitive end user's private pictures) so in one day of work I managed to download a pretrained MobileNetV3. A couple of 10k+ image datasets and I wrote <50 lines of python to tweak the last layer and fine tune the network. On rtx 2070 training took 10 minutes. Resulting accuracy on real data? 90%+. The model loads and infers in few hundreds of ms on modern phones(instantiating and loading takes longer than the inference BTW).This is priceless and 100% secure for end users. For thilose interested in the details I use ncnn and vulkan for gpu(mobile!) inference. Every commercial model maker's wet dream is to expose the model through an API, lock it behind a firewall and have people pay for access. This is not just hugely inefficient. It is insecure by design. Take copilot by example. I'm perfectly happy for all my hobby-grade code to be streamed to Microsoft, but no chance in hell I'll use it on any of my commercial projects. However faux pilot run locally is on my list of things to try. The first AI revolution was creation of those super powerful models, the second is the ability to run them on the edge devices.